A backtest finishes and prints one number: Sharpe 1.4. The scale is familiar — above 1 is usually called acceptable, above 2 is called good — and that familiarity is the risk. A Sharpe ratio is a quotient of two quantities you did not choose yourself: the average return above the risk-free rate, and the volatility of those returns, both computed over whatever sample the test happened to cover.
This article covers what the ratio actually measures, what different platforms put behind the same label, and why a high value in a backtest can be the product of a short sample, missing costs, or a search that ran more times than you remember. The point is not to discard the metric. It is to read it next to a threshold instead of alone.
What the Sharpe ratio measures
The Sharpe ratio is excess return per unit of total risk. In daily form it is the mean daily return above the risk-free rate divided by the standard deviation of those returns, then scaled to a year by multiplying by the square root of the number of trading days — √252 for most equity and crypto markets.
SR_annual = mean(R - Rf) / stdev(R) * sqrt(252)
It is not a win rate. A system that closes most trades for small gains and occasionally absorbs one large loss can show a poor Sharpe; a system that loses most of its trades by a few ticks while catching rare large trends can show a good one. The ratio sees the distribution of period returns. It does not see trade count, hold time, or hit rate.
A worked number
Take a daily strategy with a mean excess return of 0.06% per day and a daily standard deviation of 1.5%. The daily ratio is 0.06 / 1.5 = 0.04, and annualized that is 0.04 × 15.87 ≈ 0.63. Now subtract a realistic cost drag of 0.03% per day in commission and slippage — roughly one round trip every three trading days at common retail rates. The numerator halves to 0.03 and the annualized ratio falls to 0.02 × 15.87 ≈ 0.32. The rules never changed; only the accounting did, and the number halved.
| Component | Value in the example | What moves it |
|---|---|---|
| Mean excess return | 0.06% per day | Regime, signal quality, whether costs are subtracted |
| Standard deviation | 1.5% per day | Position sizing, leverage, instrument volatility |
| Annualization factor | √252 ≈ 15.87 | Assumed number of trading days per year |
| Resulting Sharpe | 0.63 gross, 0.32 net | All of the above at once |
The same label, computed differently
Two platforms can show different Sharpe values for the same equity curve, because the label hides at least three choices: the risk-free rate, cost handling, and fill timing.
TradingView's strategy() function assumes a risk-free rate of 2% by default; Composer computes Sharpe with Rf = 0. On a strong year the difference silently removes 2 percentage points of return from the numerator; on a flat year it turns a slightly negative result into a clearly negative one. Neither setting is wrong. They are different metrics sharing one name, and a comparison across platforms that does not state the assumption compares nothing.
Costs are the second fork. Some engines subtract commission and slippage only when you fill in the fields; others ignore them entirely. Fill timing is the third and the most dangerous: a backtest that buys at the close of the same bar whose close triggered the signal used information a live order could not have had. The resulting Sharpe belongs to a market that does not exist, and it survives every other check you run.
| Hidden choice | Conservative setting | Effect on the reported Sharpe |
|---|---|---|
| Risk-free rate | Stated explicitly, not defaulted | Moves the numerator by the full spread to zero |
| Commission and slippage | Entered per side | Often the largest single deduction |
| Fill price | Next bar open after the signal | Removes same-bar look-ahead profit |
| Sample | Years, not weeks | Shrinks the error around the estimate |
Where the 1 / 2 / 3 folklore comes from
The familiar bands — under 1 mediocre, 1 to 2 good, above 2 excellent — come from fund-industry heuristics built on monthly net returns of live managers: decade-long track records, real capacity constraints, audited costs. The bands were rough even there. Applied to a two-year retail backtest they lose the properties that made them meaningful. The sample is short, the returns are daily or intraday, costs are optional inputs, and the strategy was not selected by a committee — it was selected by you, out of a set of variants you tried.
A backtest that "looks like 2" is therefore not the same object as a manager running a 2. The digits match; the sampling error, the selection process, and the cost model do not. That is why the useful question is not whether a Sharpe is good. It is good relative to which threshold, produced by how many trials, over how long a sample. A plain ratio cannot answer any of the three.
Why a high Sharpe can be noise
Four mechanisms inflate a backtest Sharpe without any improvement in the underlying rules. They stack, and in practice they usually stack together.
1. The sample is short
The Sharpe estimate has its own standard error, roughly sigma_SR = sqrt((1 + SR^2 / 2) / T) for T periods. On one year of daily data with a claimed Sharpe near 1, that error is about 0.08, so a reported 1.15 carries around ±0.15 of pure sampling noise before any other objection is raised. Two years narrows the band. Two months makes the number decorative.
2. Costs are missing
The worked number above turned 0.63 into 0.32 on a modest cost drag. For strategies that trade often, commission, spread and slippage are not a rounding item — they are frequently the entire difference between the tested edge and no edge. The full list of what a backtest quietly skips lives in backtest costs.
3. Fills happen on the signal bar
Same-bar execution lets the strategy transact at the same price whose information triggered the trade. It inflates both hit rate and Sharpe, and it vanishes on the first real fill. The gap between the two worlds is measurable rather than theoretical — see backtest versus live trading.
4. The number was selected
This is the largest of the four. Bailey and Lopez de Prado showed that with five years of daily data, trying more than roughly 45 independent variants already makes a Sharpe of 1 or higher likely to appear by chance alone. Most traders do not run 45 deliberate experiments. They vary a few parameters, notice which combination looks best, and keep it — which is the same act with a smaller vocabulary. How the count is kept, why raw runs collapse into fewer effective trials, and how the bar moves with them is the subject of how many backtests is too many.
The deflated bar in one screen
The correction does not require the full deflated-Sharpe machinery to be usable. A rough bar does the job: the standard error of the Sharpe multiplied by √(2 · ln N_eff), where N_eff is the number of effectively independent variants you searched. The expression estimates how much of an observed Sharpe can be manufactured by taking the best of N draws out of noise.
As an interface example — figures the interface is designed to display, not results from a live account — a run reporting Sharpe 0.94 against a required 1.47 at 11 effective trials reads green next to zero and red next to the bar. The metric did not change between those two readings. What changed is the reference point, and only the second reading carries information about whether the result is likely to survive.
Sample length does the same work from the other side: the same search over a shorter sample produces a wider noise band, so the required bar rises even at a fixed number of trials. The check that cuts through both problems is refusing to judge on the data that selected the rules in the first place — what in-sample selection cannot tell you.
Sharpe vs Sortino, briefly
Sortino replaces the denominator with downside deviation, counting only returns below a target. On a strategy with skewed returns — many small wins, rare large losses — the two separate noticeably, because Sortino stops penalizing upside jumps that the Sharpe books as risk. Neither metric knows how many variants were tried. Sortino answers how bumpy the path was, not how likely the result is to be luck, and swapping one for the other adjusts nothing about N.
| Sharpe | Sortino | |
|---|---|---|
| Denominator | All volatility | Downside deviation only |
| Punishes upside spikes | Yes | No |
| Adjusts for the number of trials | No | No |
| Useful when | Comparing runs under identical assumptions | Returns are asymmetric |
How to check this in Normal Trading
Six checks before a Sharpe value earns any trust:
- Read it next to the threshold, never alone. A value above zero says almost nothing once trials are counted.
- Confirm commission and slippage are inside the run, not deferred to later.
- Confirm fills happen after the signal bar, not on it.
- Check how many runs produced the number, and whether near-identical variants were collapsed into effective trials.
- Check what the sample contains. A bull-only window measures almost nothing about the strategy's behavior in the regime you will actually trade.
- Compare the same run's maximum drawdown against your capital limit before any discussion of return per unit of volatility begins.
Then do one comparison run: open the Replay simulator, execute a saved graph once as saved and once with commission and slippage enabled, and read how much of the Sharpe survives the second pass. Whatever disappears there was never part of the strategy — it was part of the report.