The short answer
Mostly because the backtest found noise. In our own research, a rule we designed by searching 208,000 events made +0.10R to +0.12R a trade in the year it was found on and lost -0.06R to -0.12R the next year; the best of nine opening-range settings went from +0.13R to -0.02R; and two look-ahead bugs made levels look powerful until they were fixed. Every result on these pages is judged on data the rule never saw.
What we tested
- Markets
- Gold, major FX pairs and US indices
- Period
- Year 1 to find, year 2 to judge
- Round 9: 208,049 break-retest and rejection events on M5 and M15, a search for up to three conditions positive in both halves of year 1, frozen in writing, then judged once on year 2.
Results
| Rule | Year 1 (found on) | Year 2 (judged on) |
|---|---|---|
| Break and retest in quiet hours | +0.117R (n 396) | -0.057R (n 391) |
| Rejection in quiet hours, against the move | +0.102R (n 429) | -0.123R (n 448) |
| London opening range, best of 9 settings | +0.13R | -0.02R |
What it means
- Look-ahead bug 1: weekly bars labelled at Friday 00:00 contained the rest of Friday, so prior-week highs and lows looked like magnets. The effect vanished once fixed.
- Look-ahead bug 2: a popular Python 'smart money concepts' library marks swings using bars that come after them, so any backtest run over a full history sees the future. We replaced it with a detector that only uses closed bars.
- Data hygiene: short month-end data windows fetched for one test leaked into later ones until the loader kept only continuous data.
- Our rules now: write the rule down before testing, judge on a period it never saw, compare with random entries or fake levels, count every setting tried, and report failures.
Limits of this test
- These are our examples; they illustrate common failure modes rather than measure how often they occur.
Frequently asked questions
What is overfitting in trading?
Choosing rules or settings that fit past data by chance. In our research a searched rule made +0.12R a trade in the year it was found on and lost 0.06R the next year.
What is look-ahead bias in a backtest?
When the test uses information that was not available at the time of the trade, such as a weekly bar that already contains the rest of the week, or swing points marked using later bars.
How every test on these pages works
- Prices: the broker's own bars from an FTMO cTrader account. Fills are simulated minute by minute on one-minute bars; when one minute touches both the stop and the target, the stop is assumed to fill first.
- Costs: the broker's bid/ask spread measured by hour of day, commission per market, and the broker's own overnight swap rates for anything held past the daily rollover.
- Periods: year 1 = 24 Oct 2024 to 23 Oct 2025, year 2 = 24 Oct 2025 to 24 Oct 2026. Daily rules also use Dec 2020 to Oct 2024 ("2021-24"). Rules are written down before testing and are not tuned on the period that judges them.
- Controls: every rule is compared with entries that have no reason behind them (random sides, random days, or the same rule at fake levels), so a result has to beat chance, not just be positive.
- R is the trade's risk (entry to stop). An average of +0.10R means the average trade made a tenth of what it risked; -0.10R means it lost a tenth.
The full results for the rules RezSync Algo trades, by market and risk profile with every trade downloadable, are in the backtest report.
Published research
Related research
Research, not financial advice. These are simulated results on historical prices: they involve no real money, are produced with hindsight of the rules, and past results do not predict future ones. Costs and fills are modelled, and your broker's prices and costs will differ. Trading leveraged products carries a high risk of losing money; test on a demo account first.