Thirty Trades Is Not a Track Record

Early in a trading career there is usually one very good month. Thirty trades, a win rate somewhere in the sixties, an average result comfortably above +0.4R per trade, and a compounding calculation that takes about a minute to run out to a full year and produces a number that is genuinely worth staring at. A quiet Sunday evening, a spreadsheet, and arithmetic that makes excellent company right up until the moment somebody acts on it. Staring is harmless. Doubling the position size on the strength of it is where the money goes.

Then two losing months arrive. In absolute terms they are modest. Taken at the new size they feel like disasters, and the verdict, delivered somewhere in the second of them, is that the strategy has stopped working, which is the phrase traders reach for when the honest version would be an admission that they never established whether it worked at all.

What follows is the arithmetic that belonged in that first month, and it is also an argument for running the numbers while they can still talk somebody out of something.

The question that deflates the whole calculation is usually asked by someone with no interest in markets at all: how many trades was the figure built on?

What thirty trades can tell you

Expectancy is a sample mean, and every sample mean carries a standard error that shrinks with the square root of the sample size. The square root is the problem. Everything else in this section follows from that one fact.

Take a strategy with a true edge of +0.2R per trade. That is a respectable retail figure. Give its R multiples a standard deviation of around 1.5R, which is normal enough when the losers bunch up near −1R because the stop does its job, while the occasional winner is allowed to run to +3R or +4R. Thirty trades then produce a standard error of 1.5 divided by the square root of thirty, which comes to roughly 0.27R, and that is a wide number on which to base a decision about size.

Two standard errors either side puts the observed expectancy of a genuinely profitable strategy anywhere between about −0.3R and +0.7R. The +0.4R month sits inside that range. It also sits inside the range that luck alone produces when the underlying edge is zero, which is the awkward half of the finding, because a sample cannot separate two explanations that both predict it, and in the story above nobody asked the sample to try.

Halving the width of that interval costs four times the trades. Thirty to a hundred and twenty does not buy four times the confidence. It buys twice.

What was done, in statistical terms

A noisy estimate was read as a precise one, and the stake was raised on the strength of a precision that was never there. Sizing up after a good run feels natural. Mechanically it is the same error as sizing down after a bad one, because both use a sample too small to justify any change and both move the one variable that governs the outcome more than any other does.

What followed was the other half of a distribution that thirty trades had not been long enough to show. Held at the original size, that quarter would have finished flat and the sample would have grown by another few dozen observations. At double size it finished down. It also produced a conclusion every bit as unsupported as the one drawn from the good month, which is the part that tends to go unnoticed, since the second conclusion arrives wearing the clothes of hard-won experience.

The calculation worth doing instead

Before anything about a strategy changes, and size counts as a change, three numbers come out of the journal and go on a single line of paper. None of them needs software. The expectancy is the mean of the R multiples, the standard error is the standard deviation of those R multiples divided by the square root of the trade count, and the interval is the expectancy plus and minus roughly two standard errors.

Now read the lower bound, because the lower bound governs what happens next.

Below zero, the data has not demonstrated an edge, whatever the mean says. Above zero there is something to work with, although it remains prudent to size to the lower bound and leave the mean alone, the mean being the optimistic reading and sizing to the optimistic reading being precisely what lost the quarter described above.

The expectancy and position sizing guide carries the worked example, including the reason the Kelly formula turns this particular mistake into a larger one when it is fed a small sample.

Two uncomfortable implications

Most trading histories are too short to say anything. A hundred trades is roughly where the intervals become useful and two hundred is better still, yet a great many retail traders change strategy every few months, which guarantees that no strategy is ever measured and that the sample never reaches a size capable of answering the question that prompted the change, so the switching feels like diligence while it works as a reliable way of never finding out.

The second concerns the phrase itself. A strategy that stopped working after a bad month almost certainly did nothing of the kind, and the phrase quietly smuggles in a causal claim that the sample cannot support. Either it never worked, in which case the bad month is the truth arriving on schedule, or it works and the bad month is variance. Thirty trades cannot tell you which. More trades can, taken at a size you can survive, which is what makes small size an argument about information rather than an argument about caution, and the distinction matters most when the size feels embarrassingly small.

For anyone starting

Keep the size small for longer than feels necessary. The smallness is the price. What is being bought is knowledge of whether the edge exists at all. The currency is time spent at a size where variance cannot do real damage, and the bill falls due whether or not anybody is keeping the records that would let it be read.

Do not react to a month. Do not react to ten trades. Work out the standard error, read the lower bound, and let that number make the sizing decisions, which will be slower than instinct and considerably better at arithmetic.

Backtesting without fooling yourself is worth reading alongside this one, since the same square root that makes a live sample noisy makes a backtest optimistic, and the fix in both places is the same unglamorous one of gathering more data before believing anything.