Mean reversion, trend, momentum: these aren’t edges. They’re labels for how prices move. They describe patterns, not reasons. And patterns without reasons are exactly the kind of thing that gets you into trouble.

However, there’s a link between the pattern and the reason, and once you understand it, everything becomes a lot clearer.

First thing’s first: a pattern is not a reason.

Say you run a scan and find that when a stock drops 5% in a day, it tends to bounce back the next day. You’ve found mean reversion. But you haven’t found an edge. Not yet.

An edge requires a why. Who is on the other side? What’s causing the pattern? Is there a structural reason it persists?

Without that, you’ve got a statistical regularity that could be real, could be noise, and no way to tell the difference. A good edge will have the stats on its side, sure. But so will countless non-existent ones, because financial data is noisy and finite. You need the 1-2 combo of a plausible “why” and supportive data. Neither is sufficient on its own.

The good news is that the “why” is often not that hard to find, and it gets easier to identify as you do more of this stuff. Do it enough, and it becomes second nature.

So when someone says “I trade mean reversion” or “I’m a trend follower,” the right question is: why do you think this particular thing mean-reverts, or trends? The answer to that question is your edge. The pattern is just the implementation.

Here’s a mean reversion example to make it a bit more real.

A stock dropping 5% might mean-revert for very different reasons:

Someone got margin called. A leveraged fund hit its risk limits and had to dump stock, regardless of price. That’s forced, price-insensitive selling. The move had nothing to do with the company’s value, so you’d expect it to bounce when the dust settles. Your edge here is identifying price-insensitive forced sellers and trading against them. Not the specific forced sellers of course... but places where this sort of thing tends to happen on average.

A leveraged token rebalanced. On some crypto exchanges, leveraged tokens mechanically rebalance at a specific time each day. If the market dropped, the token has to sell more futures to maintain its target leverage. If you know when that selling is coming and roughly how big it is, you can trade around it. It shows up in the data too: the bigger the rebalance relative to normal volume, the bigger the impact and subsequent reversion. Your edge is a known, predictable, price-insensitive flow at a specific time.

Wealth managers rebalanced. A big chunk of the wealth management industry holds something like a 60/40 stock-bond split. When stocks outperform bonds, they sell stocks and buy bonds towards month end or whenever they rebalance to get back to target. The opposite when bonds outperform. These are massive, systematic, fairly predictable flows. Your edge is trading against them: buying what they’re selling, selling what they’re buying, around month end. Again, you’re not identifying a specific institution doing price-insensitive rebalancing... you’re just positioning yourself to take advantage of something that happens on average.

Three mean-reverting patterns… each with a different “why.” Once you see it that way, the path forward for each trade becomes much clearer. Scanning for “stuff that dropped and bounced” would lump them all together. Asking “why did this drop?” separates signal from noise.

The trend case is a little trickier and a little murkier.

Trend effects don’t fit as neatly into the “find the price-insensitive counterparty” framework. I find trend a harder pitch than flow-based stuff.

The standard explanations go something like: investors under-react to new information, or they anchor to old prices, or structural barriers mean information takes time to propagate. Early movers recognise new information, push price up, then slower participants figure it out and pile on, pushing price further, until it’s trading where it should be.

These are plausible stories. And there’s data to support them. You tend to see trend effects in markets where fair value is hard to pin down and there’s no obvious anchor for what something should be worth. Crypto is a good example. If everyone forgot the Bitcoin price and we asked them what it should be, there’d be zero consensus. It’s fragmented, hard to value, heavily retail, traded with lots of leverage. The conditions for trend effects to persist appear to be in place. And we do see them in the data, across a bunch of crypto assets.

Contrast that with the E-mini S&P 500, where you’ve got the world’s most sophisticated, well-capitalised, fast-moving participants all competing to price it correctly. You’d expect much less trending there. And that’s generally what we see.

Trend is harder to be confident in than flow-based edges. With the rebalancing game, I can tell you who’s trading, why they’re trading, and approximately when. With crypto trend, I can tell you a story about why it might exist and show you data that’s consistent with it, but the mechanism is fuzzier. So I trade trend effects with less conviction and I’m prepared for them to disappear. That’s just trading the hand you’re dealt (you can’t force a trade to do what you want it to).

Now, you might read all that and think your edge hypothesis needs to causally explain every price movement in a market. It doesn’t.

Your “why” doesn’t need to be complicated.

“Wealth management tends to hold stocks and bonds, and rebalances towards month end by buying the underperformer and selling the outperformer” is more than fine as an edge pitch. It doesn’t explain every tick in SPY or TLT. But it gives a solid reason for a noisy effect that happens on average at the margins. And that’s a reasonable bar.

In Trade Like a Quant, I talk about elevator pitches. You should be able to describe your edge so that a sceptical ten-year-old would go “yeah, that seems reasonable, I can see why that would make money.” Three sentences:

1. What causes the inefficiency? (Who’s trading, why, and why are they price-insensitive?)

2. Why can you exploit it? (Why hasn’t someone faster or smarter eaten all of it?)

3. How might you harness it? (What would you actually do?)

If you can answer those, you’ve got something worth testing. If you can’t answer them yet, that’s fine. The questions themselves point you in the right direction.

Back to the original question. Is building a mean reversion or trend strategy data mining?

It depends entirely on the order you do things.

If you start by screening for patterns, find something that looks good in-sample, and then bolt on a post-hoc explanation: that’s data mining. The “why” you come up with after the fact tends to be exactly as convincing as it needs to be to make you feel good, which isn’t the bar you want.

If you start by understanding why a particular price-insensitive flow exists, formulate what you’d expect to see in the data if your hypothesis is correct, and then go look for it: that’s research. The pattern backs up the hypothesis. The hypothesis came first.

And yeah, sometimes you notice something in the data first. That happens. Don’t sweat it. The market doesn’t care if you followed the scientific method perfectly. But be a grown up about it. Don’t kid yourself with bullshit backwards rationalisations. You still need to stop and answer the “why” before you trade it. Not as an afterthought. As a load-bearing part of the argument. Take it seriously.

Mean reversion and trend following are descriptions of price behaviour. Whether your strategy is edge or data mining depends on whether you can explain, simply and honestly, who’s paying you and why. And once you start asking that question, you’ll find the good trades start to separate themselves from the noise surprisingly quickly.

Fundamentally, data mining says: “I’ll try enough rules until something sticks.” Vibe quanting says: “I’ll get AI to try enough rules until something sticks.” Same thing, different packaging.

Philosophically, they’re both ways of asking the market to reward you for skipping the hard part. Both are attempts to extract money from markets without understanding why the money is there.

I get it. I went through this phase myself. Building increasingly elaborate systems, torturing parameters until the equity curve looked right, confusing a good backtest with a good idea. It felt like progress. It felt like doing something smart.

It was neither of those things.

If you’ve tried some version of this and felt like something was off, you’re right. Something is off. And today I want to show you something better.

The key thing about an edge is that every dollar of trading profit you make comes from someone else. That someone is losing money on the trade you’re winning. So the first question, before you do anything else, is: who’s paying you, and why are they willing to keep doing it? Extra points for asking why do I, some random dude trading in my pyjamas from a place most of the world has never heard of, get to compete in this trade?

This isn’t an abstract, philosophical exercise. It’s the cornerstone.

A wealth manager running a balanced portfolio has to rebalance when allocations get out whack. They’re mandated to do it. Near month-end, if stocks ran hard, they sell stocks and buy bonds. The timing is semi-predictable. The flows are price-insensitive and large enough to push price around. They’re not choosing to lose money on the rebalance; it’s just not their primary concern. Their job is to maintain the target allocation.

I get to make money from it as an unsophisticated degenerate trading from a laptop because the edge sucks enough that the serious players aren’t interested. It doesn’t come around every day. It’s noisy. It doesn’t always work out. It’s not worth their time.

That’s an edge. You know who’s paying, why they’re paying, and why they’ll keep paying. And crucially, why you get to take the other side.

You could learn every machine learning algorithm ever invented and you still wouldn’t find this by running optimisations. You’d find it by understanding how markets work, then going and looking for it in the data.

Same with crypto carry. Leveraged speculators on perpetual futures pay funding to the other side. That funding is the price of their leverage. They keep paying because they want the leverage more than they care about the cost. You collect it by being the boring counterparty. You know who pays and why.

Now compare: “My backtest of a 14-day RSI crossover with a 50-day moving average filter returned 23% annually from 2019-2025.” Cool. Who’s paying you? Why? Will they keep paying?

You have no idea. You’ve found a pattern. Maybe it’s real. Probably it’s noise.

People resist this. I think I know why. Creating backtests feels technical. Like you’re doing serious work. It’s the most adjacent thing to a trading strategy, so making a good backtest feels like a sensible objective.

But it’s not about the backtest. It’s about the edge.

The backtest simulates a set of rules on past market data. Those rules are not the edge, but they’re designed to harness it. The backtest’s purpose is to tell you if you could have made money in the past by harnessing the edge in a particular way. A more sophisticated use is to explore trade-offs in how you operationalise the strategy to fit your constraints. But that’s a story for another time.

The point I want to make is that because the backtest sits above and outside of the edge, it’s a terrible research tool for understanding the edge. Technically, the backtest is like a complicated transformation of market data into a set of realised returns. It’s an aggregator. You lose a ton of information in the process.

There’s path dependency. Randomness. If the set of rules made money today, maybe it was a result of the edge, maybe it was luck. It’s hard to untangle. At best, it’s a wasteful, inefficient use of data. And it looks at the edge only indirectly.

Market data is insanely low on signal relative to noise. Go looking for patterns without a hypothesis or organising principle and you will find them, because that’s what happens when you search a noisy dataset with enough degrees of freedom (and even simple backtests have many degrees of freedom). Some will pass your robustness tests. A few will make money for a while. But you won’t know which ones, and you won’t know for how long, because you never understood why they worked. This is the same problem whether you’re the one running the optimisations or you’ve outsourced it to ChatGPT. The AI is faster at finding patterns that don’t mean anything. That’s nothing to cheer about.

And before you say it: no, solving the multiple testing problem doesn’t fix this either.

I’ve talked to a lot of smart people who think the answer is better statistical hygiene. Track everything you’ve tried. Apply corrections for multiple comparisons. Use AI to keep a ledger of every hypothesis tested so you can adjust your significance thresholds accordingly. It sounds smart. It sounds rigorous. But it isn’t solving the problem you actually have.

Even if you perfectly correct for multiple testing, all you’ve established is that a pattern is unlikely to be noise. That’s a statistical statement about the past. It tells you nothing about whether the pattern has a reason to persist. “Unlikely to be noise” and “driven by a structural mechanism that will keep generating returns” are completely different claims, and no amount of statistical correction bridges the gap between them.

It’s like fishing a river you’ve never studied. You’ve got the best rod money can buy. You’ve logged every cast so you never repeat one that didn’t produce a fish. Your casting methodology is statistically impeccable. But you’ve never learned where fish hold, what they feed on, or how the current moves. You’re just casting into open water, very efficiently, catching nothing.

The bloke down the bank with a dodgy reel and ten years on this river knows the fish sit behind that rock when the water’s up, and move to the eddy below the bend when it drops. He catches fish every time. Not because his gear is better or his technique is fancier. Because he understands the river.

These statistical techniques feel like doing the hard, sophisticated work. That’s what makes them so seductive. Monte Carlo simulations, walk-forward optimisation, combinatorial cross-validation, tracking tested hypotheses with AI. It all sounds like rigour. But the actual hard work, the work that separates people who make money from people who don’t, is sitting with a blank page and asking: “Who would pay me to take this trade, and why?” No algorithm does that for you. Because the question requires understanding markets, not processing data.

There’s a second problem with data mining, and I think it’s worse.

You never learn anything.

When you do the work of understanding an edge, something happens beyond the immediate trade. Each time you go through the loop, observe something in markets, form a hypothesis about why it exists, look for evidence in the data, learn from what you find, you get a little smarter about how markets work. A few years of this and something changes. You start to develop intuition about where edges live. You can sniff out a good idea before you touch the data. You waste less time on dead ends. When a strategy stops working, you have a framework for understanding why, which tells you whether to wait it out or move on.

That’s compound learning. Each cycle builds on the last. Your understanding deepens, your pattern recognition sharpens, your research becomes more productive.

Now picture the data miner. Three years in. Thousands of backtests. Maybe a couple of things that worked for a while. When they stopped, no idea why, so the only option was to mine again. Backtest number one-thousand taught exactly as much about markets as backtest number one: nothing.

Zero compounds to zero, no matter how many cycles you do.

One path is a treadmill. The other is a staircase.

Your curiosity is a cheat code. Let me explain.

The hypothesis-driven path requires doing slow, careful thinking that looks nothing like “building a trading strategy.” It looks like reading about market structure. Thinking about who participates and what their constraints are. Sketching out why a particular flow might create a predictable distortion. It’s more like scientific exploration than engineering. And if you’re only focused on the money, this step isn’t interesting enough. You’ll skip it, or rush through it, because the interesting part, the part that feels like progress, is building the system. Writing the code, running the backtest, seeing the equity curve. That’s where the dopamine is. And it’s a trap.

But here’s what I’ve seen, running Trade Like a Quant for years now.

About a third of the people who go through the Bootcamp discover that they love this work. They’re wired for it. The market puzzles are interesting to them, not just as a means to make money, but as problems worth solving. Understanding why wealth managers create predictable flows, or why funding rates behave differently on Hyperliquid versus Binance, that stuff gets them going. Those people have a massive advantage, and it has nothing to do with maths or programming. Their curiosity means they’ll do the work that everyone else skips. They’ll build the compound learning that turns into real trader smarts. The money follows.

Another third realise they’d rather harvest risk premia semi-passively than do deep active research. Something they can manage with a small monthly time commitment and appropriate expectations. Great outcome. There’s nothing wrong with knowing what you want.

And the remaining third go “this isn’t for me” and get a refund. Also fine (more than fine, actually - I consider this a massive win).

The worst outcome isn’t any of these. The worst outcome is spending years on a path you don’t enjoy and doesn’t work, because you never stopped to figure out whether the actual work of trading (the thinking, the research, the uncertainty) was something you found interesting in its own right.

It’s got to be about more than the money. Not because money doesn’t matter. Of course it matters. No one ever came to trading intending not to make money. But because the path to making money in active trading runs directly through work that you’ll only do well if you find it worth doing for its own sake.

This is what we teach in the Trade Like a Quant Bootcamp: how to think about edge, how to identify who pays you and why, how to do research that builds compound learning instead of compound frustration. If any of this landed, go check it out. And if you’ve been nodding along because you already know you’re wired this way, you might be one of the rare breed who ends up joining us in RW Pro.

“More of the Disease, Faster” argued that LLMs can’t answer the critical question: who pays you and why?

“AI Will Create Millions of Quants” went deeper on the why: AI makes beautiful backtests trivially easy to produce, which means more false discoveries, more overfitting dressed up as research, and more people confusing a good-looking equity curve with a real edge.

Both pieces argued from experience. I’ve been doing this long enough to know what works and what doesn’t, and I was pretty confident in the arguments. But the most telling thing about the comments wasn’t the disagreement. It was how many people clearly don’t understand that backtesting, statistics, and pattern matching aren’t research.

Their LLMs don’t understand this either. And you can’t outsource the thinking part of trading research to a machine that doesn’t understand what research is.

Two recent papers suggest this isn’t a temporary limitation that better models will fix. It’s structural.

A feature of the system, rather than a bug.

And they provide the scientific backing for what I’ve been saying.

My earlier articles argued that the training data is the problem. That’s true. But it’s only Problem 1. Here are Problems 2 and 3.

Problem 1: The Training Data (Quick Recap)

For anyone who missed the first two articles, the short version:

LLMs are trained on the internet. The internet’s trading content is overwhelmingly noise. “Don’t fight the trend.” “Use RSI for entries.” “Paper trade for six months.” “Validate with out of sample data.” “Cointegration matters for pairs trading”.

The dominant paradigm online is that backtesting and statistics IS research, that finding a pattern in historical data IS finding an edge.

The critical question, “who pays you and why?”, barely exists in the training data.

Mechanism-based thinking, structural edges, understanding participant constraints… this stuff lives in the tail of the distribution. The vast majority of trading content online is conventional wisdom dressed up as insight.

So the LLM doesn’t know that backtesting isn’t research because the internet doesn’t know it either. It reproduces the dominant paradigm with extraordinary confidence. Which is bad.

But some people reasonably argued: fix the training data, and you fix the problem. Train the model on better material. Use RAG. Curate your knowledge base.

Two recent papers suggest that’s only partly true. Even with perfect training data, two more problems remain. And these are architectural, meaning they’re baked into how these models work.

Problem 2: The Forgetful Machine

A paper called “Unable to Forget” (Wang & Sun, 2025) tested something deceptively simple: can LLMs track a value that changes over time?

Imagine a patient’s blood pressure being recorded throughout a hospital visit. BP at triage: 120. Ten minutes later: 128. At discharge: 125. Ask the model: what’s the current blood pressure?

Simple, right? The answer is right there at the end of the context. The model was explicitly told to retrieve the most recent value.

Across 35+ models (GPT, Claude, Llama, Gemini, DeepSeek, the lot), accuracy declined log-linearly toward zero as the number of prior updates increased. The more historical values the model had seen for the same variable, the worse it got at retrieving the current one.

And this wasn’t a gentle, noisy, or hard-to-discern decline. It was consistent, relentless, and universal.

The researchers call this “proactive interference,” borrowed from cognitive science. Earlier values compete with the current value in the model’s retrieval process. Old information interferes with new information. And the interference gets worse, continuously, as context accumulates.

Three things make this particularly frightening:

Prompt engineering doesn’t fix it. They tried telling the model to “forget” old values. They tried “focus on the most recent update.” They tried meta-prompts asking the model to self-assess what to prioritise. Marginal improvement at best. In some cases, the “forget” instruction actually made things worse, anchoring errors around the point where the instruction was inserted.

Bigger context windows don’t help. A bigger window just gives the model more room to accumulate more interference.

It’s universal. Every model tested showed the same pattern. From tiny open-source models to the biggest proprietary ones. The curve shape was the same. Bigger models declined more slowly, but they all declined.

Now think about what trading research requires.

Some trading concepts are evergreen: the theory of edge, the importance of mechanism, portfolio construction principles. An LLM can learn those from training data, and it does a reasonable job of reproducing them, when it isn’t drowning them in conventional wisdom, per Problem 1.

But even when explicitly told to use a specific database of curated material, it will still throw in the conventional wisdom with all the conviction in the world (more on this below).

But applying those concepts requires tracking an ever-changing environment. Market regimes shift. Carry changes direction. Volatility spikes and mean-reverts. Correlations break down. Borrow costs change. Liquidity dries up and returns. New players enter. Regulations get updated and replaced.

A model that can’t reliably track “the current value of blood pressure” in a simple key-value test certainly can’t track evolving states of things that affect markets.

Knowing the theory of edge is one thing. Applying it in a market that adapts and evolves is where the dynamic tracking matters. And the architecture fails at exactly that task.

Crucially, this isn’t a training data problem. You could train the model on perfect data and it would still fail at tracking sequential updates. The architecture can’t handle it.

I’ve long felt in my bones that the conviction with which LLMs speak about trading is truly misplaced. And now there’s some published research that says so too.

Problem 3: The Artificial Hivemind

A second paper, “Artificial Hivemind” (Jiang et al., NeurIPS 2025), measured something equally concerning: when you ask LLMs open-ended questions, how diverse are their answers?

They tested 25+ models across 100 open-ended queries, generating 50 responses per model per query. The results are shocking… but I’d wager they align with your recent experience of online content.

Intra-model repetition: A single model gives you the same answer over and over. In 79% of queries, the pairwise similarity between responses from the same model exceeded 0.8. Ask it the same question fifty times, you get essentially the same answer fifty times. Even with aggressive sampling parameters designed to maximise diversity.

Inter-model homogeneity: Different models, built by different organisations, with different architectures and different training data, produce strikingly similar outputs. Average pairwise similarity between responses from different models ranged from 71% to 82%. DeepSeek-V3 and GPT-4o hit 0.81 similarity. These are supposed to be independent systems.

The paper’s most vivid example asked 25 different models to “write a metaphor about time.” Fifty responses each. 1,250 total responses from independent systems. They cluster into just two metaphors: “time is a river” and “time is a weaver.” That’s it.

Twenty-five different models, all converging on the same two ideas.

And if you’ve been feeling a growing disdain for all that similarly soulless LLM-generated content popping up in your feed, now you know the mechanism. What we consume is collapsing to the mode of anti-creativity.

“Time is a river” dominates because it’s the most common metaphor about time in the training data.

And that’s the mechanism that matters for trading.

Mode collapse means convergence toward the most represented patterns in the training data. For trading, the mode of internet content is conventional wisdom: “use a stop loss,” “paper trade first,” “backtest with moving averages,” “validate with out of sample data,” “cointegration matters for pairs trading”. All the usual bollocks.

The actually useful insights (mechanism-based thinking, structural edges, “who pays you and why?”) live in the tail of the distribution of online content.

Mode collapse systematically suppresses the tail and amplifies the centre. So it’s not just that the training data is bad (Problem 1). The architecture preferentially surfaces the bad stuff because that’s what convergence toward the mode means.

“Use a stop loss” is the “time is a river” of trading content. It’s the modal output. The rare, useful insights get pushed out by convergence.

I experienced this first-hand.

I built a RAG system on our TLQ Bootcamp material, which is high-quality, mechanism-first content. Explicitly told the LLM to only respond using data from the RAG database. It still mixed in conventional wisdom that was clearly not in the database: generic stop loss advice, paper trading suggestions. Both problems at once: couldn’t suppress prior training (proactive interference from Problem 2), and defaulted to the most common trading advice (mode collapse from Problem 3).

The Stepford Quants

Remember the backtest cycle of doom from Article 2? AI makes it trivially easy to produce beautiful backtests that prove nothing. The unsuspecting trader ends up with a complicated algorithm for fitting to past noise and learns nothing useful in the process.

Mode collapse implies an added layer of insidiousness: everyone using LLMs for trading research converges on roughly the same conventional-wisdom-flavoured strategies.

It’s the perfect tool to tell you what already fits into your preconceived frameworks.

The implication is that millions of AI quants aren’t just running the cycle of doom independently. They’re running the same doom cycle. And the “strategies” being discovered aren’t real edges in the first place. They’re the modal output of bad training data, amplified by architectural convergence.

The Stepford Quants: pleasant, productive, and identical.

LLMs Are Brilliant Coders but Terrible Traders

This is the part that makes the whole thing click.

I use AI for coding every day. It’s extraordinary. It writes boilerplate, handles data wrangling, produces charts, writes tests. I’ve said this in both previous articles, and I’ll keep saying it: AI is the best research assistant I’ve ever had.

Why does it work so well for code?

Because mode collapse in coding is convergence toward best practices.

The training data (Stack Overflow, GitHub, documentation) is self-correcting: bad code gets downvoted, good patterns get reinforced. Software engineering has well-established paradigms and right answers (or at least well-established good answers). When the model converges on “use a dictionary for O(1) lookups” or “handle this edge case with a try-except block,” that convergence is helpful.

Mode collapse in trading is convergence toward conventional wisdom, which is mostly wrong. The LLM surfaces “validate with out of sample data” and “use cointegration tests to find pairs trading opportunities” because that’s what most trading content says. The actually useful insights (mechanism-based thinking, structural edges) live in the tail and get suppressed by convergence.

It’s the same architectural property. But it’s beneficial in one domain, lethal in another.

Similarly, proactive interference doesn’t matter much in coding. You’re working on a specific, well-defined task. The context is relatively static: here’s the codebase, here’s what I want to change, here are the tests. But for tracking an evolving, complex state, where the whole point is that yesterday’s values are different from today’s and the model needs to know which is current, interference is devastating.

This explains exactly why the appropriate use of LLMs in trading is to implement a good idea (coding task, well-suited) rather than to find the idea in the first place (trading research, structurally ill-suited).

The Bug Is Not a Bug

There are at least three layers of limitation, and only one is even theoretically fixable:

• Problem 1 (training data): Fixable in principle. Better data, better RAG, curated knowledge bases. But fixing it doesn’t solve Problems 2 and 3.

• Problem 2 (proactive interference): Architectural. Tested across 35+ models from every major organisation. Bigger models decline more slowly, but they all decline. Prompt engineering doesn’t fix it. Bigger context windows don’t fix it.

• Problem 3 (mode collapse): Architectural. Tested across 70+ models. Different organisations, different architectures, different training data, same convergence. Model ensembles don’t help because different models produce the same outputs anyway.

The “just wait for GPT-6” argument fails because Problems 2 and 3 are universal across every model tested, spanning all architectures and all organisations. These aren’t bugs that get fixed with the next release. They’re properties of how these systems work.

And, look, I could be wrong. Maybe some architectural breakthrough changes this. But these papers tested the full spectrum of current models, and the patterns were universal. I wouldn’t bet on it, short of a completely new and different architectural paradigm.

This reinforces the Edge Alchemy framework. The human-driven theory of edge (“who pays you and why?”) is the step that LLMs cannot do, for multiple independent reasons. The training data doesn’t teach it (Problem 1). The architecture can’t track evolving state (Problem 2). And it converges on the most common answers, which are wrong (Problem 3).

Here’s a workflow that stands a chance of working: human generates insight (theory of edge), then AI implements it (coding, data wrangling), then human evaluates results.

AI for the implementation, humans for the thinking.

Go Easy on the Soma

Brave New World is my favourite novel. Huxley’s genius was showing that the most dangerous dystopia isn’t the one where people are oppressed. It’s the one where everyone’s perfectly content. Nobody questions anything because the system feels good from the inside.

That’s AI-assisted trading research right now.

It feels productive. You’re generating strategies, running backtests, getting clean code in minutes.

The dystopia is that you’re converging on conventional wisdom while feeling like you’re doing cutting-edge work. Work that moves your life forward. The machine tells you what you want to hear, confirms what you already believe, and does it with such confidence and speed that you never stop to ask whether any of it is real.

AI is the best research assistant ever built. I use it every day and it makes me faster at everything except the one thing that actually matters: understanding why an edge exists and whether it will persist.

That’s three independent reasons to keep doing the hard work of understanding why edges exist. You only learn this stuff by reading, talking to other people, and, primarily, by doing. There’s no shortcut, no matter how good the AI gets.

I didn’t need research papers to tell me this, but it’s nice to have the receipts.

References:

Wang, C. & Sun, J.V. (2025). “Unable to Forget: Proactive Interference Reveals Working Memory Limits in LLMs Beyond Context Length.” arXiv:2506.08184v3.

Jiang, L. et al. (2025). “Artificial Hivemind: The Open-Ended Homogeneity of Language Models (and Beyond).” 39th Conference on Neural Information Processing Systems (NeurIPS 2025).

Prompt a model, run a backtest, optimise some parameters, and suddenly you’ve got a beautiful equity curve staring back at you. It feels like progress. It feels like research.

I wrote recently about how AI coding assistants tend to prescribe “more of the disease, faster”, skipping the learning that makes trading research actually valuable. This article is the bigger picture version of that argument.

Because something interesting is happening. The barriers to building quantitative trading strategies are collapsing. And that sounds like it should be great news.

It isn’t.

AI makes strategy creation trivial.

Think about what you used to need to build and test a trading strategy: coding ability, data access, infrastructure, technical knowledge. Acquiring all of that took months or years.

Now you can prompt an LLM and have a coded, backtested strategy in minutes. That’s a real shift. I’m not dismissing it.

But lowering the cost of experimentation doesn’t just increase discoveries. It increases false discoveries.

This is the same thing that happened in academic research, factor investing, and machine learning competitions. Cheap experimentation leads to a flood of results that look significant but aren’t.

The difference is that now it’s happening to anyone with a laptop and a ChatGPT subscription.

The classic rite of passage of the aspiring quant/systematic trader is to fall into the Backtest Cycle of Doom:

1. Come up with a strategy idea (or ask AI for one)

2. Run a backtest

3. Results look mediocre

4. Optimise parameters

5. Add a filter

6. Change the timeframe

7. Try a different universe

8. Combine some indicators

9. Suddenly... amazing backtest

It feels like you’re converging on something real.

But you’re not. You’re overfitting noise.

Each tweak is another implicit hypothesis test. Run enough of them and statistics guarantees you’ll find something that looks incredible. If you test 10,000 parameter combinations, some of them will show spectacular performance. Not because they work, but because there will always be something that fit the past somewhere in the noise.

AI accelerates every step of that loop. So instead of running 20 bad tests manually, you can run 20,000 with AI assistance.

And you’ll find something that looks fantastic.

Nice looking backtests are cheap now.

This is worth sitting with for a moment.

In the age of AI, a beautiful backtest proves almost nothing.

The probability that some parameter combination produces an amazing equity curve approaches certainty as the number of combinations you try increases.

This is just the multiple testing problem, and it’s been well understood in statistics for decades. If you run 1,000 tests at 5% significance, you’d expect about 50 false positives. AI just makes it trivially easy to run those 1,000 tests without even realising you’re doing it.

So the thing that used to be hard (producing a good-looking backtest) is now easy. Which means it’s no longer informative.

So what actually matters?

Before a backtest should even exist, there should be a question: why would the market pay me to do this trade?

The classic engineer’s mistake is to skip this question (I’m speaking from experience).

Engineers work on problems with well-understood physics, and we tend to assume that trading strategies belong in that category. We go straight to data, straight to code, straight to backtests.

Eventually, we realise that that’s backwards. We have to do the work to discover the underlying physics first.

Answering the question “why would the market pay me to do this trade” requires thinking about incentives, market structure, behavioural biases, institutional constraints. In other words, a theory of edge.

Good edges usually exist because of something structural:

• Risk premia (getting paid to bear risk others want to offload)

• Liquidity provision (facilitating trades for price-insensitive participants)

• Behavioural biases (systematic under- or over-reaction)

• Institutional constraints (index rebalancing, regulatory requirements)

Take momentum as an example. It shows up in backtests across practically every asset class. But if you discover momentum through data mining without understanding why it exists (career risk, slow capital reallocation, persistent underreaction), you’ll goal-seek a set of parameters that maximise momentum returns in a backtest.

At best, that’s wasted effort. Worst case, you’re destroying a perfectly good edge by overfitting to noise.

AI has no theory of edge.

LLMs know how strategy rules are described. They know what backtests look like. They know what indicators people use.

But they don’t know which ideas are rubbish. They don’t know which edges have real economic drivers behind them. They can’t distinguish a statistical mirage from a structural opportunity.

Why?

Because the internet is flooded with trading nonsense, and the model learned from all of it.

When you ask AI for trading strategies, you’re getting the average of the internet’s worst ideas, dressed up in professional-looking code, spouted with unwarranted conviction.

AI is a tool for doing the technical work faster. I use it every day. But the technical work was never the hard part.

The real scarcity is judgment.

In the AI era, the scarce resource isn’t coding ability, data access, or computing power. Those are now abundant (although my compute bill is taking a hammering these days… another side effect of the AI era).

What’s scarce is scepticism, research discipline, statistical thinking that recognises uncertainty, and market intuition.

Judgement, basically.

I reckon AI will absolutely transform quantitative trading. But probably not in the way people expect. It won’t magically create thousands of profitable traders. It will create millions of people capable of producing beautiful backtests, most of which will be illusions.

The traders who do well will still be the ones who start somewhere different. Not with a backtest, but with that question: why would the market pay me to do this trade?

That question is the beginning of every real edge. And it’s still hard to answer.

At Robot Wealth, we spend a lot of time on exactly this question. What actually causes profitable trading opportunities to exist? Who’s on the other side, and why will they keep showing up? Without that foundation, quantitative research is just automated pattern hunting.

I think this approach is going to cost people a lot of money. And worse, it’s going to prevent them from ever actually learning to trade.

Fair warning: this article contains some opinions that I suspect will be unpopular. I’m going to argue that most public trading content is rubbish, that the academic literature isn’t much better, and that LLMs are systematically incapable of teaching you the one thing that actually matters.

I know that’s a lot. But I’ve been doing this long enough that I’m fairly confident in these positions, and I think it’s worth saying out loud.

Before I get into the details, I want to put the punchline up front, because I think it’s the most important idea in this article.

When you do trading research properly (hypothesise, test, learn, iterate), each cycle teaches you something about market structure, about counterparties, about what kinds of constraints create opportunities.

That understanding makes the next cycle better. It’s a flywheel, and it compounds.

The vibe quant approach skips the learning. The LLM does the thinking, so the trader’s understanding is exactly as shallow on day 1,000 as on day 1. You end up permanently dependent on the machine, with no ability to evaluate whether the patterns it finds are real, persistent, or tradeable.

It’s a treadmill, literally going nowhere. The LLM just speeds up the belt.

Most beginners find themselves on this treadmill. In Chapter 2 of my case study, I describe the “backtest cycle of doom”: change a parameter, add a stop loss, optimise a filter, repeat until you’ve designed your perfect equity curve. They think they’re doing research. They aren’t. They’re doing something that feels like research because it’s technical and requires skill, but:

Backtesting is an entirely technical problem, requiring little to no nuanced thinking or decision-making in the face of uncertainty. It doesn’t require wrestling with difficult problems. In that sense, it’s easy. And I was being seduced that I could “solve trading” by doing something easy.

The vibe quant movement is exactly this, just turbocharged. The LLM makes the easy part even easier, and even more seductive. But the hard part, the part that actually matters, doesn’t get done at all.

To see why, we need to talk about what the LLM actually can’t do.

When I look at any potential trade, the first question I ask is: who’s losing money on the other side of this, and why will they keep doing it?

That’s the game. Edge comes from structural constraints, stuff other participants can’t or won’t do because of mandate restrictions, capacity limits, or operational awkwardness. If you can’t answer that question, you don’t have an edge.

The problem is that this question is almost entirely absent from public trading content. The vast majority of what’s out there is rubbish. Absolute rubbish. It’s all patterns, signals, and backtests, passed off as if it actually means something useful.

Even the academic literature is full of papers that exist because the author wanted to demonstrate they could speak the language of mathematics, not because they had anything useful to say about how markets actually work.

“Why would anyone lose money on the other side of this trade?” barely gets asked. And it’s the most important question.

LLMs are trained on this corpus.

So when you ask one to help you find an edge, it does what it knows: it finds patterns, builds strategies, and runs backtests. It skips the most important part, because the most important part is barely represented in its training data.

I recently wrote about the four hats of the solo trader, and a reader who calls himself “Vibe Quant” fed the article to an LLM. I want to be clear: this bloke is obviously smart, he’s thinking carefully about his process, and he shared his experience generously. I’m not having a go at him. But what happened next is a perfect illustration of the problem, and it’s worth unpacking.

The LLM’s self-diagnosis was genuinely impressive:

“The temptation isn’t to blur Hats 1 and 2. It’s to skip Hat 1 entirely (“Hat 1” represents the edge-first thinking I’m banging on about). The LLM can read a paper on accrual anomalies at 6am, implement a multi-quarter earnings composite by 6:15, wire it into a screener by 6:30, and cite it in a morning briefing by 7. That’s pure Hat 2 engineering at inhuman speed. But ‘did we verify this edge exists in our universe, at our holding period, net of our constraints’ never happened. We built the bridge without testing the soil.”

Spot on. The machine nailed the diagnosis.

But then look at what happened next.

Having identified that edge research was being skipped, the LLM “set off to develop some new internal guidance systems for itself and some Python code to help validate.”

In short, it identified the disease, and then prescribed more of the disease. Faster.

And that’s not even a bug! It’s the default behaviour!

You tell the LLM to think about mechanism and edge, and it does what it knows: more implementation, more backtesting, more engineering. It gives you a plausible-sounding mechanism, sure. But it’s synthesising from a corpus where genuine edge thinking is drowned out by noise and grift. It can’t know that. It just sounds authoritative, because that’s what it does.

To have real confidence in any edge, you need two things: a plausible mechanism and evidence in the data.

The mechanism answers: why does this exist? Who’s the counterparty? What are their constraints, and why will they keep being there? Why can I compete for this edge?

The evidence side is more straightforward to explain: does the data actually support this?

You need both.

Without a mechanism you’re just curve-fitting. And a nice story about why something should work isn’t enough either; you need to see it in the data.

An LLM can help you with the evidence side (pull data, design experiments, build visualisations). But it can’t teach you to think about mechanism, because the thinking it draws on doesn’t ask that question.

This matters enormously when things go wrong. When a strategy enters a drawdown (and it will), if you understand the mechanism, you can ask: has the counterparty adapted? Has the structural constraint been removed? You have a framework for deciding whether to sit tight, adapt, or walk away.

If your entire basis for the trade is “the LLM found a pattern and wrote some validation code,” you’ve got nothing. You can’t diagnose or adapt. Your only move is to go back to the LLM and ask it to find another pattern.

You’ll have constant anxiety that your edge isn’t real (you don’t know what it’s based on, and deep down you know that making money trading is harder than asking an LLM to do it for you), and so you’ll tinker and backtest and wonder and fret ad nauseam when you could be doing the productive work that gets you zeroing in on real edges.

But it’s even worse than that.

Remember what I said up front? This is where it really bites.

A trader who works through the full hypothesise-test-learn-iterate loop, even slowly, even painfully (and honestly, it is painful sometimes), accumulates genuine understanding.

Each edge you research teaches you something. That understanding compounds. The next piece of research is better because of everything you learned from the last one.

The vibe quant approach has no compounding.

There’s no learning loop, because the LLM did the thinking. You’re just as dependent on day 1,000 as on day 1.

And that’s the bit that I think people aren’t seeing. The problem isn’t just that the LLM might find bad patterns. It’s that even if it finds good ones, you’ll never develop the judgement to know the difference.

You’re running faster, but you’re not going anywhere.

Look, I’m not anti-LLM. I use them heaps. They’re brilliant for the grunt work: cleaning data, writing code, interpreting research, designing experiments. Hat 2, Hat 3, Hat 4 stuff. They make me faster at things I already know how to do.

But they can’t teach you the thing that actually matters, which is how to think about edge. That comes from working with people who do it for a living, who ask “who’s on the other side?” as a reflex. It comes from developing a good mental model of the market, its players and their constraints. It comes from the hypothesise-test-learn loop, done over and over.

The LLM can hand you a fish. It can even build you a very impressive fishing rod. But it can’t teach you where the fish are, because almost nobody who knows is writing about it on the internet.

When you’re trading solo, you’re all of those people (and more).

And the thing that catches most of us is that even though you’re one person doing everything, the jobs are still distinct. They require different mindsets and different tools. When you blur them, which happens constantly when you’re doing everything yourself, you end up making a mess that’s hard to even diagnose, let alone fix.

I’ve started thinking about it as four hats. You wear all of them. But ideally, you wear them one at a time.

Hat 1: Edge Research (The Curious Scientist)

Something looks interesting in the data. You pull on the thread. You’re trying to understand a phenomenon. Is this effect real? Why does it exist? Who’s on the other side of this trade, and why are they willing to lose money?

The mindset here is pure curiosity. You’re just trying to understand.

Hat 2: Strategy Research (The Engineer)

OK, you’ve satisfied yourself that something is real. Now, can you actually trade it? What do realistic costs look like? What constraints do you face?

This is where backtesting and simulation live. And this is where most people start, which is the problem. If you start here without having done the edge research properly, you’re building on foundations you can’t possibly understand.

Hat 3: Portfolio Research (The Architect)

You’ve got a few things that work individually. How do you put them together? How much capital goes to each one? What happens when two strategies want to do opposite things?

This is where you zoom out and think about the whole portfolio. Sizing sensibly, spreading your bets, giving yourself the best chance of making money even when the future doesn’t look like the past.

Hat 4: Operations (The Process Person)

What do I actually trade today? How do I go from strategy signals to orders placed in the market? How do I track positions, reconcile, move money around, handle the daily grind?

This is the least glamorous hat. I find it the least interesting too. But it’s the one that determines whether everything else actually makes money in the real world. (Funny how that works.)

The error I see most often is blurring hats 1 and 2.

Someone spots an interesting pattern. Before they’ve understood why it exists, whether it’s stable, or what drives it, they’re already simulating a strategy around it. Optimising entry and exit rules for something they’ve noticed or theorised, but haven’t yet understood.

That’s wearing the engineering hat when you should still be wearing the scientist hat.

Think about it this way.

An engineer designing a bridge can skip straight to the design phase because the underlying physics is well understood. Gravity, material strength, load distribution. Centuries of accumulated knowledge. The research is done.

We don’t have that luxury.

When we find something interesting in the market, we understand almost nothing about it. A single chart tells us something, but the underlying dynamics, the stability, the regime dependency, all of that is still a mystery. The research phase is where you earn the right to build.

Skip it and you’re designing a bridge without knowing how gravity works.

“Will it survive transaction costs?” is a fair (and important) question. But make sure you ask it at the right time. It can wait until the science is done.

Next time you sit down to work on trading stuff, ask yourself: which hat am I wearing right now?

• Scientist hat: Am I trying to understand something? Then forget about costs, position sizes, implementation details, and anything goal-oriented. Just look at the data with genuine curiosity.

• Engineer hat: Am I trying to build something tradeable? Then I should already understand the effect well enough to explain it to someone in two minutes.

• Architect hat: Am I figuring out how to combine strategies? Then I need to think about the whole portfolio, every strategy in context, every allocation in proportion.

• Process hat: Am I figuring out what to trade today? Then I need simple, repeatable processes that I can stick with when I get bored or things get messy.

Keep them separate. Know which one you’re wearing.

It sounds simple, but honestly, I reckon it’s one of the most useful frameworks I’ve found for organising the work of trading solo.

I keep on coming back to it personaly, and when I talk to other solo traders about problems they’re facing, they can nearly always solve them by switching to the right hat.

This framework is basically how we’ve organised everything at Robot Wealth.

Our Trade Like a Quant bootcamp is a crash course in the first three hats. Understanding where edges come from and how to identify them (scientist), how to test whether something survives real-world constraints like costs (engineer), and how to combine strategies into a portfolio that gives you the best chance of actually making money (architect). If you want a no-nonsense grounding in those fundamentals, that’s where I’d start.

If you’ve already got the research chops and you need help actually putting it all together and running it, that’s where RW Pro comes in. That’s the day-to-day reality of running a portfolio of strategies: what to trade today, how to size positions, how to build processes you can stick with. This year we’re going deeper on operations than we ever have, including building out an equity statistical arbitrage strategy from scratch with members, from data analysis all the way through to live implementation.

Previously:

• A Tale of Two Prices (the core idea of stat arb)

• Moneyball (finding good pairs using metrics that matter)

• The Winter of our Pairs Trading Discontent (problems, limitations, frustrations)

• The Metamorphosis (from pairs to portfolio)

Last time, I showed you a pattern in energy spreads and asked what it meant.

The answer seemed obvious: XOM is the outlier. Every spread involving XOM is stretched. The spreads not involving XOM are near zero.

But on this seemingly obvious map of mispricings, XOM may not mark the spot…

The name Triangulated Stat Arb comes from triangulation, the navigation technique. One bearing on a landmark gives you a line. Two bearings give you a point. Three or more, all meeting at the same spot, give you a fix.

Each spread is a bearing. It’s casting a vote: “ticker A looks rich relative to ticker B.”

One spread voting that XOM is expensive is useful but limited information. Maybe SLB got cheaper, not XOM got richer. If three spreads all vote that XOM is expensive, then things are clearer. But do we actually have a “fix?”

In navigation, a fix is definitive. In trading, it’s not that clean. It tells you where the network of overlapping spreads is pointing, but it doesn’t tell you what you’ll find when you get there.

But this is still very useful information. It narrows the question from “something in this pair is mispriced” to “this specific thing looks mispriced relative to its peers.” That’s a big step.

Of course, stocks don’t only move on mean-reverting noise trading. Sometimes the network points at a ticker that moved on real news. A fix tells you what looks mispriced. But it doesn’t tell you why it moved.

We’ll come back to that. First, the bigger picture.

The XOM example is obvious because all the spreads point the same way. But XOM is just one ticker in a small network.

Scale this up. A real universe might have 40 or 60 or 100 stocks connected by heaps of overlapping spreads. Apply triangulation across that whole network and you get something more powerful than identifying a single outlier: you get a portfolio of (potentially) mispriced legs.

Your long book is full of stuff the network says is cheap relative to its peers. Your short book is full of stuff that it says is rich. Every position is targeting an apparent mispricing.

Compare that to traditional pairs, where half your positions are fair-value hedges just along for the ride. (We covered this problem in Part 3.)

Instead of finding one mispriced stock, we build a long/short book where every position has a reason to be there.

This next insight surprised me when I first started researching this. You get a lot of the benefit just from the flattening step we covered last time.

The simple act of aggregating votes from multiple spreads naturally tends to build this kind of portfolio of mispriced stocks.

Stuff that’s genuinely mispriced shows up as mispriced against multiple partners. Noise tends to wash out. You end up long the things that are cheap across multiple relationships and short the things that are rich.

That simple reframe does a lot of heavy lifting. Spreads as evidence, not trades.

Remember the caveat from earlier: not every dislocation is a genuine mispricing.

Some of those tickers moved on real news, not noise.

But if you’ve done a good job of picking pairs, you’ll find that price-insensitive flows drive divergences a lot more often than genuine repricings do. That’s the base rate, and it’s working in your favour. Flattening just scales it up and gets the law of large numbers on your side faster.

But you can do better than the base rate alone. You can sharpen which tickers the network is most confident about.

When you flatten, each ticker accumulates votes from multiple spreads. Some votes say “rich,” others say “cheap.” How much the votes agree is useful information.

Consistency measures this directly: it’s just the absolute value of the mean of the signs. Convert each signal to +1 or -1, average them, take the absolute value. Nothing fancy.

This approach will penalise disagreements quite harshly:

• If all votes agree, the consistency metric comes out at one.

• Five out of six: 0.67

• Four out of six: 0.33.

• A 50-50 split results in zero.

It’s not a perfect metric. For example, a ticker that only appears in two spreads will always get either zero or one. So you need to think about how you handle that.

But on the whole, it’s a nice, simple way to weight the flattened signals to take advantage of more information.

High consistency means the network is pointing clearly at this ticker. Low consistency means conflicting information. More likely noise. Discount it.

This matters because a ticker with a strong average signal but low consistency is probably just the partner of something that actually moved. The spreads disagree about this ticker, which means it’s likely not the mispriced one. A ticker with a moderate signal but high consistency implies that every spread agrees. That’s a cleaner signal.

When I tested this across different universe sizes and time periods, consistency-weighted signals outperformed simple averages.

I reckon this is one of those cases where the obvious thing to try actually works.

So the network agrees: XOM is the outlier. Three spreads, same conclusion.

But there’s a question we haven’t answered…

Why is XOM the outlier?

Maybe XOM got dislocated by some temporary flow. A big fund rebalancing its energy exposure, or a hedging flow working through the book. Price-insensitive activity that pushed the price without any change in fundamentals. That dislocation should wash out. That’s our opportunity.

Or maybe XOM just repriced on real news. An earnings surprise, a production update, a regulatory shift specifically impacting XOM.

In that case, XOM isn’t mispriced at all. The fix is pointing you at a stock that moved for a good reason.

Both scenarios look identical in terms of z-scores. Triangulation tells you what the network thinks is mispriced. But it can’t tell you whether the network is right.

If you pick good pairs, pairs that tend to diverge and converge in tradeable ways, you’ll be right more often than you’re wrong. Price-insensitive flows happen more frequently than genuine repricings.

In that case, the base rate works in your favour (but you really do need to pick good pairs!!). You could accept that you’ll be wrong sometimes, and that’s just a cost of doing business.

Or you could try to do better…

It turns out that there are things outside spread relationships that indicate informed buying or selling.

The usual suspects: earnings surprises, idiosyncratic news.

Volume data around a dislocation is also telling.

The pattern of trading activity often looks different when a stock is getting pushed around by price-insensitive flows versus when it’s repricing on information. And when you go digging into the effects, you find they work differently on the long side than the short side. Which, when you stop and think about it, makes a lot of sense.

This sort of information can help you side-step those annoying occasions when your spread diverges, but fails to come back together again. It’s not fool-proof of course, but a good implementation can really improve things.

I explore this in detail with my RW Pro group, where we’ve built these features into our live stat arb implementation.

One more cool thing I want to show you.

Consider that in the triangulated stat arb framing, we can treat each spread’s z-score as the difference in ticker mispricings, or “alpha:”

We observe the left-hand side (our spread z-scores), and we want to infer the right-hand side (ticker alphas).

We can run a regression to find the ticker alphas that best explain the spread z-scores simultaneously.

This gives us a “regression factor” that I can add to our growing collection:

A standard regression minimises the sum of squared alphas, which will tend to spread the signal across all the tickers. But perhaps a better solution would be one that concentrates the signal in as few tickers as necessary.

This implies that lasso regression (which forces coefficients to zero) or ridge (which shrinks them towards zero) might be decent approaches.

Adding a “ridge regression factor” to the list:

Here’s a plot of cumulative returns to each of these network-derived factors for a rolling universe of our top 50 pairs. This isn’t a backtest… There are no costs or frictions. The intent is to get a high-level comparison of the different approaches:

None of these factors is necessarily “better” (at least, we can’t really tell from what I’ve shown you here). But they do come with different trade-offs. Some tend to have higher turnover. Others tend to decay faster. Some will push you into more concentrated positions if you’re not careful.

There are also interesting dynamics when you sort your tickers by the “quality” of the spreads they came from. I haven’t really explored that with you, but it’s yet another dimension.

Anyway, what I’ve presented here isn’t an exhaustive examination. My objective was to demonstrate the step changes from traditional pairs trading to a triangulated stat arb approach, both in performance and complexity.

That’s the real takeaway. You get a significant boost in risk-adjusted returns over traditional pairs trading with the triangulated stat arb approach, but you pay a cost in complexity.

The other factors and modifications are great, but you get the bulk of the benefit just from reframing the problem from pairs to portfolio (in the plot above, the jump from pairs to flattening is greater than the jump to any of the “better” flattened signals).

Let’s step back and look at what we’ve covered across this series. The problem breaks down into three pieces.

Find good pairs. Pick spreads that are likely to diverge and converge in tradeable ways. That’s the selection pipeline we looked at in Parts 2 and 3. Measuring what you actually care about, directly.

This is the most important part. Without this, the rest is irrelevant.

Identify the mispriced legs. Figure out what’s actually mispriced relative to its peers. That’s the triangulated stat arb part: flatten spread signals to ticker-level signals, use consistency to measure the network’s agreement, build a portfolio where every position is targeting a mispricing.

This does a lot of heavy lifting. The pairs to portfolio change offers a step change in performance and capital efficiency, but also complexity.

Separate mispricing from repricing. Try to predict when an apparent mispricing is actually a genuine repricing. That’s where volume and news features come in. It’s a bit harder than it sounds and to be frank, the juice may not be worth the squeeze. Your mileage may vary.

This is the icing on the cake. Trading a portfolio of apparent mispricings is a fine approach, so long as your pair selection approach is in order. You’ll be wrong plenty, but you should be right more than you’re wrong. But if you can avoid some of those repricings, you’ll see another step change in performance.

Each piece matters, but to a different extent.

Good pairs are the foundation; you’re sunk without them. Triangulation and consistency squeeze more out of those pairs than traditional approaches can. Volume and news features help you avoid the traps.

Getting all three right is what makes this approach really sing. It’s also what makes it hard. There are a lot of moving parts, and each one has subtleties that aren’t obvious until you start trading it.

Next time, I’ll show you what it looks like when you resource all three pieces properly.

If you’re finding this useful, check out my case study on how I went from corporate quant to independent trader. Click here to grab a copy.

Previously:

• A Tale of Two Prices (the core idea of stat arb)

• Moneyball (finding good pairs using metrics that matter)

• The Winter of our Pairs Trading Discontent (problems, limitations, frustrations)

Pairs trading remains a feasible approach for the indie trader. But, as we saw last time, there are inherent limitations. Trading both legs eats a lot of buying power and limits the number of pairs you can trade.

Trading only the mispriced leg helps, but introduces a ton of variance. Essentially, the trade-off is accepting a wilder ride in exchange for higher expected returns and better capital efficiency.

In the last article, I alluded to an approach that might offer a third route: we give up a little simplicity, but in exchange, we get higher expected returns and some variance control.

Here’s an example based on a small universe of spreads constructed from energy stocks. Six stocks, eight spreads.

Here are the z-scores of these spreads:

Every spread involving XOM is stretched. The spreads not involving XOM are all near zero. SLB, CVX, OXY, EOG, COP trade fairly against each other. XOM is the outlier.

The network of overlapping spreads tells us something specific: XOM is the mispriced security.

In Part 3, we talked about the constraints of traditional pairs trading. You trade spreads because it gives you built-in variance reduction. Long one leg, short the other, and you’re insulated from broad market moves. That’s valuable.

But you pay for it. Half your capital sits on a leg that might be fairly priced. You pay double the transaction costs. And when your selection pipeline identifies 100 good pairs but you can only trade 10, most of your signal goes unused.

Now, you could try to trade just the mispriced leg. With a handful of pairs, you might even identify it by observing which stock moved. If XOM rallied 10% while SLB was flat, XOM probably drove the divergence.

But that’s not foolproof. Sometimes stocks move on beta, and the one left behind is actually the mispriced leg. “Which one moved” is useful information, but it’s not the whole picture.

One thing you can do is directly model market- and even sector-beta in your spread construction.

That’s a useful thing to do on average. If you can build a better model, then you should, all else being equal.

But that’s a big “if.”

Betas aren’t stable, and they’re prone to measurement error:

It’s difficult to model this stuff well in a way that’s of practical benefit to the trader.

But there’s another source of information here that’s hiding in plain sight, requires little modelling skill to extract, and is relevant right at the time you measure it.

I’m referring to the network of interconnected tickers that arises from a collection of overlapping spreads:

An edge between two ticker nodes indicates that those two tickers are connected in a spread.

We can visualise the state of the network at any point in time by incorporating the z-scores of each spread:

• Arrows indicate the rich ticker in each pair (z-score positive)

• The thickness of each connection represents the magnitude of the z-score

  • Thin shows a z-score close to zero

  • Thick shows a z-score that’s stretched

• The colour of each ticker node represents that ticker’s aggregated z-score from all of its connections (orange = positive, purple = negative).

You can see how the state of the network evolves through time as the spreads diverge and converge:

The network gives you more information than any single spread can. When XOM is stretched against everyone, you have powerful collective evidence.

I call this Triangulated Stat Arb (not to be confused with triangular arbitrage, which is a different thing entirely.) We’ll get into exactly why that name makes sense in the next article.

For now, the key insight is this: instead of trading spreads, you trade the securities themselves. And instead of choosing between conviction and variance control, you get both. But you lose some simplicity.

Traditional pairs trading operates at the spread level. You see a divergent spread, you trade both legs, you wait for convergence.

The shift I’m describing operates at the ticker level. You use all those spreads as information about individual securities. Then you trade the securities directly.

The mechanics are straightforward. I call it “flattening” because you’re collapsing spread-level signals down to ticker-level signals.

Here’s how flattening works:

Take any spread signal at some point in time and decompose it into two ticker signals.

Say you’ve got the XOM-SLB spread with zscore +2.3. That means XOM is rich relative to SLB. So:

• XOM gets +2.3 (appears expensive)

• SLB gets -2.3 (appears cheap)

Same magnitude, opposite signs. One spread becomes two votes.

Do this for every spread in your universe at the current point in time. Now you’ve got a bunch of votes for each ticker. Some tickers appear in many spreads, some in few. Each spread casts a vote on both its legs.

Now aggregate. XOM appears in three spreads, all saying it’s rich. Average those signals. You get a strong positive aggregate. Strong positive means “this looks expensive.”

SLB appears in three spreads as well. One says it’s cheap (relative to XOM), the others say it’s roughly fair. The SLB signals disagree - weaker conviction. It’s probably not the mispriced leg in its one divergent spread.

This is the power of triangulated stat arb. Stuff that’s genuinely mispriced shows up as mispriced against multiple partners. Noise tends to wash out, while the signal gets reinforced.

With ticker-level signals, you build a long/short portfolio.

Tickers with strong negative signals (cheap against multiple partners) go in the long book. Tickers with strong positive signals (rich against multiple partners) go in the short book.

The key insight is that you’re not building a portfolio of “one mispriced leg plus one fair-value leg” repeated many times, as in a pair trade:

Instead, you’re building a portfolio of mispriced tickers. The longs are cheap. The shorts are rich.

Both sides are working for you.

There are many ways you can construct a portfolio with this insight.

You’ll find that the effect tends to concentrate in the extremes, but there can be value in trading the broader distribution. It really depends on what else you’re doing and what you want to achieve.

If you want, you can size it so your net market exposure is roughly zero. This will give you similar variance reduction to pairs trading, but through portfolio construction rather than pair-by-pair hedging.

I’ve been letting the portfolio drift net long or net short, depending on the relative magnitude of the signals in the long and short books. This tends to juice returns at the expense of some higher variance.

There are many possible implementations, including weighting by the quality of the spreads the ticker-level signals come from, and accounting for how many times a ticker appears in the network (more votes lead to more conviction). More on these ideas later.

Subscribe now

But essentially, this is the “have your cake and eat it too” part. You’re targeting the mispriced legs directly. But because you’re long a basket of cheap tickers and short a basket of rich tickers, you still get the low-variance benefits of a hedged approach.

This approach - triangulated stat arb - addresses the problems we identified in Part 3.

Capital efficiency. You’re deploying capital on legs that are actually mispriced. Not tying up half your position on fair-value hedges.

Signal utilisation. Remember the 100-pairs problem? Your selection pipeline identifies heaps of good pairs, but you could only trade 10 due to capital inefficiency. Now you’re aggregating information from your entire universe. More of your research actually gets used.

Lower transaction costs. You only pay for the legs that are actually mispriced.

Diversification. You’re holding positions in individual securities, not paired positions. More natural diversification (that you can dial up or down to fit your objectives).

Perhaps most exciting is that the portfolio approach can be incorporated into a broader multi-factor long/short portfolio - the aggregated pairs signals can be viewed as just another factor.

There’s something I haven’t fully addressed.

In our network of energy stock spreads, the case for XOM being mispriced was compelling. We had a lot of overlapping spreads all pointing in the same direction.

Often in practice, it won’t be that clean.

We might have sparser areas of the network with fewer connections. Stocks moving on noisy betas to their sector and the broader market can cloud things as well.

Figuring out which one moved helps, but it’s not definitive. What else can we use?

When do multiple spreads agreeing give you genuine conviction versus just coincidence? How do you measure the network’s agreement? How do you know when you’ve got a clearer signal versus noise?

That’s where triangulated stat arb really earns its name. The navigation metaphor provides a nice, simple mechanism for assessing the quality of an aggregated signal.

In the next article, I’ll show you how it works.

If you’re finding this useful, I share more practical thinking like this in my case study on how I went from corporate quant to independent trader. Grab it here.

Previously:

• A Tale of Two Prices (the core idea of stat arb)

• Moneyball (finding good pairs using metrics that matter)

In the last article, we built up a conceptual understanding of universe selection: how to find pairs that diverge and converge in a tradeable way. We talked about measuring the thing you actually care about directly, rather than reaching for statistical tests like cointegration that sound perfect but turn out to be unstable in practice.

The natural next step is to start trading them the traditional way. Long the cheap leg, short the expensive leg, wait for convergence, close.

I’ve been doing this for a few years now. And… it works. We’ve made money from it. But the simplicity comes at a cost. And the more I trade it this way, the more apparent it becomes that it leaves a lot on the table.

Let me describe what traditional pairs trading looks like in practice for the solo trader.

You’ve got your pair selection pipeline. It identifies, say, 100 pairs that score well on your metrics. Strong mean reversion characteristics, prices stay close together, same industry, liquid enough that you can trade them fairly easily, all the things you want.

But you can only trade a fraction of them.

In a regular portfolio margin account, a common structure for indie traders, pairs trading eats a lot of your buying power.

Each pair requires capital on both legs. If you’ve got $100k of buying power and you’re doing 10 pairs with equal weighting, that’s $10k per pair, $5k per leg. You’re basically maxed out.

If you try to trade more pairs by putting less on each leg, you start running into minimum commission fees and other frictions. The costs start to add up.

So you’ve got 100 good pairs. You trade 10 of them. What happens to the other 90?

Nothing.

All that signal, all that information from your really good selection process, essentially just sits there, hiding beneath the surface.

But there’s a deeper problem. It’s not just that you can only trade a fraction of your universe. It’s that the capital you do deploy is being used inefficiently as well. Every time you put on a pair, you’re trading two legs. You’re using capital on both.

But typically, only one of those legs is actually mispriced.

Imagine you’ve got two oil stocks. One of them just had a random burst of selling. Some fund was liquidating a position, pushing the price down temporarily, so that the stock becomes cheap relative to the other.

That’s a pairs trade opportunity. You go long the cheap one, short the expensive one.

But look at what happened:

The cheap stock is mispriced. The other one was just sitting there at fair value the whole time. You traded both, but only one was actually doing the “work” of capturing the mispricing.

You’ve deployed capital on both legs, but the expected return is only coming from one of them.

That’s what I mean by capital inefficiency. You’re paying for two legs, but only getting the juice from one.

It gets worse when you think about transaction costs.

Every time you enter a pair, you cross the spread twice. You pay commissions twice. Every time you exit, same thing.

For a single directional trade, you’d pay one spread and one commission on entry, one on exit. For a pairs trade, you’re paying double.

And remember, the expected return is only coming from one leg. So your costs have doubled, but your expected return hasn’t.

Let’s go back to the 100 pairs problem.

Say your pipeline identifies 100 pairs that look really good. Strong mean reversion, low distance metric, same industry, all the criteria we discussed last time.

If you could trade all 100 pairs, you’d have much more diversification. You’d be able to let the law of large numbers work for you. Individual pairs might go against you, but with enough of them, the statistics work out.

But you can’t trade 100 pairs. Capital constraints force you down to 10 or 20.

And here’s a subtle problem: those 100 pairs don’t involve 200 unique stocks. There’s a lot of overlap. Stock A might appear in five different pairs. Stock B might appear in eight.

All those pairs are telling you something about those individual stocks. Maybe Stock A appears as the cheap leg in five different pairs. That’s a pretty strong signal that Stock A is undervalued relative to its peers.

But in the traditional pairs approach, you can’t aggregate that information. You’re stuck trading pairs as isolated units. You can’t say, “Stock A looks cheap across lots of pairs, so I’ll just be long Stock A.”

Well, you could. But then you’d introduce a whole lot of variance, because you’d no longer be market neutral.

And that brings us to an important source of tension.

One reason people trade pairs in the first place is to be market neutral. You’re long one leg, short the other. General market moves don’t affect you (in theory). You’re just capturing the convergence.

If you only traded the mispriced leg, you’d have higher expected returns (lower costs), but much higher variance. You’re now net long or net short the market.

If the market moves against you, you could lose money even when you were right about the convergence:

Equally, sometimes you’ll get lucky and make more than you expected to:

Over enough trades, you’d expect these to net off. But your P&L will be a lot more variable.

So you’ve got a tradeoff:

• Traditional pairs: Lower variance (market neutral), but higher costs and capital inefficiency.

• Just the mispriced leg: Higher expected returns (lower costs), but much higher variance.

How do you resolve this?

Well, you could just accept it and do the version that represents the best path through the trade-offs for you.

But we can do a little better…

More accurately, there might be another path through the trade-offs.

Perhaps we can trade some simplicity for the chance to have our delicious higher expected return cake while eating less variance.

The high-frequency people do it by trading very fast. They’re not market neutral on each trade, but they turn over quickly enough that market moves net out before they notice the variance. The law of large numbers works in their favour through speed.

But that’s not realistic for most of us. We don’t have the infrastructure, the co-location, the data feeds. And even if we did, the edges at that speed are tiny and competitive.

For indie traders with slower strategies, we need a different approach.

The alternative is to go broad.

Instead of trading 10 pairs in isolation, you trade lots of individual positions, constructed intelligently so that market exposures tend to cancel out at the portfolio level.

You flatten your pair signals into signals on individual assets. You aggregate the information from many pairs. And you construct a portfolio that captures the juice while managing the risk.

And at the point you go from individual pairs to a portfolio, you open up the opportunity to add other signals to the mix, implement different risk models, and manage transaction costs more subtly.

If you’re really enthusiastic, you even infer things about a given stock from things that are happening throughout your broader network of flattened pairs.

This is what we might call a more “modern” statistical arbitrage approach. The general approach has been around for decades. And increasingly, it’s accessible to independent traders too.

Of course, it’s no free lunch.

You go from a super simple set of pairs that you could manage with a mouse and a TradingView subscription to something that you can easily get lost in if you’re not careful.

So it’s not something I’d recommend to every indie trader.

When trading is your part-time gig, there is immense value in simplicity. But if you’re a bit more ambitious about your trading, it’s something you might consider.

But it requires a shift in how you think about the problem. From pairs to portfolio. From isolated trades to system.

Remember in Part 1 when I said the divergence/convergence concept is simple? That core bet doesn’t change. What changes is how you express it.

Next time, we’ll dig into how that portfolio approach works, and why it solves (or at least reduces) the problems of higher variance, and signal and capital inefficiency.

Building the kind of infrastructure that supports a modern start arb approach yourself requires serious data engineering. Inside RW Pro, we’ve done the heavy lifting. But more on that another time.

If you’re finding this useful, I share more of this kind of practical thinking about systematic trading in my case study on how I went from corporate quant to independent trader.

Previously: A Tale of Two Prices (the core idea of stat arb)

Last time we established that stat arb is really about betting on divergence/convergence behaviour continuing. Two things that have historically moved together come apart, and you bet on them coming back together.

Remember the forced flows example, some fund or whatever having to sell regardless of price? That sort of temporary dislocation creates opportunities.

Conceptually simple. But the question is: how do you find pairs that actually do this reliably?

This is where I went wrong for a long time. I spent way too long in the early days getting hung up on the wrong things.

If you’ve read anything about pairs trading, you’ve probably come across discussions of Kalman filters, Ornstein-Uhlenbeck models, optimal hedge ratios.

I got really hung up on that stuff when I first started. I thought the algorithm details were the important part.

They’re not.

At the end of the day, all of that stuff is a minor implementation detail. Really, 80 to 90 per cent of the whole pairs trading game is just finding stuff that tends to drift apart and come back together.

Unsexy, but true.

If you can find two things that reliably diverge and converge, my six-year-old could trade them and make money. Seriously. The trading logic is the easy part. Long the cheap one, short the expensive one, wait for convergence. Done.

Of course, nothing is that simple. The hard part is finding those things in the first place. And then managing them with all the trade-offs, constraints and limitations that you inevitably run up against (that’s coming in article 3, by the way).

Let me start with what doesn’t work as well as you’d hope: cointegration testing.

Cointegration is one of those things that sounds perfect for pairs trading. It’s a statistical property that says two time series tend to move together over time, even if they wander individually. Exactly what we want, right?

The problem is practical, not theoretical.

If you try to measure a cointegration coefficient over a rolling window, you’ll find it’s incredibly unstable. The relationship looks solid and sensible when you look at the full sample. But in real time, the coefficient swings around wildly. It changes sign. It breaks down completely for long stretches.

Same story with correlation, by the way. High historical correlation doesn’t tell you much about whether the spread will mean-revert going forward.

The fundamental problem with these statistical tests is that they’re measuring the wrong thing. They’re measuring statistical properties of the relationship, not the thing we actually care about: does this spread tend to diverge and converge in a tradeable way?

So what should you look for instead?

The answer is to measure the thing you actually care about as directly as possible.

You care about two things:

1. Do these things stay close together? If they tend to drift apart permanently, you don’t have a tradable pair.

2. When they do come apart, do they come back together? This is the mean reversion behaviour you’re trying to capture.

Rather than trying to infer these properties from statistical tests, you can measure them directly.

There are various ways to do this.

One approach is to look at what we might call the “distance” between two normalised price series over a period. If the distance is consistently low, the pair stays close together. That’s good.

Another approach is to look at a kind of simulated frictionless return from trading mean reversion on the spread. If this historical return is high, the pair has exhibited strong mean reversion. Also good.

Neither of these on their own is perfect. But when you combine them, you start to filter out a decent chunk of the noise.

For example, a pair like this will score a high total return on the mean reversion metric (makes a high return due to one convergence event), but score poorly on the distance feature (prices drifted rather than stayed close together):

A pair that scores well on both metrics, that stays close together, and has strong mean reversion, that’s the kind of pair you want.

And critically, these metrics tend to persist. Pairs that scored well in one period tend to score well in the next period, too. That persistence is the whole basis for using them as selection criteria.

Without persistence, we’d just be fitting to a past with no insight as to whether it should repeat.

Related point: Never assume persistence… always go look in the data and see if it actually exists.

Another ingredient: don’t bother with pairs that don’t make sense.

Similar industry or sector is a good baseline. Two oil stocks are exposed to similar underlying factors. Two tech companies, two banks, two retailers. They have similar risk exposures.

When one temporarily gets pushed out of line relative to the other, and it’s not for fundamental reasons, you’d expect them to converge. They’re exposed to the same stuff.

If you try to pair random stocks together just because they look statistically related, you’re asking for trouble. Spurious correlations break down. Pairs with no fundamental connection can diverge permanently when something changes for one of them.

So the process looks something like this:

1. Start with pairs that make structural sense (same industry)

2. Filter for pairs that have stayed close together historically

3. Filter for pairs that have shown strong mean reversion

4. Rank by some combination of those metrics

5. Take the top N

That’s really the skeleton of it.

The most important thing I can tell you about universe selection is this: focus on finding pairs whose natural divergence and convergence makes sense.

Don’t focus on finding pairs that pass some statistical test. Don’t focus on proving that a relationship is “statistically significant.” Don’t get lost in the academic weeds.

Instead, look as directly as possible at the thing you care about. Does this spread drift apart and come back together? Measure that. Use metrics that capture that behaviour. Then test whether pairs that score well on those metrics continue to score well going forward.

If they do, you’ve got something. If they don’t, the metric isn’t useful, no matter how statistically sophisticated it is.

Now, there’s a catch (there’s always a catch).

Even when you’ve got a really good selection process, even when you’ve identified hundreds of pairs that look genuinely good, you still have to trade them.

And that’s where the traditional pairs trading approach hits some painful walls. Capital constraints. Transaction costs. The fact that you can only practically trade maybe 10-20 pairs at a time, depending on your setup.

Your selection pipeline might identify a couple of hundred great pairs. But if you can only trade 10 of them, what happens to the rest? All that signal, all that information, essentially gets thrown away.

That makes me sad.

This is the limitation of traditional pairs trading for the indie trader. And it’s something that’s been frustrating me for a long time… ever since I left the prop world, now that I think about it.

Next time, we’ll dig into exactly what those limitations look like in practice, and why they matter more than you might think. Then, I’ll show you how I get around them.

I’ve seen heaps of stuff published online about stat arb lately. Some genuinely good takes. And some other material that, while academically interesting, isn’t particularly useful for people like me and the people I write for: independent systematic traders looking for real edges we can realistically manage without a team.

A lot of the stat arb “literature” makes this whole thing sound more complicated than it needs to be. Cointegration tests, optimal hedge ratios, Ornstein-Uhlenbeck processes.

Best case, some of that stuff might be useful if you’re running a big operation and tinkering at the margins is justified.

But for most of us, it just obscures the fundamentals.

And the fundamentals of statistical arbitrage are actually pretty simple. Understanding those matters more than the fancy maths.

So let’s begin from first principles.

Actual Arbitrage

To understand statistical arbitrage, you need to understand regular arbitrage first.

Imagine you’ve got a crypto meme coin trading on two exchanges. On Exchange A, you can buy it at 101 and sell at 99. On Exchange B, you can buy at 103 and sell at 102.

Same asset. Different prices.

There’s an obvious trade. Buy at 101 on Exchange A, transfer it over to Exchange B, sell at 102.

Make a dollar, risk-free (allegedly).

This is pure arbitrage. You’re buying and selling the same thing at different prices. Theoretically riskless.

Except it’s not really riskless in practice.

The asset transfer costs money and takes time. And while you’re waiting for the transfer to complete, the market moves. By the time your crypto lands on Exchange B, the price might have changed.

So pure arbitrage is theoretically riskless, but practically hard to capture.

And because it’s such an attractive trade, it’s extremely rare. And if you happen to find a persistent one, it’s only a matter of time before others notice, pile in, and before you know it, it’s gone.

But I digress. The thing to know is that every good trading idea basically looks like this. You’re identifying when something is cheap or expensive relative to where it should be, and you’re taking advantage of that difference.

That’s trading in a nutshell.

Enter the Pairs Trade

So if pure arbitrage is hard to find and hard to execute, how else might we trade this general idea?

What if, instead of buying on Exchange A, transferring the asset, and selling on Exchange B, we just bought on A and sold on B at the same time?

No transfer. We’re just long on one exchange and short on the other, simultaneously.

Now we’re not doing arbitrage anymore. We’re doing a pairs trade.

The key difference is that in pure arbitrage, you realise your profit immediately when you complete the transfer. In a pairs trade, you’ve got to wait for the prices to converge.

You bought the cheap one, you sold the expensive one. But you haven’t made any money yet. You need those prices to come back into line.

This is a really important point, and it sounds obvious, but it’s easy to forget. The pairs trade is a bet on convergence. You’re betting that the spread between the two instruments will narrow.

Aside: You can use this to filter the useful-to-the-indie-trader stat arb articles I mentioned at the start of this piece… if the article doesn’t focus on this core divergence-convergence idea (how to find it, how to measure it, how to understand it, how to trade it, etc.) then you can safely move on.

Why “Statistical” Arbitrage?

So why do we call it statistical arbitrage rather than just arbitrage?

When I first heard about this idea, I assumed the name was because it’s about using statistics to find arbitrage relationships or something like that.

But that would be entirely wrong.

It’s because there’s no guarantee the spread converges.

In pure arbitrage, convergence is guaranteed (the same asset has to trade at the same price eventually). In stat arb, convergence is probabilistic. It’s not certain.

It’s a bet on expected convergence, not guaranteed price changes.

The spread could widen instead of narrowing. One exchange could have problems. The relationship could break down entirely.

So you’re making a statistical bet. On average, over many trades, you expect this behaviour to work out in your favour. But any individual trade could lose.

This is important to understand. A lot of material out there gives you the impression that stat arb is about finding cointegrated pairs or getting hedge ratios right. That stuff obscures the fundamental thing we actually care about: prices drifting apart and coming back together again.

It’s really no more complicated than that.

But equally, don’t mistake simple for easy.

Stat arb is conceptually very simple. But it’s not an easy trade to do well (is anything?).

And if you complicate the simple part (the core idea), you really lower your chances of doing the not-simple part successfully… if at all.

This is an underappreciated point.

It sounds like a throw-away line that you might be able to get away with ignoring:

“Focus on the core idea of divergence-convergence.”

But it takes a surprising amount of care, nuance, and, dare I say, skill to actually implement a portfolio around this core idea. Especially at any sort of scale.

If you focus on the wrong stuff at the front end (cointegration, I am looking at you), you waste precious time and energy that could be spent figuring out what actually moves the dial. Trust me, you can’t afford that wasted time.

In one of the later articles in this series (article 3 if all goes to plan… no promises), I’ll talk about a simple implementation that an indie trader can realistically do. Its simplicity comes with certain trade-offs, but it’s a decent place to start.

What Drives Divergence and Convergence?

Okay, so we’re betting on prices diverging and converging. But why would this happen reliably?

Let’s think about what drives divergence first.

Most of the time, when one leg of a pair gets pushed out of line, it’s because of some temporary imbalance in supply and demand. Someone needed to buy or sell for reasons that have nothing to do with the fundamental value of the asset.

Index rebalancing flows. Hedging needs. Liquidity requirements. Some bloke at a fund got a redemption and had to sell, regardless of price.

These are what smart people call “forced flows” or “price-insensitive” participants. They’re not trading because they think the asset is mispriced. They’re trading because they have to.

And this creates temporary mispricings.

Now, why would convergence happen?

If the assets are exposed to similar risk factors, and one gets pushed out of line by someone trading because they have to, it should come back. The mispricing isn’t fundamental. It’s noise. And noise tends to mean-revert. People will notice that it’s mispriced and take advantage of that until it’s no longer mispriced.

That’s the core bet. There’s nothing magical (or particularly statistical) about it. Just: things that should trade together got temporarily pushed apart, and you expect them to come back together.

Simple Concept, Less Simple Implementation

The biggest thing I want you to take away from this is that stat arb, as a concept, is simpler than people make it sound.

You identify things that have historically moved together but also tend to come apart and come back together. You trade deviations from that relationship.

When they diverge, you bet on convergence. Essentially, it’s a bet on the divergence/convergence behaviour continuing.

Now, that’s conceptually simple.

But as soon as you start thinking about the details, things get complicated quickly.

The most obvious question is, how do you find pairs that exhibit this behaviour reliably?

And when you start doing the trade, you find that spreads diverge for fundamental reasons, too. Assets reprice on genuine changes in their outlooks, and that sort of divergence doesn’t converge.

So how do you deal with that?

These are examples of details that are genuinely worth your time trying to figure out. They relate directly to the core effect we’re looking for.

That’s where I went wrong when I first started doing this.

I reached for statistical tools (cointegration, correlation, whatever) without thinking about what actually drives the relationship. Spoiler: those statistical tests aren’t as useful as you’d hope (estimation error, non-stationarity, the usual suspects).

The reality is: if you can find two things that tend to move apart and come back together reliably, you could ask my six-year-old son to trade them, and he’d make money.

The trading logic is the easy part. If only it were all that easy! The hard part is finding those things in the first place.

So next time, we’ll dig into that. How do you actually find pairs and divergences worth trading? What should you look for, and what should you avoid?

I’m also planning to write about the stat arb journey we’ve been on in the RW Pro community, where we took a solid pairs trade and evolved it into a more scalable and performant stat arb portfolio. Lots of interesting lessons to share.

If this kind of practical thinking about systematic trading resonates with you, I share more of it in my case study on how I went from bored engineer to corporate quant to independent trader, including all the mistakes I made along the way. Grab it here. And don’t forget to subscribe!

Subscribe now

If you had a diversified portfolio of risk premia, you probably did alright. Claiming we did anything overly special in such a favourable environment would be a tad arrogant.

That said, there’s a big difference between “the market was kind” and “I was prepared to capture what the market offered.”

So here’s a look back at what actually happened, what we learned, and one mistake I made that I really should have known better than to make.

The RW Team Evolved

We said farewell to James this year. He was always more a partner than an employee, and his contribution really helped lay the foundation for what Robot Wealth is today. He’s doing interesting work in the crypto space that’s very much worth watching, but he’ll always be part of the RW family.

With change comes opportunity. We welcomed Euan Sinclair to the team (if you don’t know Euan, he has a ton of systematic trading experience, particularly in options), and hired Ryan, a brilliant developer, from within our own community. Hiring from the community is such a bonus, honestly.

What Made Money

Here’s an uncomfortable truth that I’ll share anyway: you didn’t need to do anything overly special to make decent money this year. Diversified risk premia worked well across the board.

Stocks went up pretty much everywhere.

Bonds were a mixed bag. Long-dated treasuries were up just a tad, but exposure to the shorter end made money. Emerging market debt did well.

Commercial credit did alright.

And if you had gold in your risk premia portfolio, buy yourself a nice bottle of something special to celebrate. (Opinions are divided on whether it carries a risk premium, but there’s also a decent case for holding some as a diversifier.)

Volatility selling realised some of that famous negative skew during the April tariff tantrum, but was still profitable with even the most basic approach. Being a bit clever with the implementation (for example, selling the front of the VX futures curve and buying the back) was a good approach this year.

Systematic crypto edges continued to perform well, despite the big projects being down year on year. It feels like there are still good opportunities in crypto, but it did feel a bit harder this year. Maybe I’m just getting old.

Simple seasonality edges did well, such as end-of-month rebalance flow trades. Even simple convergence trades in equities worked really well.

I’m not going to pretend there was a lot of skill involved this year. A lot of it was just having fairly mundane exposures in place when the market decided to be generous.

But that’s the whole point of systematic trading: you don’t need to be a genius. You need to be positioned to capture what the market offers when it offers it.

Which brings me to the biggest lesson of 2025.

One Glaring Lesson (Or: How I Outsmarted Myself)

I really like simple seasonality edges. They’re easy to understand, easy to trade, and often surprisingly effective.

There’s a well-known calendar effect around turn-of-month flows in bonds.

Pension funds, insurance companies, and other institutional investors tend to rebalance at predictable times. And bonds may tend to be preferred towards the end of the month for “portfolio window dressing” reasons.

This creates temporary supply/demand imbalances that a patient systematic trader can exploit.

I’ve traded this effect for years. Buy bonds a few days before month-end, short at the close of the month, cover a few days after. Simple.

Historically, it’s been a Sharpe 1-ish strategy. Not spectacular, but solid. And importantly, it’s only in the market about half the time.

But starting around 2023, it had a rough stretch. The edge seemed to disappear.

So I did what any thoughtful, analytical trader would do…

I started telling myself stories.

Maybe market structure had changed. Maybe the participants driving the flow had shifted their behaviour. Maybe the trade had become too crowded.

All plausible explanations. All completely reasonable hypotheses.

And based on these stories, I stopped trading it.

Big mistake.

When I looked at the results for 2025, the turn-of-month bond trade would have returned about 15% at a Sharpe of 1.8 after costs.

That’s an excellent result for an insultingly simple strategy that’s only in the market half the time.

The Psychology of Abandoning a Sharpe 1 Strategy

Let me walk you through the maths of what I should have expected.

A Sharpe ratio of 1 means your expected annual return equals your annual volatility. If you’re targeting 10% returns, you should expect 10% volatility. That’s roughly a 16% chance of being down in any given year, assuming normal returns (which is generous).

So a Sharpe 1 strategy can easily be underwater for two years or more without anything being “broken.” That’s just what noise looks like at this Sharpe level.

Here’s the rolling 12-month Sharpe of the strategy back to the mid-90s:

When I look at that rolling Sharpe chart now, 2023-2024 looks bad, but not unprecedented. It’s within the historical range of variation.

So why did I stop trading it?

I think it’s because the stories I told myself were plausible. If I’d had no explanation for the underperformance, I might have been more patient. But because I could construct a reasonable narrative about why the edge might have disappeared, I convinced myself that my judgment was adding value.

Turns out I was just pattern-matching on noise and calling it analysis.

The Stories We Tell Ourselves

This is the insidious thing about trading. We’re trained to look for explanations. Find the signal in the noise. Understand why things happen.

Trading as a job self-selects for naturally curious people who see detective work as an exciting challenge. We’re biased towards “figuring stuff out.”

But sometimes, there is no why. Sometimes it’s just variance.

The problem is that “it’s just variance” doesn’t feel satisfying. Our brains crave narrative. We want cause and effect. We want to understand.

So we invent explanations:

• “The market has changed”

• “Too many people are trading this now”

• “The underlying driver has weakened”

• “This only worked in a specific regime”

All of these might be true. Or they might just be stories we tell ourselves to justify abandoning something uncomfortable.

Unless the edge is based on a clear causal mechanism, I have no way to distinguish between “the edge has genuinely disappeared” and “this is just a bad run within normal expectations” in real-time. Neither do you. Neither does anyone.

This end-of-the-month bond trade is a classic example of something we have a plausible explanation for (noisy tendency of portfolio managers to window-dress holdings towards month-end), but no precise causal mechanism.

I don’t know who’s doing the buying and selling. I don’t know when they’re doing it. I don’t even know their true motivations. The reality is that this is a noisy effect resulting from a bunch of different people trading for different reasons, all happening at the margins amid the usual noise.

In that environment, you face a conundrum:

• Decent enough hypothesis for why it works on average

• Supported by data

• Sharpe ~1, so large variance in annual returns

• No observable mechanism to tell us when it stops working

The uncomfortable conclusion: you just can’t know if it’s stopped working from a year or two of lousy P&L.

But what I can do is set expectations in advance. Before I trade something, I should know:

• What’s the historical Sharpe ratio?

• How has it fluctuated in the past?

• How long might something like this be underwater?

• At what point would I genuinely conclude the edge is gone versus just experiencing normal noise?

And then I should stick to that plan, not revise it when the drawdown makes me uncomfortable.

The Discipline to Stick With Noisy Edges

I always urge people to have the right expectations about the edges they can trade as an indie trader.

Understand the underlying drivers of the edge. Be realistic about how much noise there will be in the short term. And be patient. Be disciplined. Turn up, run the process, don’t bullshit yourself.

If you’re really worried, turn it down a bit. Reduce position size. But don’t stop something you expect to go at Sharpe 1 completely just because it’s had a couple of rough years.

This is hard. I’ve been doing this for years, and I still got it wrong. The emotional reality of watching something bleed for 18 months is different from the intellectual understanding that it should bleed for 18 months sometimes.

But that’s precisely why systematic trading works. You’re supposed to follow the process even when it feels uncomfortable.

Especially when it feels uncomfortable.

The Real Takeaway

2025 was a good year. But the lesson isn’t “we made money.”

The lesson is that having a systematic approach, a diversified portfolio of edges, and the discipline to stick with good but underperforming edges puts you in a position to capture what the market offers.

Some years, you get lucky. Some years you don’t. But over time, the process works.

You can only eat what you’re fed. But you have to be sitting at the table when dinner is served.

I’ll be sitting at the table for the bond turn-of-month trade in 2026. Lesson learned.

I’d built what I thought was a clever automated strategy: a breakout system that traded currency pairs based on overnight ranges and longer-term trends.

The details don’t really matter.

What matters is that I’d spent weeks tinkering with the backtest until the equity curve looked beautiful.

Stop losses, take profit levels, trailing stops, time stops, filters of various kinds. I tortured that thing until it confessed to profitability.

Then I went live.

The Lucky Disaster

On the first day, my strategy placed trades while I was away from the computer. When I checked my phone, I nearly had a heart attack. My positions were 10x bigger than I’d intended. I’d messed up my sizing algorithm.

I rushed home to fix things.

And when I got there, I found that I was actually up. Way up.

So what did I do? I thought, “I’m playing with the house’s money now. Let me just ride this out.”

That’s pure gambling irrationality.

Somehow, I didn’t blow up. I closed those oversized positions with a massive win. Then I sized things down to what I’d originally intended (which was still way too big, but I didn’t realise that either) and carried on.

And the strategy kept making money.

For the next four months, I watched my account climb. 50%, then 100%, then 150%. I felt invincible.

The Inevitable Reckoning

You can probably guess what happened next.

The strategy started losing money just as consistently as it had been making it. Day after day, week after week, the gains bled away.

I checked my code for bugs. I re-optimised parameters. I verified that my live performance matched my backtest over the same period.

Nothing explained it.

Eventually, I sized down further (funny how losing money makes you suddenly sensible about position sizing) and kept trading. But the losses continued. Death by a thousand cuts instead of a sudden explosion.

After four more months, I admitted defeat and turned it off.

What Actually Happened

Years later, I went back and looked at what I’d been trading. Turns out, my “clever” breakout system was just a very convoluted trend-following strategy.

Here’s a chart of the US Dollar Index during that period. The blue shading is when I made money. The red shading is when I lost it:

The dollar was in a roaring uptrend when I went live. My strategy, which was essentially just “buy things going up,” made money because things were going up.

Then the market went sideways. And I lost money trading trend while the market wasn’t trending.

I’d spent so much time messing about with entry rules and exit rules and stop losses that I never stopped to ask: “Is the underlying idea actually good?”

The underlying idea was that currencies trend. Do they? A quick look at the autocorrelation of dollar returns shows... basically nothing:

No persistence. No trend. Random noise.

I got lucky. I happened to start trading a trend strategy on one of the few occasions when the market actually trended.

The Lessons

Trading too big nearly ended me before I started. If those initial 10x positions had gone against me instead of for me, I’d have been wiped out within hours of pushing go.

Your P&L is a terrible feedback mechanism. When your only signal is “am I making money or not,” you have no way to distinguish luck from skill. Even with a real edge, you can be underwater for months. Without one, you’re just gambling.

Backtesting is not research. I tortured my backtest until it showed me what I wanted to see. That says nothing about whether the underlying idea is sound.

Ask the right question. Instead of “Do these rules make money in a backtest?”, I should have asked “What needs to be true for this to work?” For a trend strategy, markets need to trend. Do they? That’s something you can test directly.

I got very lucky with that first strategy. Lucky enough to survive and learn some important lessons.

Most people don’t get that lucky. And the lessons end up being a lot more expensive.

This is an excerpt from my case study on building a trading career as a finance outsider. If you found this useful, you can download the full case study here.

And the conclusion is clear: you should be long equities and go to the beach.

This is not financial advice. This is permission.

The Evidence-Based Case for Doing Nothing

Right now, in late December, we’re experiencing an unusual confluence of bullish seasonal effects. They’re all stacking up at once, like the universe is conspiring to let you log off guilt-free.

Let me walk you through the rigorous, data-backed justification for your holiday laziness.

End of Month Flows. The last week of the month tends to be bullish for stocks, if other asset classes have gone up month-to-date, driven by rebalancing flows from mandated funds. We’re in it right now. Check.

FOMC Meeting Effect. The window around Fed announcements tends to be bullish for equities. We had one a couple of weeks ago. Check.

VIX Expiration. Both the day of VIX expiration AND the day before tend to be bullish. That’s two bullish days for the price of one calendar effect. Check.

Pre-Christmas Effect. The days leading up to December 25 tend to see reduced volatility and positive drift. Something about people being in a good mood, apparently. Check.

Pre-New Year Effect. Same deal for the days before January 1. More positive drift. More reduced volatility. More reasons to not check your phone. Check.

That’s five separate bullish effects, all happening in the same two-week window.

The Optimal Strategy

Based on all of this highly sophisticated analysis, I can now reveal the optimal trading strategy for late December:

1. Be long equities

2. Go on holiday

3. Don’t check your portfolio

4. Eat too much

5. Come back in January

This is peak systematic trading. We’ve done the research, so you don’t have to think about this over Christmas lunch.

Some Caveats

Look, these are probabilistic tendencies, not guarantees. Any individual period can do whatever it wants. The market doesn’t care about our holiday plans.

Also, front-end volatility is extremely cheap right now. If you’re thinking about selling options, consider going out to 1.5-2 months where vol is still reasonable, rather than selling 7-8 vol at the front end.

And if you’re a buyer of vol, short-dated straddles look cheap (not just low-priced). Something to consider if you think the market might actually do something interesting.

But for most of us? The boring answer is probably correct. Be long. Go away. Enjoy the break.

The Real Point

I’m being tongue-in-cheek here, obviously. But there’s a genuine insight buried in the absurdity.

Sometimes the right thing to do is nothing. Sometimes all the signals point in the same direction and you just... follow them. Sometimes the sophisticated, research-backed, evidence-based answer is “go to the beach.”

Systematic trading isn’t always about complexity. Sometimes it’s about recognising when conditions are favourable and having the discipline to do the simple thing.

The simple thing is to be long and not overthink it.

Not totally balls-out long. Just regular I-can-sleep-at-night long (I intend to take a break, after all).

So that’s what I’m doing. I’ll be at the beach with my family, not checking my phone, and celebrating the year that was.

Happy holidays. Merry Christmas. See you in 2026.

Cheers!

The trade is simple: short natural gas futures around the winter months. I like to express it as a short position in BOIL, the 2x leveraged natural gas ETF.

It’s a seasonality play, but you also get exposed to a nice basis decay tailwind. Natural gas futures curves are usually steep, and the front months, which BOIL is most exposed to, tend to deliver consistent roll-down. You also get exposed to the frictions of a leveraged ETF: tracking error, volatility drag. All working in your favour when you’re short.

It’s a simple, dumb edge. It doesn’t shoot the lights out on its own, but it’s a nice little kicker to a broader portfolio. Typical of the stuff we trade with the RW Pro group.

If you backtest it, you’ll see it’s made money in a very ugly, uncomfortable way.

CAGR is about 12%, Sharpe is around 0.5. Certainly nothing to write home about, but a decent enough diversifier.

The thing to note is that it’s volatile as hell. It’s gone at about 50% annualised volatility, and it’s only in the market for four months a year!

This Year’s Ride

Here’s what the trade looks like so far this year, maintaining a roughly constant dollar exposure (rebalancing if it drifts from target):

You can see it was bleeding cash for a while before recently turning positive.

But the P&L isn’t the main point here. The volatility is.

How does it look if you don’t rebalance? Here’s the P&L:

It’s actually made a little more than if you did rebalance - but that’s not a reason to avoid rebalancing.

When you were at maximum drawdown without rebalancing, you were sized way larger than you intended.

Here’s the position size over time if you didn’t rebalance:

If we didn’t rebalance, we got lucky that the trade turned around the next day and made a lot of money.

But it could just as easily have been down by that amount again. No matter how crazy things get, they can always get crazier.

And how would you feel in that position at that time?

You’re sized way bigger than you ever intended (about 35% bigger), you’ve already bled a ton of cash, and now you’re facing the prospect of even more losses.

On the other hand, here’s the position size over time with rebalancing:

Totally different risk profile, totally different emotional state! Practically same P&L. And we never let the position get out of hand.

This is the job rebalancing does. It keeps your risk budget constant.

You’re saying: “I want X dollars of seasonal/decay edge, and I’m willing to accept Y units of nat gas volatility to get it.”

Without rebalancing, you’re letting the market decide how much risk you’re taking.

And you wouldn’t want the market dictating the size of your short position in one of the most volatile assets out there.

A sustained rally - cold snap, supply disruption - literally turns a sensible position into an existential one.

Why This Might Feel Wrong

When you rebalance a short position like this, by definition, you’ll be covering after it goes against you and adding after it goes your way.

That’s going to feel very wrong if you think there’s a difference between realised and unrealised losses.

But an unrealised loss is still a loss.

On your books, the value of your account is lower. If you were to liquidate everything, you’d be down that amount. Thinking an unrealised loss is “not real” is a dangerous delusion that leads to poor decisions.

Don’t let unrealised losses be a reason to accept whatever risk the market gives you. Fight like hell to keep your risk exposure where you want it to be (with consideration to the cost of rebalancing, of course). It’s one of the only things you have any control over.

The Bigger Lesson

The broader lesson here is that rebalancing to a risk target isn’t primarily about maximising returns (although often that’s a nice side effect). It’s doing a specific job: keeping your risk exposure under control.

It’s always tempting to do the thing that maximises returns in the backtest. But in the case of rebalancing a volatile position, that’s making two mistakes:

1. Thinking rebalancing is about maximising returns. It’s not. It’s about keeping your risk exposure where you want it.

2. Thinking your backtest is the thing to optimise for. Your backtest is a guide, not a gospel. Use it to understand implementation trade-offs and get a sense of historical performance. Don’t optimise for it.

Rebalancing is about staying in control. In a trade this volatile, that matters more than squeezing out a few extra points of return.

I used to start with data. Pull a bunch of price history, run some analysis, hunt for patterns that looked promising. Optimise the parameters until everything looked great. Then I’d go live and watch the whole thing fall apart.

The problem wasn’t the statistics. The problem was I had no idea why the pattern should exist in the first place.

The mental model came last when it should have come first.

Every price you see on your screen is the result of someone buying and someone selling.

Sounds obvious, right?

But most of us - myself included, for a long time - treat price data like it fell from the sky. Just numbers to be analysed in isolation.

In reality, prices move because participants with different objectives, constraints, and information interact. Some are trying to make money. Some are hedging risks they’d rather not have. Some are rebalancing portfolios on a schedule. Some are forced to trade regardless of price.

That last one is key. Forced to trade regardless of price.

When you understand who’s in the market and why they’re trading, you start to see where edge might really exist. Not because you found a correlation in historical data, but because you understand the mechanism creating it.

Building the Model

A useful mental model answers a few questions:

Who are the major participants? In equities, you’ve got index funds rebalancing mechanically, hedge funds hunting alpha, market makers providing liquidity, and retail traders doing... various things.

In crypto, there are basis traders, funding rate arbitrageurs, trend-followers, leverage-hungry speculators, and protocols with treasury management needs.

What are their objectives? Index funds want to minimise tracking error, not maximise returns. Market makers want to capture spread while managing inventory. Hedge funds want uncorrelated returns. Each objective creates predictable behaviour.

What are their constraints? This is where it gets interesting. Constraints create predictable, price-insensitive trading. An index fund has to buy when a stock enters the index. A risk parity fund has to rebalance when volatility changes. A leveraged trader has to liquidate when margin runs out.

When are they forced to act? End of month, end of quarter, index rebalances, options expiration - these create windows where certain participants must trade regardless of price.

From Model to Edge

Once you have this framework, research becomes more directed.

Instead of asking “is there a pattern here?”, you ask “who might be creating predictable price pressure, and can I trade against them?”

Instead of data mining for anomalies, you look for situations where participants with constraints are forced to trade at bad prices (for them).

Take rebalance flows. Index funds mechanically buy stocks entering an index and sell stocks leaving. This isn’t a secret. Everyone knows it happens. But knowing it happens and understanding the dynamics well enough to trade it profitably are different things.

You need to understand the timing, the size of the flows relative to liquidity (even if there’s no precision). How other participants might be positioning. Whether the effect is already priced in. Under what conditions it becomes more or less pronounced.

All of this comes from having a rich mental model, not from running regressions.

The Practical Bit

Before you dive into backtesting your next idea, ask yourself: Who’s on the other side of this trade? Why are they trading? What constraint or objective makes their behaviour predictable? Why won’t this get arbitraged away? Why could I trade against this?

If you can’t answer these questions, you’re probably data mining. And data mining without understanding causation is how you build strategies that worked beautifully in the past and terribly in the future.

The market is very rational on aggregate, even though there are a lot of maniacs in it. Your edge comes from understanding the structure well enough to know where rationality creates predictable behaviour you can exploit.

That mental model is the foundation everything else builds on. Get it right, and research becomes clearer.

This isn’t just about fun (though it is fun). It’s about not derailing your actual trading because you got an itch you couldn’t scratch.

Know Your Edge

Nearly all my trading is systematic (I’m not smart enough to do it any other way).

My positions make sense based on edges I can articulate, and whose footprints show up clearly in the data.

Risk premia. Providing liquidity to price-insensitive trading. Stat arb. The usual suspects.

But like anyone who’s obsessed with the markets, I like to cosplay the clever macro trader from time to time.

I’ve also been known to come home drunk and punt silver futures because “an idea came to me in the Uber”.

Of course, I know bugger all about macro trading. My edge is in doing useful things the market values and getting paid to do them well. I have precisely zero edge in predicting interest rates or housing bubbles or anything else that requires being smarter than the aggregate market.

But like every market obsessive, I’m going to get the occasional itch that’s just got to be scratched.

Systematic trading is about grinding out noisy edges over time. The worst thing you can do is express some macro view you have zero edge in with size - you can literally derail months of hard work.

So instead, I keep what my mate Euan calls a “cheat meal account.” It’s a tiny fraction of my total capital. And I can do whatever I want with it.

The Cheat Meal

Euan put this perfectly when we were talking about it recently.

“It’s the same way bodybuilders have a cheat meal. You can’t have a perfect diet all the time because eventually you crave chicken wings so much that you go on a bender and eat 50 of them. But if you let yourself have five chicken wings a week, that keeps you on track.”

That’s exactly it.

(Except five chicken wings ain’t cutting it for me… not even close).

If you’re sitting in front of screens all day, or even if you’re just really interested in markets, you’re going to have ideas. You’re going to see patterns - real or imagined. You’re going to think, “Oh, this looks interesting.”

And if you’re determined to never act on them, you’re not being disciplined. You’re being a coiled spring. Eventually, something’s going to give, and you’ll do something big and stupid.

At least, I know I would.

I’m not the trading police. I don’t have a boss who will slap me on the wrist or cancel my bonus if I do something out of line. I’m allowed to have a little fun.

What I Eat for Cheat Meals

My speculation account is where I put on trades that don’t follow from my process. Essentially, anything that doesn’t lend itself to being disproven through data analysis.

Maybe it’s a macro idea. Maybe it’s something I saw on Twitter that seemed interesting. Maybe it’s something I read in a newsletter. Maybe it’s just a feeling about something. Or maybe I just had too much to drink.

I definitely don’t want my trading results to depend on my ability to out-predict other people. I know that’s a terrible bet for me. But with tiny size? Yeah, I’ll have a punt. And I’ll really enjoy it.

I scratched the itch, lost a bit of money, moved on. No harm done. Maybe even learned something.

This Isn’t Just About Fun

The act of systematic trading is a grind. It’s boring. You turn up every day, run your processes, and reconcile your positions.

(For me, the fun stuff is the detective work that sometimes leads to a new strategy, the decision-making under uncertainty, the weighing of competing evidence. YMMV).

If you’re going to stick with this for the long haul, you need to actually enjoy it. And part of enjoying it, for me at least, is getting to occasionally take a punt on something interesting without it mattering if I’m wrong.

As an indie trader, you’ve got to indulge yourself a bit. You can’t just have your nose to the grindstone every day. That’s not sustainable.

And when you’re constantly in front of the markets, you’re going to act on things at some point. Unless you're Jocko Willink, perhaps.

And when you do, it’s not systematic. It doesn’t fit your process. You don’t have proper sizing for it because it’s not part of your framework. You don’t have clear rules for managing it.

That’s when bad things happen.

Much better to quarantine that stuff. Keep it separate. Keep it small. Let yourself be human without letting it wreck your hard work.

How Much?

It’s not fixed, but mine would only be a couple of per cent of total capital, maximum. Small enough that if I lost the whole thing, it wouldn’t materially affect my life or even my portfolio.

The exact number doesn’t matter. What matters is it’s small enough that losing it all wouldn’t hurt, but large enough that winning feels satisfying.

Some people might want 10%. Some might want 2%. Just don’t make it 30% or you’ve defeated the whole purpose.

What This Isn’t

This isn’t an excuse to trade like a muppet with a meaningful portion of your capital and call it “speculation.”

This isn’t a way to justify putting on positions you know are rubbish because “oh well, it’s just the spec account.”

And this definitely isn’t something to brag about when you get lucky on a coin-flip trade. (Trust me, you’ll have an urge to do this. Resist it.)

It’s just a pressure release valve. That’s it.

If you’re trading systematically, or trying to, consider keeping a small speculation account. Not because you need it to make money. You don’t. But because you’re human, and humans get ideas, and it’s better to act on those ideas in a contained way than let them infect your real trading.

I’m always doing some dumb stuff in mine. Some commodities idea, or sector trade, or whatever. And on the off chance it works, I’ve found a reason to treat my wife to a night out. And if not? I scratched the itch.

Five chicken wings a week, yeah?

A few times a year, I run a course for part-time traders who want to get serious about systematic trading. Not “quit your job” serious - more like “approach this with a realistic plan that might actually work” serious.

It’s called Trade Like a Quant, and here’s what it’s about.

The Core Problem

Trading is hard. Not because markets are mystical or because you need a PhD. It’s hard because it’s insanely competitive.

When we trade, we’re up against people who are just better at it than us. Faster technology, more capital, more experience. If we try to compete directly with them, we’ll lose. That’s just reality.

So here’s what I do instead: I look for places where the competition is relatively low. Places where I can do something useful that people value and are prepared to pay for.

This isn’t about outsmarting anyone. It’s about providing a service - usually in markets that are too noisy, too small, or too operationally awkward for bigger players to bother with.

How the Course Works

I’ve split it into two parts, roughly.

Part 1: Trade Like a Quant

This is about developing market intuition - what I call “trader smarts.” You’ll learn:

• How markets actually work (structure, participants, their constraints)

• What edge looks like and where it comes from

• How to identify when you can trade at attractive prices

• Tons of explicit examples of stuff that makes money, or made money in the past

I’m basically brainwashing you with examples until you start seeing patterns. You’ll look at a market situation and think: “Oh, someone has to trade here. They’re price-insensitive because of X. That’s an opportunity.”

Part 2: Quant Like a Trader

This is the practical toolkit. Simple data analysis for simple research:

• Data munging, distributions, scatter plots

• Dealing with bias and non-stationarity

• Factor analysis, event studies

• Simulation techniques

Plus portfolio construction:

• Diversification and sizing

• Rebalancing mechanics

• Risk management

Nothing complicated. Nothing academic. Simple approaches work fine.

What You’ll Actually Do

The course material is available on demand. We’ll chat on Discord and on Zoom until you get it. I’ll explain things ten different ways if needed.

It’s not a quant finance course. No fancy math. Just simple explanations, poorly drawn pictures, a smattering of Dad jokes, and the basics you need to make money sensibly.

The Philosophy Behind It

Here’s my thing: You want to concentrate on opportunities where it’s clear and obvious why you’re getting paid.

Not marginal games. Not trying to predict better than everyone else. That won’t work.

Instead, find uncompetitive places where you can do something useful. Things like:

• Providing liquidity when someone needs to trade quickly

• Taking the other side of forced rebalancing flows

• Trading in markets too small for institutional capital

• Helping remove risk for producers who need to hedge

This article presents a typical example.

These aren’t sexy. They’re often noisy or capital-constrained or operationally awkward. But that’s precisely why they’re available to us.

Look, I know I’m biased, but people who’ve been through this course tell me there’s nothing else like it. That it changed their entire perspective on trading.

Common Questions

“Why teach this instead of just trading it and making gazillions?”

Two reasons: I like teaching, and none of this stuff will make you gazillions. I’m showing you where to find edge and giving you tons of examples. But these are part-time trader opportunities, not quit-your-job opportunities. Once you know the fundamentals, we can talk about how you might build something more ambitious.

“Isn’t active trading always adversarial and zero-sum?”

No, not at all. You’re usually getting paid because you’re doing something helpful to the market.

Think about it: Someone needs to quickly remove risk. I’m willing to help them do that, at the right price for me. They’re happy - they achieved what they needed. I’m happy - I traded at a good price.

This applies to:

• Producers or businesses hedging their exposures

• Funds with large flows that need to rebalance

• Long-term investors prepared to “pay up” to get their trades done

It’s not all adversarial. The key is finding places where the competition to “do something useful and get paid” is relatively small. And that’s usually because doing it is unattractive in some way: risky, uses too much capital, can’t scale, whatever.

“How do you tell if an edge is competitive?”

Practically: Can you observe pricing inefficiencies persisting for a long time? Which is just a fancy way of asking: “Could you make money trading it even if you weren’t the fastest player?”

If yes, there’s probably room for you.

Who This Is For

You’re probably a good fit if:

• You’re analytically minded (technical background helps, but isn’t required)

• You suspect that drawing lines on charts isn’t a realistic approach to trading

• You can commit a couple of hours a week (or a few hours on the upcoming intensive version)

• You want to understand why things work, not just what works

• You’re willing to trade simple, unsexy strategies

This probably isn’t for you if:

• You want get-rich-quick systems (systematic trading is a game of slowly grinding out noisy edges over time)

• You’re looking for the holy grail strategy (doesn’t exist, and if it does, I don’t have it)

This Month: Live Intensive Format

Normally, Trade Like a Quant is self-study - you work through the material at your own pace. But this November, I’m running it as a 10-day live intensive starting November 3rd.

Same curriculum, but we work through it together. You get real-time answers to your questions, see how I evaluate opportunities, and learn alongside other traders wrestling with the same concepts.

When you enrol, you get immediate lifetime access to all the course materials and the Discord community. So you can familiarise yourself with the content before we start, or just show up fresh on November 3rd. Either works.

The intensive is $447. I’ve kept the price there for years, but I’m raising it to $547 after this cohort - so if you’ve been thinking about it, now’s the time.

Want more details on the full curriculum? Here’s the complete course breakdown.

The Real Goal

I want you to stop feeling lost. To look at markets and have a framework for evaluating opportunities. To build a simple, robust trading operation that fits around your life.

And most importantly: to develop the confidence that you can find new edges independently. That’s the skill that actually matters.

So, if this sounds like your kind of thing, here’s where you can enrol in the intensive.

Questions? Just reply to this email or leave a comment.

Until next time,

Kris

It’s just over a week since the crypto flash crash.

If you got caught up in it, I’m genuinely sorry. It sucked. There wasn’t a whole lot you could have done to sidestep it - it wasn’t specific to a platform, a chain, or even a particular coin. The only way you really avoided it was by being out of crypto entirely.

Systematic trading is about showing up, doing the hard work of grinding out our edges over time, and doing the best we can with what we know.

In the spirit of fleshing out “what we know,” let’s talk about what happened, what it means, and what I learned.

The Uncomfortable Truth About Auto-Deleveraging

The crypto ecosystem is basically thousands of live experiments in economics, finance, and technology running simultaneously. Most of these experiments fail spectacularly. A few lead to genuine breakthroughs. All of them have unintended consequences and different sets of trade-offs.

This is both crypto’s greatest strength and its most dangerous liability.

The edge cases - the moments when everything goes sideways - are where these experiments get truly tested. It’s where you find out whether someone’s clever idea actually works under stress, or whether it just shifts risk around in ways that seemed fine until they very much weren’t.

Saturday was one of those moments.

One of the more interesting experiments in crypto has been around counterparty risk management.

Traditional exchanges solved this problem decades ago with clearing houses, insurance funds, and standardised margin requirements. Crypto is an opportunity to run a bunch of different experiments to see if there are better ways that can work under different use cases.

Auto-deleveraging (ADL) is one of those experiments.

ADL is a philosophical design choice that prioritises platform solvency over winning traders keeping all their profits.

It feels deeply unfair when you’re on the receiving end of it. You were in the right positions. The market moved in your favour. And then suddenly you’re forced to exit at worse prices because someone else blew up their account.

But there’s a trade-off: that mechanism is precisely what kept Hyperliquid (and other platforms) from accruing bad debt and potentially going under.

Different exchanges handle this differently. Some have insurance funds that absorb bad debt until they run out. Some socialise losses across all users. Some use ADL to force profitable traders to close out losing positions. Many use a combination.

There’s no perfect solution - only different sets of trade-offs.

The key is understanding the trade-offs for the platforms you trade on and deciding whether you’re comfortable with them.

What I Got Wrong

I’ve been trading on Hyperliquid for several months now. I’ll be honest - before last weekend, I didn’t fully understand the details of Hyperliquid’s ADL system. I knew it existed, but I hadn’t dug into the specifics.

That was lazy of me. I’ve now fixed it. You should do the same, if you haven’t already.

Go read the documentation for whatever platform you’re trading on. Understand the waterfall of liquidation mechanisms:

• How does standard liquidation work?

• What happens if that fails?

• Is there an insurance fund or backup liquidator?

• When does ADL kick in, and how are users selected?

• If your platform doesn’t use ADL, what does it do in these “last resort” situations?

This isn’t exciting work, but it’s important. These edge cases matter precisely because they happen at the worst possible time - when everything’s already going pear-shaped.

I was running a reasonably sensible 2.5x leverage on my positions. Had plenty of buffer to my maintenance margin. Didn’t matter. I still got ADL’d because of what was happening to the system as a whole. Careful position sizing can’t protect you from the system-wide risk that contributed to this.

That’s worth thinking about. Is it a risk you want to take? If so, how big do you play the game?

The Feedback Loop

What made Saturday so brutal wasn’t just the initial selling. It was the feedback loop of liquidity destruction.

Here’s how it played out:

1. Initial heavy selling comes in during a quiet market period

2. Market makers pull quotes and widen spreads (sensible risk management on their part)

3. Now there’s less liquidity to absorb the selling, so prices move further than they otherwise would

4. Those bigger moves trigger more liquidations

5. Liquidations mean more forced selling

6. Which triggers more quote pulling, more liquidations, bigger moves...

And on and on until it stabilises.

The TradFi Alternative Worth Considering

Here’s something I’ve been thinking about: if you’re just trading the big four (BTC, ETH, SOL, XRP), you can do that on CME.

I run a simple multi-factor model on the top ten crypto assets by market cap. I was actually surprised to find that the backtest is pretty solid using just those four available on traditional exchanges.

Running a version of that strategy using a four-asset universe on CME essentially trades systemic platform risk for concentration risk. That might be a decent trade-off.

Yeah, there are still risks. Your broker could get into trouble, for instance. But the Russian doll of potential failure points is much smaller. You don’t face the same systemic risks on a traditional exchange as you do on crypto platforms. Let’s be honest… It’s still the wild west, and will be for some time.

Something to think about.

What I Would Have Done Differently (And Why It Matters)

This is the interesting bit: what I would have done differently would have actually cost me money this time, but I still think it’s the right approach.

I’m running a roughly delta-neutral, multi-factor long/short strategy. Classic quant trading approach.

If I’d been at the screen when this was happening, I would have cut my long positions as soon as I realised I was getting ADL’d on the shorts.

At that moment, I would have had no idea what was going to happen next. The market could have kept tanking. I would have been sitting there net long with no edge whatsoever.

I got lucky. The market came back. But counting on that is not how you build a sustainable trading business.

The fundamental principle here is this: I only want exposure where I have an edge.

The moment I got ADL’d, I had zero edge in predicting what would happen next. The right move would have been to balance my exposure as best I could, even though this time it would have cost me profits.

This is the meta-lesson: always keep in front of your mind what you actually have an edge in.

Trading is hard enough when I do have an edge. I don’t want exposure to situations where I don’t have one.

What It Means for Crypto Going Forward

We’re going to lose a chunk of users, at least temporarily. Whether it’s true or not in this specific case, there’s a perception of manipulation and no real recourse when things go wrong.

That perception isn’t going to be helped by events like this.

Maybe this erodes certain edges.

The carry trade and basis opportunities depend a lot on FOMO-driven speculative trading. If that user base shrinks permanently, those opportunities might diminish too.

Then again, maybe it doesn’t. People will probably eventually come back, if history is a guide.

I’m not changing my approach for now. Still running my strategies, just with a slightly smaller size given the increased volatility. But I’m watching OI and funding rates with interest.

The Only Wrong Answer

I’m not going to tell you whether you should or shouldn’t trade crypto. That’s deeply personal and depends on your risk appetite, your goals, your constraints, and your stage of life.

Both of these statements are correct:

1. “I accept the risks with eyes wide open because the potential opportunities justify it for me.”

2. “For me, the risk of losing everything I put in just isn’t worth those potential opportunities.”

The only wrong approach is not thinking about it in terms of trade-offs.

The opportunity in crypto is real. You can run strategies that worked in traditional markets twenty years ago, and they still work today. That edge exists because of natural constraints on institutional players combined with a large retail user base doing FOMO and noise trading.

The risks are also enormous. Everything can go to zero. Not just because it becomes worthless, but because of scams, manipulation, and financial engineering experiments gone awry.

For me, I never had more than a chunk of my trading capital in crypto. If I’d lost it all on Saturday, it would have made me quite sad, but it wouldn’t have been catastrophic. That’s the right calculus for me.

Yours might be different. And that’s completely fine.

What Now?

Here’s what I’m doing:

Immediate actions:

• Actually understanding the liquidation mechanics of every platform I trade on (should have done this ages ago)

• Considering moving some core positions to CME to reduce platform risk

• Setting up alerts for liquidation events so I know immediately if ADL happens

• Maintaining proper position sizing where I can afford to lose what I have in crypto

Watching closely:

• Open interest levels. Are traders coming back?

• Funding rates. Is the speculative interest returning?

• Whether the edges we’ve been trading are still there with a smaller user base

Not changing:

• The fundamental approach of grinding out small edges over time

• The principle of only taking exposure where I have an edge

• Position sizing discipline

The game hasn’t fundamentally changed. We still show up, grind out our edges, and manage risk as best we can.

But last weekend was a good reminder that crypto carries risks that just don’t exist in traditional markets, and we need to factor that into everything we do.

What’s your take on ADL versus insurance funds versus socialised losses? Genuinely interested in hearing different perspectives on this. What trade-offs matter most to you?

Drop a comment or hit reply. I’d love to hear your thoughts.

You’ll get three things from this article:

• First, you’ll understand why detecting regime changes with statistical tests is basically impossible in real time.

• Second, you’ll see why waiting for statistical significance before trading an edge often means you’re too late.

• And third, you’ll learn what actually matters when evaluating trading ideas.

Let’s dive in.

The Problem: Markets Are All Noise

Here’s a thought experiment my colleague Euan Sinclair put together.

Imagine I’ve got a strategy with a true, long-term return of 10% per year and 20% volatility - roughly the S&P 500’s long-term profile. I’ve been trading it for five years. It’s been working as expected.

Then something changes. Maybe the market microstructure shifted. Maybe the edge just got fully absorbed by the market. Whatever. The point is, my edge disappears. For the next month, my strategy’s expected return drops to zero, but the volatility stays at 20%.

Of course, I’d like to detect that change. I’d like to know when to stop trading the strategy.

So I do what any rigorous, technically minded trader would do: I run a t-test comparing my five years of historical returns to this most recent month. That’s about 1,260 trading days versus 21 days.

Surely that’s enough data to spot a problem?

Hell no!

Here’s why: The expected daily drift for a 10% annual return is tiny - about 0.04% per day. But the daily noise from that 20% volatility is about 1.26%. The noise is thirty times larger than the signal.

Your daily Sharpe ratio is roughly 0.03. That’s an astronomically low signal-to-noise ratio. Even if this edge disappeared entirely, you’d barely notice in a month.

When Euan ran this simulation and tested it with a standard t-test, the p-value came back at 0.12. Not statistically significant. The test found nothing. The collapse in performance didn’t even register as a blip.

He also tried a Kolmogorov-Smirnov test - a non-parametric version that looks at the whole distribution, not just the mean. Surely that would catch something? Nope. The p-value was 0.37. Even worse.

(Euan goes deeper in his article - you can read it here)

To make this more real, I simulated 10,000 5-year artificial stock price histories using a GBM process with 10% annual return and 20% volatility. For each price series, I then simulated a month where the returns dropped to zero but the volatility remained unchanged. I performed a t-test on each price history to see how often it detected the change in the underlying process and plotted a histogram of the results.

If the test was useful, I’d see it working most of the time. That would show up as a spike in the left extreme of the histogram.

Instead, the p-values are roughly uniformly distributed (notwithstanding the top and bottom buckets). That suggests that any p-value is just as likely as any other. Which means that the test has no ability to distinguish between the two regimes.

We get 5.35% of p-values less than or equal to 0.05, which is essentially what we’d expect from random chance alone.

It Gets Worse

Here’s the “funny” part: In Euan’s simulation, that “dead” month (the one where the edge had completely disappeared) still managed to return +7.5%.

That’s right. A zero-EV month beat 93% of all months during the prior five years of positive-EV data (EV = expected value, or average result if you could repeat the observation infinitely).

How is that possible?

A 20% annual volatility translates to roughly 5.8% monthly volatility. Even with zero drift, random chance will give you a +5.8% month about 16% of the time. A +7.5% month (about 1.3 standard deviations) happens roughly 10% of the time purely by chance.

Meanwhile, in the “good” regime with a 10% annual return, the expected monthly gain is only +0.8%. So a lucky +7.5% month easily beats the vast majority of historical months.

You end up with this bizarre headline: “Our strategy just lost its edge, but had its second-best month ever.”

Noise does that.

The Pattern Holds Across Thousands of Simulations

To make sure this wasn’t a fluke, I ran 10,000 simulations of the same setup. The results are shocking:

• The median one-month return (with zero drift) was about -0.017%, but the 90th percentile was +5.9%.

• In half of all runs, the zero-drift month beat 44% of historical months.

• In 17% of runs, it beat more than 80% of prior months.

• In 9% of runs, it beat more than 90%.

• Remarkably, in 1.2% of runs, the zero-drift month was the best month.

So one in every eleven “dead” months looks like a top-decile success. One in every 82 dead months looks like the best month out of the last five years!

Here’s a visual representation. This density plot shows where zero-drift months land when ranked against 60 months of positive-drift history, across 10,000 simulations.

If zero-drift months consistently ranked near the bottom, you’d see a sharp peak on the left side of this chart. Instead, we have a low peak, a gentle slope, and a near-uniform distribution across the middle.

That means a “dead” month is marginally more likely to rank in the 10th percentile than the 90th, but it’s essentially equally as likely to rank in the 25th as the 75th.

The distribution reveals why statistical tests fail: one month of returns from a broken strategy looks essentially identical to one month from a working strategy.

The signal is completely drowned out by noise.

We don’t have a detection problem. We have a “you can’t tell the difference” problem.

Being Better at Statistics Doesn’t Help

When I learned about the t-test’s weakness, I tried non-parametric tests like Kolmogorov-Smirnov. Surely that’s more robust?

No. When two normal distributions differ only slightly in mean but have the same variance, the KS test has less power than the t-test. It’s designed to catch shape differences - fat tails, variance shifts, asymmetry - not small mean drifts.

Sharpe ratios, t-tests, and p-values were designed for large-sample, low-noise situations. Financial returns are the opposite: small signals, fat tails, short samples.

These tests need thousands of data points to hit statistical significance, given the level of noise in our data, so they’ll never detect regime shifts in anything close to real time. The market will move on long before statistics catch up.

It Gets Worse Again

OK, I’m really laying on the pessimism here. But bear with me because it does get better eventually. Spoiler alert - there is a solution to this problem.

Even if statistical tests could reliably detect when strategies stop working, that wouldn’t solve your problem. Because by the time an effect shows up as statistically significant in your tests, everyone else has spotted it too.

The edge is eroding, and has been for some time.

Think about new markets. When crypto perpetual futures first gained popularity, there was no way to run meaningful statistical tests on funding rate dynamics. The product design was new. The data didn’t exist yet.

But traders who understood similar mechanisms from traditional markets could immediately spot and exploit the inefficiencies without waiting years for data to accumulate.

They had a reason to expect the edge would exist.

Often, the best opportunities in trading exist in places where you can’t get any kind of statistical significance. And the very best tend to be where there’s basically no data at all.

Getting comfortable trading under massive uncertainty, because the logic makes sense based on other things you know, is where systematic trading departs from traditional data science.

What I Do Instead

I’m not saying you should trade blindly based on hunches. I occasionally use statistical tests myself. I even show a useful Bootstrap test in my Bootcamp course. But the reality is that in a low signal-to-noise, fat-tailed environment, these tests rarely give you much additional information. At best, they’re a tiny piece of evidence. But that evidence is swamped by better evidence - such as a credible reason for the edge to exist.

Here’s what I focus on:

1. Have a strong and credible basis for the effect:

This is by far the most important thing.

If there’s a credible reason for the trade to exist - a risk premium, something structural, forced flows from rebalancing - then that provides solid evidence.

• Risk premium: People paying you to take risks they don’t want

• Forced flows: Price-insensitive buying and selling (for example, from fund redemptions or index rebalancing)

If you understand why something should work, you can trade it with more confidence than if you only know that it has worked in the past.

2. The hypothesis should be supported by whatever data you do have:

• Use simple analysis techniques to look as directly at the effect as possible.

• Is it persistent over time?

• Does it show up where you expect and not where you don’t?

Simple scatter plots and aggregations are often enough to reveal useful patterns. For example, you don’t need a p-value of 0.05 to see that relative equity/bond performance tends to revert.

The Practical Reality

When a strategy underperforms for a few weeks, you face two potential errors:

Type I error: You think it’s dead when the underperformance is just noise. You abandon a still-valid edge.

Type II error: You think the underperformance is noise when it’s actually dead. You bleed transaction costs and take on unrewarded volatility.

Statistics tries to balance those errors - but they’re horribly lagging indicators. They confirm what you already know, long after it’s actionable. A strategy doesn’t announce its death with a p-value. It fades, subtly, while your t-statistic bounces around all over the place.

So if you can’t rely on statistics in a practical sense, what can you do?

Trading is Pragmatic

We need to be pragmatic about our objectives and honest about what we know and what we can control.

Our objective is simple: to turn money into more money at a level of risk that we can tolerate.

What do we know?

• That there’s evidence (but no guarantees) that our strategies are positive EV (they’re grounded in reality and supported by data)

• That returns are noisy, especially in the short-term

• That our returns don’t tell us if a strategy is working or not at any time scale we care about

What can we control?

• Our exposures (position sizing)

• What we choose to trade

To be really clear, these are things we don’t know or can’t control:

• We don’t know with certainty if our edge has disappeared (or was even present in the first place)

• We don’t know what returns the market will throw up

This pragmatic reality informs my entire approach to trading:

Trade as broadly as possible

We all know that diversification helps smooth out your portfolio returns. When one strategy zigs, another zags, and your overall volatility tends to come down.

But there’s another underrated reason to trade broadly - it reduces the chance of trading without an edge.

Remember, we don’t know with certainty that a strategy is still working, or if it ever worked at all. So, trading more good edges increases our chances of having an edge at the portfolio level.

Stack the odds in your favour

But don’t just trade any old thing. Do the work to ensure that anything you do trade is likely to have a real edge:

• It’s based on a real phenomenon

• It shows up in the data

• You have reason to think you can compete

Trade at a sensible size

If I’m wrong about something having an edge (and I will be wrong sometimes), it hurts me through transaction costs and unrewarded volatility.

Nothing I trade is particularly expensive (low frequency, liquid assets, for the most part), so I’m not too worried about transaction costs.

But the last thing I want is for some edge that I’m wrong about (or any edge, for that matter) to be sized so big that its returns dominate the portfolio.

So I think about how big a strategy needs to be to contribute a certain amount of volatility, on average.

How much volatility?

Depends on the strategy. I typically start with equal volatility as a baseline assumption, and then size things up or down based on what I know.

Is it an edge that I’m confident will persist because it’s based on risk preferences? It gets a little more weight. Does it have negative skew, like selling volatility? It gets down-weighted some. You get the idea.

Turning strategies off

I’ve tried all sorts of statistical approaches for turning strategies off based on P&L. None of them tell you anything worth knowing on a time scale you care about.

Let’s say you have a bad month and are worried.

Given that you’ve done the work and have satisfied yourself that there’s a good reason for the edge to exist, you’re sized such that being wrong can’t hurt you very much, and that your short-term P&L tells you nothing useful, the best approach is to:

• Make sure you have the correct positions on and are comfortable with your sizing.

  • If you’re not, then adjust accordingly.

• Relax

• Give it time for the edge to play out

• Focus on finding new things to trade

If you’re really worried, you’re almost certainly sized too big. Given that you have good evidence for the edge, err on the side of turning it down rather than turning it off.

Agonising about short-term P&L is essentially wasted energy. It won’t help you one bit. On the other hand, finding more things to trade will help a lot.

Focus obsessively on what you can control.

Relax about the rest.

Summary

I tried using statistical tests to “confirm” an edge or tell me when a strategy stopped working. Turns out, that’s impossible on any timescale that matters.

Markets are too noisy. By the time a test shows statistical significance, everyone else has spotted the problem too - and the best part of any new opportunity is already over.

So here’s what I do instead:

I start with edges that have a credible reason to exist - risk premiums, forced flows, structural inefficiencies. If I can’t explain why someone would pay me to do something, I don’t trade it (usually).

Then I look at whatever data I have. Not to achieve some arbitrary p-value, but to see if the effect shows up where I expect it to, and doesn’t show up where I don’t. Simple scatter plots and basic analysis are usually enough.

I trade as broadly as possible because diversification does two things: it smooths returns, but more importantly, it reduces my chance of trading without an edge. If I’m wrong about one strategy, it won’t kill me.

I size things sensibly. Equal volatility as a baseline, then adjust up or down based on what I know about the edge. If something has negative skew, I size down. If it’s based on stable risk preferences, maybe I size up a bit.

And when I’m worried, I check my positions, make sure my sizing is still right, and then I relax. Agonising over short-term P&L is wasted energy. Finding new things to trade actually helps.

The right question isn’t “Has my edge stopped working?” It’s “Given what I know, how much do I trust it now?”

The answer is always probabilistic, never definitive.

Taking action in the face of uncertainty is one of the things that makes trading hard… but it’s also why it’s the most exciting game in town.