How StrategyProof tests a strategy, and how to read every result.
StrategyProof is built on one idea: a strategy is not defined by one backtest. An optimizer can always find
settings that look perfect on past data. The question is whether the strategy keeps working on data it has never seen.
1. Walk-forward testing
The history is split into consecutive windows. In each window the strategy is optimised on a
TRAIN period (for example 12 months) and then judged, without any change, on the following
OOS period (out-of-sample, for example 3 months). The window then moves forward by the STEP
and the process repeats.
Only the OOS periods count. Joined together they form the OOS equity curve: the closest thing to how the
strategy would have behaved if you had traded it live, re-optimising on a fixed schedule.
2. Key metrics
- Windows / traded windows / no-trade windows: how many OOS tests were run and how many produced trades.
- Total trades and trades per 1,000 bars: how much evidence there is. Below about 300 OOS trades every
statistic is fragile. - Net profit, profit factor, win rate, average OOS Sharpe: measured only on OOS trades.
- Max drawdown (global OOS): the deepest fall of the joined OOS equity curve.
- Read with caution: warnings shown when trades or windows are too few to trust the numbers.
3. Variant map: plateau or peak?
For each window StrategyProof looks at the parameter combination chosen on TRAIN and at its neighbours
(one grid step away on each parameter). If most neighbours are also profitable on TRAIN, the choice sits on a
plateau: small changes do not break it. If the neighbours fail, it is a lonely peak, and peaks
are usually the product of chance. Robust strategies show plateaus in most windows.
4. Variant cloud
The cloud compares walk-forward with every fixed combination of the grid, each traded unchanged over the
same OOS periods. The percentile tells you how many fixed combinations walk-forward beat. A high percentile means the
re-optimisation process adds value; a low one means a random fixed setting would often have done better.
5. Certification
A result is marked certified only when the evidence is consistent: enough trades, enough windows and
choices that sit on a plateau in most windows. “Not yet certified” is not a failure; it means the evidence is not strong
enough yet. Certification describes the past data tested and is never a promise of future profit.
6. Strategy families
- B01 Breakout: entries on the break of recent highs and lows.
- T01 Trend: trend following with a fast and a slow moving average.
- M01: RSI-based entries with fixed stop loss and reward-to-risk target.
- F01 Fair Value Gap: entries on price imbalances (FVG) left by strong moves.
7. Supporter tools
- Signal Filters: test whether a higher-timeframe trend filter, a volatility regime filter or an ONNX
model improves a completed research run. Every verdict is measured out of sample. - Monte Carlo: reshuffles the OOS trades thousands of times to estimate the range of drawdowns and the
probability of ruin. - Prop-firm challenge simulator: estimates the probability of passing challenge rules with your
results. - Robustness / stress test: higher costs, execution delay, removing the best trades, market regimes and
neighbouring parameters.
See also: Data · FAQ · Risk Disclaimer