Three more research axes from autonomous exploration
Adding extra filters to an existing trading system often feels like trying to squeeze more juice from a lemon, but my recent testing shows that most…
Adding extra filters to an existing trading system often feels like trying to squeeze more juice from a lemon, but my recent testing shows that most of these attempts simply lead to dead ends. After evaluating seven different research paths to improve my autonomous trading logic, I have found that almost every new idea I tested was either redundant or actively harmful to performance. Here is the breakdown of the failed and redundant approaches:
- Pair-specific stock filtering: I hypothesized that filtering each currency pair individually using stock market data would improve results. This failed because the existing logic already accounts for indices and gold, making the extra layer redundant.
- Connors-style mean reversion: I tried using mean reversion to fill the “empty” time when the core system was idle, aiming to capture risk-off periods. This proved useless; the mean reversion strategy only made money on risk-on days, and the risk-off performance was a negligible 1.2% over 11 years.
- Heikin-Ashi smoothing: By attempting to denoise entries using Heikin-Ashi, I only introduced lag. This resulted in worse performance for two out of every three pairs compared to a standard breakout.
- Asymmetric stock overlay: I tested increasing leverage during strong risk-on market conditions. While this increased total returns, the drawdown grew even faster, dropping the Calmar ratio from 1.08 to 1.01. In other words, the risk taken was not worth the reward, and the existing symmetric v1.4.0 version remains the superior choice. I did test one approach that showed promise: using the ratio of currency pairs trending in the same direction to adjust leverage. Although this worked for the core FX logic and yielded a 0.33% gain compared to 0.26% from stock signals, it was essentially redundant. It functioned as a risk-off detector, just like the stock market filters already integrated into v1.4.0. Looking back at this research track, it is clear that out of the seven axes I tested, only the intermarket stock signal (v1.4.0) provided a genuine improvement: specifically a 7.6% boost. The rest were either ineffective or covered by mechanics I had already implemented. Some readers previously suggested that my system had hit a “ceiling.” While I disagreed at the time, these results prove that finding a new, non-redundant edge is difficult. I have effectively exhausted the “risk-off reduction” approach. Moving forward, the only significant untapped areas involve complex sizing models (such as machine learning or reinforcement learning), Renko bars, or incorporating data points outside of price action that fall outside of typical prop-firm constraints.
How this connects
This verification builds on earlier ones (what failed before and what I tried this time, comparisons between approaches).