Sample-size check

How many trades does it take to prove an edge? There is no universal number. The trades required scale with the square of your edge's dispersion-to-size ratio, so a coarse edge shows itself in hundreds and a fine one hides for thousands. This computes your own figure — the luck horizon — from the log you already keep, and says whether the record you have is evidence yet or still consistent with luck.

Why there is no universal number

"Does my strategy work?" is, statistically, "is the average of my per-trade results distinguishable from zero, given how much they vary?" Each trade is one noisy observation of your edge.

The signal is the mean result per trade; the noise is the spread around it; and a record's power to separate them grows only with the square root of its length. Halve the edge and you need four times the trades. That is why every fixed number in this genre — a hundred trades, two hundred, five hundred — is wrong in both directions at once: enough for a coarse edge, nowhere near enough for a fine one. The reasoning in full is how many trades prove a trading edge, of which this page is the arithmetic.

What this computes

From your log: the mean result per trade in R, the standard deviation of those results, and their ratio — the trade quality, your edge's signal-to-noise per trade. The evidence a record carries is that quality times the square root of its length, and the conventional bar for taking a result seriously sits near two. Rearranged, the luck horizon is four divided by the trade quality squared: the number of trades at which a real edge of your recorded size would typically clear the bar. A quality of one fifth puts it near a hundred trades; one twentieth, nearer sixteen hundred.

The interval is placed on the mean result, not on the win rate, because the win rate is the wrong object: it cannot see the shape of frequent small wins and rare large losses that produces a proud percentage and a negative expectancy at the same time. If all you have is a win rate and a payoff ratio, the summary input builds the smallest dispersion those two numbers allow — every win exactly the payoff, every loss exactly one R — so the horizon it returns is a floor; a real log disperses more and needs more trades.

Where the sample cannot support the arithmetic — fewer than twelve pasted results, a series with no spread, a mean at or below zero — the panel refuses and says why, rather than printing a number with nothing behind it.

What this does not do

It assumes your trades are comparable draws from one process. A strategy revised mid-record restarts its own clock, and positions held at the same time in correlated instruments count as fewer effective observations than the row count suggests — so the horizon here is optimistic for any record with either. It prices only the recorded past: an edge measured over one regime says nothing yet about the next, which is the harder question of regime shift.

A record past its horizon establishes that the recorded process had an edge over the recorded period. It does not establish that the record was honestly selected: a strategy chosen from many tried variants carries a multiplicity this arithmetic cannot see, which is the domain of backtest overfitting and data snooping. And results in R are net of whatever you netted out of them — if execution costs are not in the log, they are not in the answer.

The survival question this leaves open — whether an account can afford the distance to its horizon — is the risk of ruin calculator, and the stake that record supports is the Kelly criterion calculator. The same sample-size honesty applied to a returns series rather than a trade list is the Sharpe ratio calculator. Whether a record that has cleared its horizon was also honestly selected is what the Overfit Auditor bounds. All the calculators are listed under calculators. And when the inputs cannot support an answer, the panel above refuses and says why rather than rendering a zero — why an instrument refuses to answer is that design, written down.