How many trades before a backtest means anything?

July 31, 2026 · 8 min read

"You need at least 100 trades" is the usual answer, and it's wrong in both directions — far too few for a scalping strategy, more than necessary for one with a large, consistent edge. The honest answer falls out of one calculation, and it's a calculation you can run on your own results in about a minute.

The arithmetic

Your backtest gives you n trades with a mean P&L per trade of m and a standard deviation of s. The uncertainty in that mean — the standard error — is s/√n. Dividing the mean by its own standard error gives a t-statistic:

t = √n × (m / s)

The m/s term is the strategy's per-trade signal-to-noise ratio, and it is a property of the strategy, not something you choose. Rearranging for the sample size that gets you to a given t:

n = (t / (m / s))²

The square is the whole story. Halving your edge relative to its noise doesn't double the trades you need — it quadruples them.

Per-trade m/sTrades for t = 2Trades for t = 3Typical of
0.0210,00022,500High-frequency scalping
0.051,6003,600Intraday futures, tight stops
0.10400900Intraday with a real edge
0.20100225Swing strategy, wide targets
0.3045100Rare — verify before believing
Direct evaluation of n = (t / (m/s))², rounded up. Not a backtest — this is what the statistics require, independent of any particular market. t = 2 is roughly the conventional 5% threshold; t = 3 is the level to hold yourself to once you've tested multiple variants (see below).

Reading it off your own results

You don't have to guess which row you're in. Take your trade list, compute the mean and standard deviation of per-trade P&L, and divide. That ratio, plugged into the formula above, tells you how many trades this specific strategy needs — which is usually a larger number than the one you already have.

The correction nobody applies

The table assumes you tested one strategy once. If you tried thirty parameter combinations and are reporting the best, the threshold moves: the best of thirty random-but-worthless variants will clear t = 2 routinely, because you gave it thirty chances. That is what the t = 3 column is for — a rough, cheap adjustment for having searched.

Better than adjusting the threshold: hold out data from the beginning, freeze the strategy, and test once on the part you never looked at. A single honest out-of-sample test beats any correction applied after the fact.

Trade count isn't the only axis

Even a large n can be less informative than it looks, because the formula assumes trades are independent draws. Three things break that assumption:

  • Overlapping and clustered trades. Five trades taken inside the same forty-minute move are largely one observation of one event. If your strategy fires in bursts, your effective sample size is closer to the number of bursts than the number of fills.
  • Regime coverage. Two thousand trades all taken between 2023 and 2026 tell you about a particular volatility regime. A strategy that has never seen a sustained low-vol grind, or a genuine volatility shock, is untested against those regardless of trade count. This is the argument for measuring history in market conditions, not just years.
  • Skewed distributions. The t-statistic behaves badly when a handful of trades dominate the total. If deleting your best five trades materially changes the mean, the normal-approximation arithmetic above is flattering you, and you should be looking at the distribution directly rather than its summary.
Practical floor: for an intraday index-futures strategy, expect to need something in the four-figure range of trades spanning several distinct volatility regimes before the result is worth acting on. If your strategy can't generate that many trades over the history available, that is itself a finding — it means you will not be able to distinguish it from luck within any reasonable time, and it should be sized accordingly.

What to do when you don't have enough

Under-powered isn't the same as worthless. It changes what you're allowed to conclude:

  • Report a range, not a point. "Profit factor 1.4" over 120 trades is really "somewhere between losing and quite good" — say that.
  • Loosen the strategy until it trades more, then test. A less selective version with 1,500 trades often teaches you more about whether the underlying effect is real than the tuned version with 90.
  • Test the same idea on correlated instruments. ES and NQ are not independent, but a result that appears on both is more credible than one that appears on either.
  • Size for the uncertainty. A strategy you can't distinguish from noise gets the position size of a hypothesis, not a conviction.

For the biases that make an under-powered result look better than it is, see why your backtest lies, and for what the resulting numbers should plausibly look like, what a realistic profit factor looks like.