Losses Arrive Together
Published 2026-07-27 · By Drew Shelem
Why consecutive trades aren’t independent — and how that makes your drawdown twice as deep at the same Sharpe
Article 7 of the cycle. The earlier articles removed traps one at a time: the win rate, the fee, the search across configurations, the costs of execution, the random out-of-sample slice. One assumption still sits quietly underneath all of them, and underneath the significance formula itself — that each trade is independent of its neighbors. On 5–15-minute markets that’s false, and the false part costs you the most exactly where you actually lose money: in the drawdown.
Where this topic even comes from
A familiar picture. A bot on 5-minute BTC contracts made money for a month. In the backtest its Sharpe was 1.8, the equity curve was smooth, and the deepest drawdown over the whole test history was 8%. You take it live, and in three days the account drops 18% — twice as deep as the backtest showed over its entire history. And nothing broke: the average size of a win is the same, the share of profitable trades is the same. The losses that were scattered across time in the backtest, not getting in each other’s way, arrived live as one clumped run.
This isn’t rare bad luck or a malfunction. It’s a predictable property of markets like these, and here’s where it comes from.
Every earlier article in the cycle — on sample size, on counting trials, on the fan of out-of-sample results — quietly leaned on one premise: that each trade is independent of the one before it, like a separate coin flip that remembers nothing about the last flip. The formula everyone uses to decide whether an edge is real rests on the same premise. On 5–15-minute markets that premise is false: trades don’t arrive one by one, they arrive in linked runs. This article works through why that happens, and what it costs you — both in how significant your edge is and, above all, in your drawdown.
What this is about — briefly, no jargon
Below I’ll show three things, and all of them follow from one fact: consecutive trades resemble each other rather than being independent.
First, because of that resemblance, your real sample is smaller than you think. Two thousand trades arriving in runs carry about as much information as a few hundred genuinely independent ones. Which means the edge is confirmed more weakly than the formula claims.
Second — and this is the main one — the drawdown comes out noticeably deeper. At exactly the same average win and the same share of profitable trades, the account falls twice as deep, because the losses come in a run and don’t let the account catch its breath between them.
Third, the usual ways of judging how reliable a strategy is (including that fan from the previous article) understate the true spread, because they too treat trades as independent.
The good news: this is not hard to fix, and I’ll lay out how at the end. The bad news: if you don’t know about it, a smooth backtest confidently lies about the parameter most dangerous to you — the depth of the pit.
How to check I’m not making this up. I took two strategies with exactly the same average win, spread, and share of profitable trades. The one difference between them: in the first, trades are independent; in the second, they come in runs. If the runs changed nothing, the two would behave identically. They behave differently, and I’ll show by how much, in numbers. All the numbers come from synthetic data, where I control the degree of runs myself and hold the per-trade metrics fixed — that way the pure effect is visible, unmixed with anything else. How much your own trades run in streaks is a separate question, and at the end I’ll say how to measure it for yourself.
Let’s break it down step by step
Step 1. What “trades come in runs” means, and why it happens
Start with the fact itself. The market doesn’t switch between calm and storm randomly and instantly. Calm stretches stay calm and turbulent ones stay turbulent: if it’s volatile now, then five minutes from now it’s most likely still volatile. This property has a name — volatility clustering — and it’s one of the most reliably confirmed facts about financial markets in general: quiet and turbulent periods come in chunks, not shuffled together.
Now watch what that does to your trades. Inside a turbulent stretch your strategy produces similar trades — say, a run of entries on wide moves — and inside a quiet stretch a different run, smaller ones. So the trades don’t arrive at random; they arrive in groups tied to whichever regime the market is in right now. Knowing how the last trade turned out, you already know something about the next: they’re probably from the same regime. That “a trade remembers its neighbor” is called serial correlation, or, more simply, inertia.
How strong that inertia is for you specifically depends on your strategy and your market, and it has to be measured (how, at the end of the article). But the fact that volatility clustering exists and pushes trades into runs is not a hypothesis — it’s a general property of markets. Everything that follows is stronger the stronger your inertia is, and it collapses to zero only if you have no inertia at all.
Step 2. Why runs shrink your real sample
An analogy from surveys helps here. Suppose you want to gauge a city’s opinion and you polled 2,000 people. But you polled them not one at a time, but by whole households — several people from each home, and inside a home people mostly agree with one another. How many genuinely independent opinions do you actually have? Not 2,000, but roughly as many as there are households. Each additional family member adds almost nothing new: they repeat what you already heard from their relatives.
With trades it’s exactly the same. Trades inside one regime are “one household”: they resemble each other, and each next one adds little to what you already know from the previous. So for the question “is my edge real or did it just look that way,” a run of ten similar trades weighs less than ten independent ones.
Let’s check it in numbers. I set a moderate inertia (such that neighboring trades are noticeably, but not completely, alike) and held the average win and spread exactly the same as in the independent case. Here’s what happened to the sample:
| What we count | Value | What it means in practice |
|---|---|---|
| Trades, nominally | 2,000 | that’s how many you counted in your log |
| Of them, effectively independent | ≈ 664 | the other ~1,340 mostly repeat their neighbors |
That is: your 2,000 trades over two weeks carry about as much evidence as 664 genuinely independent trades would. Two-thirds of the sample is, from the standpoint of “am I sure about the edge,” a repetition of neighbors, not new evidence.
From there it hits significance. The significance of an edge is usually measured by a quantity I’ll just call the “confidence number” (in formulas it’s t = Sharpe × √N): roughly, how many steps your edge sits away from zero; above about two, it’s conventionally taken to be non-random. The formula uses the nominal number of trades. Plug in the honest one, and the picture changes:
| What we count | Value | What it means |
|---|---|---|
| Confidence number using the nominal 2,000 trades | 2.30 | above two — “the edge isn’t random” |
| The same, using the honest 664 trades | 1.32 | below two — “could just be luck” |
Your confidence in the edge turns out about 1.7 times smaller than the formula claims. The same edge that is “significant on paper” is, on the honest sample, indistinguishable from luck. This is a direct continuation of the cycle’s first article (which was about how many trades you need to tell an edge from noise) and its third (where the same formula appeared): both treated trades as independent — and this is what that indulgence costs, in numbers.
Step 3. The main one: same statistics — but the pit is twice as deep
This is where inertia hits hardest, so let’s go slowly.
First, what we’re measuring. Maximum drawdown is the deepest fall of the account from a reached peak down to the following trough: the worst plunge you’d have lived through trading this strategy. It’s this, not the average result, that decides whether you survive long enough to reach the edge, and whether you get knocked out first.
Now, how the experiment is set up, so the numbers are clear. I took a strategy with a thin but real positive edge: on average it makes about 0.05 of a “unit” per trade, while a typical single trade swings the account by about 1 such unit, up or down. (The unit here is arbitrary — think of it as the size of one ordinary trade.) I ran 2,000 trades many times over and looked at how deep the worst plunge comes out, on average and in the worst cases — first for independent trades, then for trades in runs, with the same average and spread.
| Maximum drawdown (in sizes of one typical trade) | Independent trades | In runs | How much deeper |
|---|---|---|---|
| In a typical run-through (median) | 27.6 | 56.1 | twice |
| In an unlucky run-through (worst 5%) | 48.8 | 108.1 | 2.2 times |
How to read this. The number “27.6” means the plunge was about 28 ordinary trades deep — the account gave back about what 28 typical trades make or lose. For the strategy whose trades come in runs, that same plunge is about 56 trades deep, twice as much. And in the unlucky run-throughs (one in twenty), the strategy-in-runs plunges 108 versus 49 — more than double.
Why, if their average win and share of profitable trades are identical? Back to the weather. Independent losses are separate raindrops, scattered across time: between them the account recovers, and no deep pit builds up. Losses in runs are a downpour: they arrive in a clumped series, one after another, the account never catches its breath, and the plunge stacks into a deep pit. The average trade over the whole period is the same — but the trades arrived in a heap, not spread out.
The practical takeaway is harsh. Two strategies with the same Sharpe survive differently: the one whose losses clump wipes out the account where the other merely dips and carries on. A smooth equity curve in a backtest, computed as if trades were independent, shows a pit twice as shallow as the real one. Your position size and stop have to be set for a clumped run of losses, not for the average loss — otherwise you get knocked out on a drawdown you never saw in the backtest.
Step 4. How to see the truth — reshuffle in runs, not one by one
A strategy’s reliability is usually checked with a bootstrap. The idea is simple: you take your trade history and reshuffle it many times, each time reassembling a random version, and see how much the result wanders from version to version. The more it wanders, the less reliable your single real result is.
The catch is that an ordinary bootstrap reshuffles trades one at a time. And because it breaks them up individually, it tears the runs apart and thereby pretends the trades are independent. The result is a spread narrower than the truth — you look more confident than you are.
The fix is a block bootstrap: reshuffle not one trade at a time, but in consecutive chunks, keeping neighbors together. Then the runs stay whole inside the chunks, the inertia doesn’t vanish, and the spread comes out honest. The chunk length is taken to be about the length of a typical run.
| Way of estimating the spread | Width of the interval for Sharpe | What it quietly assumes |
|---|---|---|
| Ordinary bootstrap (one trade at a time) | 0.076 | trades are independent — which is false |
| Block bootstrap (in chunks) | 0.131 | runs preserved — honest, interval 1.7 times wider |
And here’s the direct link to the previous article. Remember the fan of 924 out-of-sample slices, and the caveat that it understates the true spread? The reason is exactly this: the slices are built on inertial data, while the method treats them as independent. The block approach fixes the fan too — the chunks keep the runs intact, and the fan widens to its honest width.
What would refute this — and where the method is powerless
The condition on which the whole article would collapse (and didn’t). If inertia didn’t deepen the drawdown and didn’t shrink the sample at the same average win and share of profitable trades, there’d be nothing to write about. But at fixed per-trade metrics it deepens the drawdown twofold and shrinks the sample threefold. That was my attempt to kill the thesis; it didn’t die.
The honest boundary — and it’s about you personally. The strength of everything above depends on how much your trades come in runs. I took a moderate inertia as an example. If your strategy’s trades are nearly independent, the effect will be small, and the article barely concerns you. So you can’t take it on faith — you have to measure your own inertia (how, below). The direction of the conclusion is nonetheless ironclad: any positive inertia deepens the pit and weakens the significance, never the other way around. The magnitude — 2× on drawdown, 3× on the sample — is from my example; yours will be your own.
What the example does not claim. I modeled inertia as a smooth decay: today’s trade strongly resembles yesterday’s, the day-after’s less so. In reality the dependence may be built differently — for instance, abrupt switches between regimes rather than a smooth fade. That changes the specific numbers, but not their sign: any positive dependence pushes in the same direction.
What I don’t know, and where the limits are
- All the numbers are from synthetic data. I set the degree of inertia myself (a moderate one). This is a demonstration of the mechanism, not a measurement of your strategy.
- The size of the effect is your own. The twofold drawdown and threefold sample shrinkage are at the inertia I set. Weaker inertia, weaker effect.
- The sign of the conclusion is more reliable than its shape. Real dependence may be abrupt rather than smooth. Same direction, different numbers.
- You have to measure the right quantity. The direction of BTC over 5 minutes may itself be nearly independent from window to window — and your strategy will still be inertial, because its trades are driven by the volatility regime, not by the direction of price. What you measure is the correlation of the returns of your own trades, not the raw “up or down” of the market.
What to do tomorrow
- Measure your inertia. Write out the returns of your trades in order and see how alike neighboring ones are (this is called autocorrelation; it’s enough to look at a shift of one trade, and a few following ones). Close to zero — this barely concerns you, carry on as you are. Noticeably above zero — read the points below.
- Cut your sample down to the honest one. If there’s inertia, your N trades are worth less than N independent ones. Divide your confidence accordingly: the same edge is less significant than the
Sharpe × √Nformula promises. - Size the drawdown for a clumped run of losses. A smooth backtest curve understates the pit. Set your position size and stop with a margin for a run of losses in a row, not for one average loss.
- When judging reliability, reshuffle in chunks, not one trade at a time. Otherwise the interval comes out narrower than the truth, and you’ll be more confident than you should be. Take the chunk length to be about the length of a typical run.
- Don’t drop the tools from the earlier articles — they aren’t canceled, they’re refined. Count your trials (article 3), build the out-of-sample fan (article 6), but remember that both understate the spread on inertial data. You can run a strategy on real execution without risking money, in paper mode — free, before any subscription: the signal engine there is the same one used in live trading, so runs of losses are reproduced honestly rather than smoothed away.
FAQ
The backtest showed a smooth curve and an 8% drawdown, but live it was 18%. Did the strategy break? Probably not. If the average win and the share of profitable trades are the same but the pit is deeper, the losses arrived in a run. A backtest that treated trades as independent smeared that run across time and showed a drawdown shallower than the real one.
How do I tell whether I have inertia, and how strong it is? Compute how alike neighboring trades are by return (autocorrelation at a shift of one trade). Noticeably above zero — there’s inertia, and it has to be accounted for. The higher it is, the stronger everything described in this article.
Isn’t the direction of BTC over 5 minutes nearly random from window to window? The direction itself — possibly, yes. But the inertia sits not in the direction of price, but in the volatility regime, and that’s what drives your trades into runs. So what you measure is the returns of your trades, not the movement of the market.
What do I do if the inertia is large? Three things: judge significance on the smaller, honest sample; size your position for a run of losses rather than the average; and when checking reliability, reshuffle in chunks rather than one at a time.
Does this cancel the combinatorial check from the previous article? No, it complements it. Inertia was precisely the reason the fan from the previous article understated the spread. Block resampling fixes that too.
Disclaimer
This material is educational and is not financial advice. Past results do not predict future results. Trading on prediction markets carries real risk, up to and including the total loss of capital. Access to Polymarket is restricted or prohibited in some jurisdictions — verify legality where you live before trading.