Backtests and their limits

Methodology

A backtest runs a rule over history as if it had been live, and is the only honest way to check a rule before it is. Every model here was backtested before activation, and the historical statistics quoted on the platform come from those tests.

What the tests covered

• Daily data from 2015 for the index models; ten years of daily data for the 2026 round that produced the two rare Elite models.
• Entries at the close of the day a rule fired, never earlier in the day.
• Transaction costs deducted on every entry and exit.
• A held-out period: the most recent years were kept out of any tuning and reported separately as out-of-sample.
• No parameter fitting to the test period. A candidate had to clear a fixed bar on risk-adjusted return, maximum drawdown, trade count and out-of-sample performance before activation, and most candidates did not.

Findings that carry

• Volatility-regime filters hurt trend-following rules and help currency mean reversion. That is why the regime check sits inside a checklist rather than switching a trend model off.
• Seasonality on the JSE is real enough to be a check. In the backtest (JSE Top 40 proxy, 2015 to 2026) April was the strongest month, averaging +3.32% with 83% of years positive, and September the weakest across all eleven assets tested, averaging about -1.4%. That is the Not September chip on a model page.
• Costs matter more than filters for fast mean reversion. Rules that trade often lose their edge to the spread, which is why the live models trade rarely.

What a backtest cannot tell you

• It assumes a fill at the closing price. Real fills differ, and on the JSE the difference is larger.
• Costs are estimates. Slippage in a stressed market is not.
• Regimes change. A rule that worked through 2015 to 2026 has seen one rate cycle and one pandemic.
• A rare model produces few trades, so its statistics carry wide uncertainty. A rule that fired a handful of times in ten years cannot be judged by its win rate.
• The screener universe is today's constituents. A company that fell out of the top 150 is not in the history, which flatters any backward look.
• The live record is short. The statistics on Models and History combine positions from the rule applied to past data with the live record since the models went live in 2026, and the count beside a percentage is the number to read first.