How many months of backtest or track record are enough?
At a true annualised Sharpe of 1, about 48 months to reach t = 2 for a single result; at Sharpe 2, 12 months. If the result is the best of many tries, it needs more: the best of 100 needs t ≥ 3.51, which at Sharpe 1 is 12.3 years.
The formula
This is Lo (2002) under i.i.d. returns, so it is a floor: autocorrelated returns need longer.
The table
| True Sharpe | One result (t ≥ 2) | Best of 100 (t ≥ 3.51) | Best of 48,407 (t ≥ 4.90) |
|---|---|---|---|
| 0.5 | 16.0 years | 49.2 years | 96.2 years |
| 1 | 4.0 years | 12.3 years | 24.0 years |
| 2 | 12 months | 3.1 years | 6.0 years |
| 3 | 5 months | 16 months | 32 months |
Which months count
Only months the result was not fitted on. For a backtest run with a language model, months before the model’s training cutoff cannot be ruled out as seen; only months after it are clean. Check a backtest splits the window for you. For a trader you are about to copy, only the record after the day you picked them tests the pick.
Count every variant tried, including ones that were dropped. An agent can try fifty in a minute; the plugin in the repository counts them as it goes.
Related questions
- Is this Hyperliquid trader skilled or just lucky?
- How many months of backtest or track record are enough?
- Why isn't the #1 trader on a leaderboard evidence of skill?
Numbers rebuild daily from latest.json. Method: Method. Not investment advice.