TESTING · 7 MIN READ
Walk-forward analysis explained
Walk-forward analysis fits a strategy on one window of history, tests it unchanged on the next window, then rolls both forward and repeats. A single backtest assumes you knew the right parameters in advance. Walk-forward removes that assumption, which is why its results look so much worse.
How walk-forward analysis works
- Fit and tune the strategy on a first window, say months one to six.
- Test it, unchanged, on the window straight after it: month seven.
- Roll both windows forward by one period and repeat.
- Join every test period into one equity curve.
Every point on that curve comes from a period the strategy had not seen when its parameters were set, so the whole curve is out-of-sample. That is far stronger evidence than one held-out period, and much stronger than an optimised backtest.
What it shows that a single split cannot
It tells you whether the strategy needs refitting to keep working, and how fast its parameters go stale. A system that only does well when refitted every month is chasing a moving target. Treat that as a warning, however good the joined-up curve looks.
It also shows parameter instability. If the best setting jumps around from one window to the next, the strategy is fitting noise even when the total is positive. That is the most common way overfitting gets past a naive test.
Walk-forward needs more trades than most bots take
Each window uses up history, and a strategy that trades rarely gives you very few trades per window. A system averaging ten trades a month produces about ten results per monthly test window, far too few for any one window to mean anything.
This limits our own bot too. Its four strategies take about five trades a month, about 59 with the three gold clock rules added, and pretending otherwise would misrepresent what the method can do here. That is why our published figures are 22 separate monthly replays with the trade count beside each one, instead of a walk-forward. See sample size.
How we run a replay
Our replay engine feeds real broker bars through the same signal code the live engine runs, with the real performance tracker choosing the strategy. There is no separate backtest version of the code, so a whole class of hidden differences between test and live cannot creep in.
Persistence is hard-disabled during a replay, so a test cannot change live state, and the full gate chain (session, spread, bias, concurrency) runs exactly as it does live. See backtest does not match live results for the gaps that remain.
An out-of-sample test is also where our own bot failed. We wrote the test down in advance and then ran the four strategies in the bot on their own, on gold from 2019 to 2024, years they were not tuned on. None of the six years was positive. We publish that, and we never call the bot's edge proven.
Keep reading
- How to backtest a strategy without fooling yourselfA backtest is the weakest evidence worth having. Here is what makes one trustworthy, and the…
- Curve fitting, and how to recognise itOverfitting is the largest single cause of trading bot failure. Here is how it happens throu…
- Reading a verified track record properlyA verified badge proves the trades happened. It does not prove the account is live, the only…
- The 34 strategies we tested
Further reading
Educational information only, not financial advice. Trading leveraged products carries substantial risk of loss. Last updated 2026-11-11.