What is the runs test — and how do you tell whether your losing streaks are real?
Every trader has had the week where it felt as if the market had learned their setup. Five, six, seven losses in a row. The question that week is never asked out loud, but it decides what you do next: is this bad luck that will pass, or is something real going on? Your eye cannot answer it — as you will see below, pure luck produces streaks long enough to fool anyone. A two-minute calculation on your trading journal can. This page shows you how, how many trades it needs before it can see anything, and what the answer should change about your sizing.
KEY TAKEAWAYS
- Pure luck streaks hard. With a 45% win rate and every trade independent, a 50-trade journal has a 46.5% chance of containing six or more losses in a row, and a 200-trade journal an 8-loss streak at the median.
- The runs test counts how many times your results switch from win to loss or back, and compares that with what luck would give. Too few switches means your losses clump.
- Our 40-trade worked example has 14 runs where luck predicts 20.6: a z-score of −2.15, the kind of gap luck produces about 1.6% of the time.
- Strong clumping is visible early (caught 75% of the time at 50 trades). Mild clumping is not (28% at 50 trades, 71% at 200).
- When losses do clump, “halve the size after three losses” cut the chance of a 50% drawdown from 24.7% to 2.6% and raised the median result. When they don’t, the same rule lowered the median result from 1.93× to 1.83×.
What is the runs test, in one paragraph?
The runs test is a simple statistical check, published by the mathematicians Abraham Wald and Jacob Wolfowitz in 1940, for whether a sequence of two outcomes is in random order. Write your trades out as W and L in the order you took them. A run is each unbroken block of the same letter: LLL WW L WWW is four runs. If your results are independent — if the last trade tells you nothing about the next — then, given how many wins and losses you had, there is an expected number of runs and a known spread around it. Far fewer runs than expected means wins and losses are stuck together in clumps. Far more means they alternate suspiciously, which is rare in trading and usually a sign of something mechanical.
It matters because nearly every risk number you have ever seen assumes independence. The 1–2% rule, our risk of ruin simulator, the Kelly criterion: all of them treat each trade as a fresh coin. Our risk of ruin page shows that clumped results can multiply ruin by 81 at the same win rate. That page shows what clumping does. This one shows how to find out whether you have it.
Why can’t you trust your eye on streaks?
Because random sequences are far streakier than people expect, and the streaks you remember are the painful ones. Here is a system that wins 45% of the time, every trade independent of the last — no clumping at all, by construction — and the chance that a journal of each length contains a losing streak at least as long as the one in each column. 200,000 simulated journals per row.
| Journal length | Median longest losing streak | 4+ in a row | 6+ in a row | 8+ in a row | 10+ in a row |
|---|---|---|---|---|---|
| 30 trades | 5 | 76.4% | 29.7% | 8.9% | 2.5% |
| 50 trades | 5 | 91.6% | 46.5% | 15.6% | 4.7% |
| 100 trades | 6 | 99.4% | 73.0% | 30.7% | 10.1% |
| 200 trades | 8 | 100% | 93.1% | 53.0% | 19.8% |
Read the bold cell twice. Nearly half of all perfectly honest, perfectly independent 50-trade journals contain six losses in a row. Over 200 trades — a few months for an active swing trader — an eight-loss streak is the typical outcome, and one journal in five sees ten. None of those streaks means anything. They are what a 45% coin looks like.
So a long streak, on its own, is not evidence of a problem. That is the first half of the lesson, and our 1–2% rule lesson builds the whole risk rule on it. The second half is that the opposite mistake is just as expensive: deciding every streak is luck when your results really do clump. A streak tells you nothing. The pattern of switches across the whole journal does.
How do you run the test on your own journal?
You need four numbers, all of which you can count by hand or in one spreadsheet column. Here is a 40-trade journal, invented for this example, written out in order:
Step one — count. L, the number of losses: 23. W, the number of wins: 17. n, the total: 40. R, the number of runs: 14. The quickest way to count runs is to count the switches between one trade and the next and add one; here there are 13 switches.
Step two — work out what luck would give. For a random order of 23 losses and 17 wins, the expected number of runs is
Spread (standard deviation) = √[ 2LW(2LW − n) ÷ (n²(n − 1)) ] = √[ 782 × 742 ÷ 62,400 ] = 3.05
Step three — compare. The z-score is the gap in units of spread: (14 − 20.55) ÷ 3.05 = −2.15. A negative z means fewer runs than luck, which is clumping. The usual one-sided cut-off is −1.645: past it, luck alone would produce a gap that size less than 5% of the time. At −2.15 that figure is about 1.6%. This journal’s losses clump.
In a spreadsheet with W and L in column A, the switch count is =SUMPRODUCT(--(A3:A41<>A2:A40)), and the rest is three cells of arithmetic. Our trading journal exports your trades as CSV in the order you logged them, which is exactly the column this needs. Two practical rules when you prepare it: put trades in the order they closed, because that is the order your account felt them; and decide in advance what to do with break-even trades — either drop them or count them as losses, but do it the same way every time, before you look at the answer.
Isn’t “how often does a loss follow a loss” enough?
It is the most intuitive version of the same question, and worth computing alongside, but on its own it is noisier than it looks. In the example journal every one of the 23 losses is followed by another trade, and 16 of those 23 next trades were losses again: a loss followed a loss 69.6% of the time, against an overall loss rate of 57.5%. After a win, the next trade lost only 37.5% of the time. That looks damning.
Now look at how much that number wobbles when there is no clumping at all. We simulated 200,000 independent journals at a 45% win rate, where the true chance of a loss after a loss is exactly 55%:
| Journal length | True chance of a loss after a loss | Middle 90% of what journals actually show |
|---|---|---|
| 50 trades | 55% | 37.5% to 69.0% |
| 200 trades | 55% | 46.7% to 62.4% |
At 50 trades, an honest coin shows anything from 37.5% to 69% one time in twenty either side. Our example’s 69.6% sits right at that edge. The single conditional percentage, read alone, can barely tell clumping from noise at this size. The runs test does better on the same data because it uses every switch in both directions — wins sticking to wins as well as losses sticking to losses — and it already adjusts for how many wins and losses you happened to have. Use the conditional percentage to describe the pattern and the runs test to decide whether there is one.
How many trades before the test can see anything?
That depends on how strong the clumping is, and the honest answer is: more than most people log before drawing conclusions. We generated journals in which a loss follows a loss with a fixed probability, tuned so that the overall win rate stays exactly 45% — so the only thing that differs from pure luck is the order. Then we ran the one-sided runs test at the 5% cut-off on each. 40,000 journals per cell.
| A loss follows a loss… | 20 trades | 30 | 50 | 100 | 200 | 300 | 500 |
|---|---|---|---|---|---|---|---|
| 55% (pure luck) | 5.4% | 5.1% | 5.4% | 5.0% | 5.0% | 4.8% | 5.0% |
| 62% (mild) | 16.9% | 20.1% | 28.3% | 44.5% | 70.8% | 85.3% | 96.6% |
| 65% | 24.4% | 31.4% | 45.8% | 70.0% | 93.3% | 98.6% | 100% |
| 70% (strong) | 41.0% | 53.7% | 75.3% | 95.2% | 99.9% | 100% | 100% |
| 75% | 59.2% | 75.3% | 93.0% | 99.7% | 100% | 100% | 100% |
The top row is the test keeping its promise: on journals with no clumping it raises a false alarm about 5% of the time at every length, which is what the cut-off is for. The rest of the table is the part to plan around. Strong clumping — the kind where a bad week genuinely tends to stay bad — shows up in a 50-trade journal three times out of four. Mild clumping, 62% instead of 55%, is missed nearly three times out of four at 50 trades and needs around 200 before the test catches it more often than not.
That gives a practical reading of a result. A clear negative z at 50 trades is worth acting on: whatever is causing it is strong. No signal at 50 trades rules out strong clumping but not mild clumping — keep logging and run it again at 100 and 200. A trader who checks once at 30 trades and concludes “my results are random” has learned very little.
What should you do if your losses do cluster?
Use it, because clumping is one of the few patterns a trader can turn directly into a sizing rule. Here is the test. A system that wins 45% of the time and pays 1.5 times what it loses, 2% of current equity at risk per trade, 300 trades, 100,000 simulated runs. “Ruin” is a 50% drawdown from the running peak, the same definition as our risk of ruin simulator. We compare flat 2% sizing with one simple rule: after three losses in a row, risk 1% until the next win.
| How often a loss follows a loss | Sizing | Chance of a 50% drawdown | Median worst drawdown | Worst drawdown, 1 run in 10 | Median ending account |
|---|---|---|---|---|---|
| 55% — pure luck | Flat 2% | 1.4% | 25.5% | 38.3% | 1.93× |
| Halve after 3 losses | 0.7% | 23.8% | 35.7% | 1.83× | |
| 65% — moderate clumping | Flat 2% | 7.7% | 32.3% | 47.9% | 1.93× |
| Halve after 3 losses | 1.5% | 26.3% | 38.9% | 2.12× | |
| 75% — strong clumping | Flat 2% | 24.7% | 40.7% | 59.1% | 1.93× |
| Halve after 3 losses | 2.6% | 28.5% | 41.7% | 2.75× |
With strong clumping the rule is the best kind of risk control there is: it cuts the chance of losing half the account from roughly one in four to one in forty and lifts the median result from 1.93× to 2.75×. That is not magic. When losses stick together, the trade after three losses is genuinely more likely to lose than an average trade, so betting less on it is betting less on worse odds. The rule is simply reading the pattern the runs test found.
Two refinements once you know you clump. First, the trigger: three losses was a round choice, not an optimised one. Look at your own journal — after how many consecutive losses does the next trade’s loss rate actually rise? Set the trigger there, and do it from data you collected before choosing it. Second, ask why. Clumping in trading usually has a cause you can name: a strategy that works in trends and bleeds in ranges, or a trader who sizes up or trades faster after losses. The first is a market problem and a trend-or-range filter may fix it better than a sizing rule. The second is a behaviour problem, the one our revenge trading lesson measures, and the trade autopsy tool is the quickest way to see whether the losing trades in a clump share a cause.
And if they don’t?
Then the same rule is a ritual with a price tag. Look at the top two rows of the table again. With independent results, halving after three losses still roughly halves the chance of a 50% drawdown — but only because it lowers your average size, and a plain lower size does that job far better: flat 1% sizing on the same system took the chance of a 50% drawdown to effectively zero (0.0%). Meanwhile the rule cut the median result from 1.93× to 1.83×, because under independence the trade after three losses is exactly as good as any other trade, and the rule bets less on it for no reason.
This is why the test is worth the two minutes. “Step back after a bad run” is advice you will hear everywhere, and it is right for some traders and quietly expensive for others. Which one you are is not a matter of opinion. It is in your journal.
What the runs test is NOT
It is not a test of whether your system has an edge. A journal can be perfectly random in order and still lose money, or clump badly and still be profitable. Edge is measured by expectancy; the runs test only looks at sequence.
It is not a prediction of the next trade. Even with strong clumping, a trade after three losses in our 75% scenario still wins a quarter of the time. The test tells you the odds have shifted, not that the next trade will lose.
It is not the “hot hand” argument in reverse. The famous 1985 basketball study by Gilovich, Vallone and Tversky asked whether success breeds success; later work found the original method was biased in small samples. The same trap exists here, which is why this page leans on a test with a known false-alarm rate rather than on eyeballing conditional percentages.
It is not a substitute for position sizing. If your flat size is too large, no streak rule rescues it. Start from position sizing and the position size calculator; the runs test only fine-tunes what happens around a losing run.
Where this reasoning breaks down
1. The model of clumping is the simplest one. We let each trade depend only on the one before it. Real clumping can come from market regimes lasting weeks, which a runs test still detects but which a three-loss trigger may react to too slowly. If your journal shows clumping, plot your losses by date as well; long dry spells point to regimes, short bursts point to behaviour.
2. Changing the rules mid-journal breaks the test. If you changed strategy, timeframe or size halfway through, a run of losses after the change is not clumping — it is a different system. Run the test only on a stretch where the rules were constant.
3. Looking for patterns until one appears. Test wins and losses, then long trades only, then Mondays only, then one setup, and one of them will clear the 5% cut-off by luck roughly one time in twenty. Decide the one question before you look, and treat any second test as a hypothesis for the next 100 trades, not a conclusion.
4. When the advice is wrong: very high-frequency or very few trades. With hundreds of trades a day, overlapping positions make “the order trades closed in” ambiguous and the test loses meaning. With fewer than about 20 trades, the formula’s normal approximation is rough and the test can barely detect anything, as the table in section five shows.
5. When the advice is wrong: the rule has a cost you cannot see in a table. Some traders find a mechanical size cut after losses calms them and keeps them executing; for them the median-return cost under independence may be worth paying. That is a legitimate choice — as long as it is made knowing the price, not believing it is free.
Where should you go from here?
Export your trades from the trading journal, write them out as W and L in the order they closed, and count the runs. If you have fewer than 50 trades, keep logging and put a reminder in for when you reach 50, 100 and 200. Then read what risk of ruin is for why clumping matters so much to survival, position correlation for the other way losses arrive together — across open trades at the same moment — and the trading journal lesson for the columns that make a test like this possible in the first place.
FAQ
What is the runs test in trading?
It is a statistical check on whether your wins and losses come in random order. You count the runs, meaning the unbroken stretches of wins or of losses, and compare that count with what luck would produce for the same number of wins and losses. Too few runs means your losses tend to clump together.
How long a losing streak is normal?
Longer than most people think. For a system that wins 45% of the time with independent trades, the median longest losing streak is 5 in 50 trades and 8 in 200 trades, and a 50-trade journal has a 46.5% chance of six or more losses in a row. A long streak alone is not evidence that anything is wrong.
How do I calculate the runs test?
Count losses L, wins W, total n and runs R. Expected runs are 2LW/n + 1, and the standard deviation is the square root of 2LW(2LW − n) divided by n squared times (n − 1). The z-score is (R − expected) divided by the standard deviation. Below −1.645 means fewer runs than luck would give 95% of the time, which suggests clumping.
How many trades do I need for the runs test?
About 50 to catch strong clumping reliably. In our simulations, losses that follow losses 70% of the time instead of 55% were detected 75% of the time at 50 trades, while milder clumping at 62% was detected only 28% of the time at 50 trades and 71% at 200.
Should I reduce my position size after a losing streak?
Only if your results actually clump. With strong clumping, halving size after three losses cut the chance of a 50% drawdown from 24.7% to 2.6% and raised the median result. With independent results, the same rule lowered the median result from 1.93 to 1.83 times the starting account, and a plain smaller size reduced drawdown risk more.