How to read statistics without fooling yourself

Small samples lie

Ten trades with seven winners give 70% — a figure that looks convincing and means nothing. A fair coin lands seven heads out of ten roughly once in six attempts.

To tell an edge from luck you need at least a hundred trades. That is why the platform shows no percentage until a hundred signals have settled, and writes “collecting” instead of a pretty number.

Standard error

Every measured percentage carries an error, and the smaller the sample the larger it is. On a hundred trades the error is around five points: a measured 56% actually means “somewhere between 51% and 61%”.

The practical rule: if the measured win rate exceeds breakeven by less than two standard errors, the edge is not proven. It may exist, or it may not.

Overfitting

The most common mistake when testing strategies is to try a hundred parameter sets on one history and pick the best. It will almost certainly be luck — out of a hundred variants, one will look excellent by chance alone.

The defence is a split test: parameters are tuned on one part of the history and the result is measured on another the algorithm never saw. If the edge vanishes on the second part, it was never there.

Why other people's percentages are worthless

A service promising 90% accuracy shows neither the sample, nor the period, nor the method. Such a claim cannot be checked, which means it is worth nothing.

The only statistics worth trusting are those computed from your own closed trades. They are in your dashboard and update themselves.