The regime-switching EA lost to plain long trend following
Adding complex logic to switch between different trading strategies based on market conditions does not improve performance.
Adding complex logic to switch between different trading strategies based on market conditions does not improve performance. My latest research into a regime-switching system confirms that trying to be flexible often just dilutes your results. I designed this system to act as a router, assigning specific sub-logic (trend-long, trend-short, or mean-reversion) to different market regimes. To ensure the test was fair, I used fixed rules for routing that were determined in advance, keeping the evaluation entirely out-of-sample across eight currency pairs on the one-hour timeframe from 2017 to 2026. Here is how the different sub-strategies performed across all currency pairs:
| Strategy | Performance |
|---|---|
| Trend Long | 6 of 8 pairs profitable (Avg PF 1.12) |
| Mean Reversion | 1 of 8 pairs profitable (Avg PF 0.91) |
| Trend Short | 3 of 8 pairs profitable (Avg PF 0.87) |
| The data shows that a consistent edge only exists in the trend-long logic. When I compared a simple, permanent trend-long strategy against more complex switching versions, the results were clear: |
- Baseline (Trend Long only): PF 1.12, Sharpe ratio +0.72, profitable in 8 to 10 years, and a total return of +140%.
- Switching Variants: Adding logic to switch to short or mean-reversion strategies caused performance to deteriorate, with Sharpe ratios falling between +0.18 and -0.39. A full-switching model actually resulted in a -61% return. I also tested various filters designed to avoid high volatility, sidestep ranging markets, or only trade during upward trends. These filters did lower the drawdown, but they reduced returns proportionally. In other words, they failed to improve the risk-adjusted return. The reason for this failure is that the non-trend strategies have a negative expected value. Switching to them only serves to dilute the gains from the trend-long core. Furthermore, adding extra filters often destroys the edge entirely. The trend-long strategy already incorporates its own SMA (Simple Moving Average) filter, and adding external gates meant missing out on genuine opportunities. While this project did not yield a better trading system, it produced a valuable framework. I have built a robust “Regime Router” that allows for forward-testing and strict, objective evaluation without the risk of hindsight bias. If I discover a new, genuinely profitable edge in the future, I can plug it into this framework to verify it immediately. For now, however, the simplest approach remains the best. The core trend-following strategy continues to be my most reliable system.
How this connects
This verification builds on earlier ones (what failed before and what I tried this time, comparisons between approaches).