How Many Backtest Trades Before You Go Live?

Ask how many trades a backtest needs and you get a ladder. Twenty is noise, fifty is a hint, one hundred is meaningful, two hundred is confidence. It is repeated everywhere, it was repeated on two pages of this site until we rewrote them today, and nobody who quotes it shows where the numbers came from.

They came from nowhere. Sample size is not a property of backtesting, it is a property of your strategy, and two strategies can differ by a factor of twenty. This post does the arithmetic instead of the ladder.

Why One Number Cannot Be Right

You are trying to answer a specific question: is the result I just measured real, or is it what a strategy with no edge at all would produce on a good run?

That question has two moving parts.

How big your edge is. The average result per trade, measured in R rather than dollars so position size cannot distort it. Half an R a trade is a large edge. Four hundredths of an R is a real one, and it is nearly invisible.

How spread out your results are. A strategy that produces +1R and −1R over and over has tight, well-behaved results. A strategy that produces one +5R for every three −1R is far noisier, even though it might be the better strategy.

A big edge in tight results shows up fast. A small edge in wild results can hide for thousands of trades. No fixed count can serve both, which is why every ladder you have read is wrong for most of the people reading it.

The One Line of Arithmetic

Write your average result per trade in R as E, and the standard deviation of your per-trade R outcomes as s.

The uncertainty in a measured average is the standard error:

As you add trades, uncertainty falls, but slowly: it takes four times the sample to halve it.

You want your measured edge to sit at least two standard errors above zero, the usual two-sigma bar that corresponds to roughly 95% confidence under a normal approximation. So you need:

Rearranged, that is the whole answer:

Two inputs, one number. Now watch what it does to four real strategy shapes.

Four Strategies, Four Very Different Answers

Each of these risks 1R per trade, so a loss is −1R and a win is the R multiple of the target. Writing the win rate as p and the reward multiple as R, both inputs come out of the same pair of formulas every time:

Strategy A. Scalper, 1:1, wins 52%. Average per trade: 0.52 − 0.48 = 0.04R. Average squared outcome: 1.0, so the variance is 1.0 − 0.0016 = 0.9984, and s = 1.00R.

Strategy B. Intraday setup, 1:2, wins 40%. Average per trade: (0.4 × 2) − 0.6 = 0.20R. Average squared outcome: (0.4 × 4) + (0.6 × 1) = 2.2, so the variance is 2.16 and s = 1.47R.

Strategy C. Swing runner, 1:5, wins 25%. Average per trade: (0.25 × 5) − 0.75 = 0.50R. Average squared outcome: (0.25 × 25) + (0.75 × 1) = 7.0, so the variance is 6.75 and s = 2.60R.

Strategy D. Strategy B, but the win rate is really 38%. Average per trade: (0.38 × 2) − 0.62 = 0.14R. s works out at 1.46R.

Three things fall out of that table, and all three matter more than the ladder ever did.

The famous 100 only fits one row. It is roughly right for the swing runner, half of what the intraday setup needs, and off by a factor of twenty-five for the scalper. Anyone who went live on 100 scalping trades has measured almost nothing.

Higher R needs fewer trades, not more. This surprises people, because the swing runner's results are the most spread out of the four. But sample size is decided by the ratio of edge to noise, and going from 1:2 to 1:5 multiplied the edge by two and a half while the spread only went up by three quarters. The edge grew faster than the noise.

Two points of win rate doubled the requirement. Rows B and D are the same strategy, measured slightly differently. A 40% win rate that is really 38% needs twice the sample. This is the honest reason thin edges are hard: not that they fail, but that they take so long to confirm that most traders abandon or oversize them first.

The 2 Is a Choice, and It Is the Permissive One

That 2 is the two-sigma bar. It sets how sure you want to be that a positive result is not noise. What it does not set is how likely you are to see that result in the first place, and those are different questions.

Read the formula literally. n is the point where a measured edge of E would just clear the bar. If your true edge is exactly E, then in a fresh sample of n trades you land above the line about half the time and below it about half the time. Clearing the bar at n is a coin flip, not a confirmation. Every number in the table above is a floor.

Statistics handles this with a second term, called power, and the convention is 80%. It raises the multiplier from 2 to 2.8:

which works out at almost exactly double the sample.

A third bar matters if you arrived at this strategy by testing several. Harvey, Liu and Zhu showed that once many candidates have been tried, a t-statistic of 2 stops meaning what it appears to mean, because the best of twenty coin flips also looks skilled. Their threshold for a genuinely new factor is 3, which multiplies the floor by 2.25.

Pick your column honestly. If this is the only strategy you have ever tested and the question is whether to keep going, the floor is a fair answer. If you compared six setups and are about to fund the winner, you are in the third column, and the scalper in row A is asking for over five thousand trades before it has told you anything.

Four Things This Formula Does Not Know

The arithmetic is exact. The assumptions underneath it are not, and every one of them fails in the direction of needing more trades, not fewer.

1. It assumes your average behaves normally, and that is weakest exactly where the numbers look best. Strategy C's 108 leans on the central limit theorem, and convergence is slow when three quarters of your trades are small losses and the entire result comes from an occasional large win. Treat a low number from a high-R strategy as a floor to clear, not a target to stop at.

2. You have to plug in the thing you are trying to measure. You do not know your true win rate; you only know what the sample says so far. That is circular, and the working fix is to recompute the requirement as you go. If the number of trades you need keeps climbing while you add trades, your measured edge is drifting toward zero and the sample is telling you something.

3. It assumes trades are independent. Four positions open at once on correlated pairs during one London session are close to a single observation. So are twenty trades taken in one trending fortnight. The formula counts rows; edge lives in independent decisions.

4. It only tests whether your edge is above zero. Clearing the bar means you probably have something. It does not mean the number you measured is the number you will get. Sizing a live account off a measured 0.5R when the truth is 0.2R is a separate and more expensive mistake than going live too early.

Three Tests That Beat Counting

Once you are near the number, stop looking at the count and run these instead. Each one takes a minute and each one has ended a strategy that had passed on volume alone.

Delete your best trade and recompute. If a single winner is carrying the result, you have one good trade and a lot of filler. The expectancy article works this through: a single 5R winner supplies 0.17R of expectancy all by itself in a 30-trade sample, which is most of a typical edge.

Split the sample by date and compare halves. First half against second half, on the same rules. A real edge is boring across both. A curve-fitted one usually lives in one period, and it is almost always the period you replayed first, when you already knew what the market did next.

Count conditions, not trades. Two hundred trades taken from six weeks of one clean trend is one observation repeated two hundred times. Ask how many distinct months, sessions and market states your sample covers. Two hundred trades spread across a trend, a range and a news-driven month is worth more than a thousand from a single quarter. If you want your sample to include the days that break strategies on purpose, backtesting through NFP, CPI and FOMC releases is where those days live.

And when the sample finally passes: the count says the edge is real, it says nothing about the order the wins and losses will arrive in. A positive edge can still fail an evaluation on a bad sequence, which is a drawdown problem rather than a sample problem, and the arithmetic for it is in how to pass a prop firm challenge.

What That Means in Calendar Time

This is the part that decides whether anyone ever collects a real sample.

Take Strategy B and its 216 trades. At two setups a week, that is 108 weeks. Slightly over two years of live trading before the number means anything, and that assumes the strategy, the market and your execution all stay still for two years, which none of them do. At the honest bar of 424 it is four years, which is longer than most traders keep the same strategy.

That is not an argument against sample size. It is the argument for bar replay. Replaying historical price bar by bar produces the same 216 decisions in a handful of afternoons, on markets and periods you choose deliberately rather than whichever ones happened to arrive.

Two conditions make those trades count:

No hindsight. Start from a date you have not studied, reveal one bar at a time, and take the setup when it appears rather than when you remember what followed. The bar replay backtesting guide covers the method, including the mistakes that quietly inflate the exact number you are measuring.

Same tagging as live. Entry, exit, stop, costs, and the conditions behind the trade. Untagged replay trades give you a win rate and nothing to do with it. Tagged ones let you ask which condition is carrying the edge, and profit factor and expectancy can then be computed on the same footing as your live results.

One practical note on tooling: TradingView's Bar Replay is restricted to daily and higher timeframes on the free plan, and intraday replay depth is tied to your subscription tier, which caps how large a sample you can build on the timeframe most day traders actually use. We broke the tiers down in TradingView bar replay limits. Checked against TradingView's own Bar Replay support documentation on 7 September 2026.

Collecting the Sample in TradingSFX

The backtester is open on every plan, including the free one.

• Replay with your own data. Import a CSV from any broker or data source and replay it bar by bar. The file is parsed in your browser and the candles never touch our servers, so the market coverage is whatever you can get data for rather than whatever a vendor licenses.
• Practice trades with a real R outcome. Buy and sell with draggable stop and target, and the R result on close, on every plan. Saving a backtest trade into your journal is a Pro feature. Below Pro the close form still runs in full, and the last 10 practice trades are held for 30 days so they can be imported into the journal if you upgrade rather than being lost.
• Backtest trades sit in the same statistics as live ones. Nothing in the app filters them out, so expectancy, win rate, profit factor and the confluence breakdowns are computed across both. That is deliberate: a sample you cannot compare to your live results is not much of a sample.
• Tag every replay trade with its conditions. The same confluence checklist runs in the close form, which is what turns a win rate into an answer about which condition produced it.
• Custom timeframes in the backtester are on Pro and above.

You can run the whole thing on sample data with no account and no sign-up in the backtester demo.

Bottom Line

The number of backtest trades you need is not 100. It is (2s ÷ E)², and until you have put your own two numbers into it you do not know whether you are looking at 108 or 2,496. The calculator at the end of this page does that arithmetic for your own win rate and reward multiple, gives you the floor and the honest number side by side, and turns both into how long they would take to collect live.

Work it out before you start, so you know what you are committing to. Recompute it as the sample grows, because a requirement that keeps rising is a strategy quietly telling you its edge is smaller than you hoped. Then, before going live, delete your best trade, split the sample in half, and count the market conditions rather than the rows.

If all three survive, the count was the easy part.

TradingSFX Basic is free forever at 10 trades a month, and the backtester is included on it, which is enough to start building the sample and computing your own two inputs.

Published September 7, 2026. Updated September 9, 2026 to state the statistical power the formula assumes and to add the 80% power and t = 3 columns. Every figure here is arithmetic worked openly from the stated inputs, and the sample-size formula itself is the standard one-sample calculation, equivalent to requiring a t-statistic of 2. No survey data or user statistics are cited. Two external claims: TradingView's Bar Replay plan restriction, verified against TradingView's own support documentation on 7 September 2026, and the t = 3 multiple-testing threshold from Harvey, Liu and Zhu, "... and the Cross-Section of Expected Returns", Review of Financial Studies 29(1), 2016. No competitor pricing is quoted.