Sweeps and walk-forward tests Simulated
A single backtest shows how one setting did. These show whether it would have held up: a sweep runs every combination of up to two of a strategy's numbers and measures how much of the best result is luck; a walk-forward picks the best setting on a training window, then runs it on the months after, which it never saw - and repeats, window by window.
Sign in to test a strategy.
How to read the results
- Deflated Sharpe ratio
- The chance the best combination's Sharpe ratio is really above zero, once you allow for how many combinations were tried, how widely their results spread, and how lopsided and fat-tailed its daily returns are. Try enough settings and one will look good by chance; this measure charges for the trying.
- Probability of overfitting
- The sessions are cut into ten blocks. For each of the 252 ways to pick five of them, the best combination on those five is found and ranked on the other five. The probability is how often that winner lands in the bottom half. Near 50% means picking the best told you nothing about the rest of the period.
- Peak shape
- Whether the best combination's neighbours on the grid also did well. A plateau is more believable than a lone spike.
- Walk-forward efficiency
- The annual return on the test windows divided by the chosen settings' annual return on their training windows. Well under 1 means the tuning mostly fitted the past.