Portfolio-level vol targeting was not robust enough
Revisiting the aggregate volatility-targeting approach for the v1.4.0 update has confirmed that the decentralized version I currently use is the more…
Revisiting the aggregate volatility-targeting approach for the v1.4.0 update has confirmed that the decentralized version I currently use is the more robust choice. In my previous research, I tested aggregate volatility-targeting, where all strategy sleeves are controlled by the volatility of the total account equity. At the time, it showed a 14% performance boost over the decentralized version. However, I initially rejected it because it required a master controller, which created a single point of failure and unnecessary complexity. Since my deployment strategy has shifted to a single MT5 EA that already tracks total account equity, the cost of implementing this aggregate approach effectively dropped to zero. I decided to re-evaluate it using the v1.4.0 architecture, which includes stock filters and a refined volatility cap.
| Metric | Decentralized (Current v1.4.0) | Aggregate (L2) |
|---|---|---|
| Monthly Return | 1.59% | 1.68% |
| Drawdown | 10% | 10% |
| M1 Worst Drawdown | 1.43% | 1.43% |
| MC Pass Rate | 94% | 94% |
| The aggregate method showed a 6% improvement in monthly returns across the full data set. However, this is significantly lower than the 14% gain I observed in my earlier research. It appears that the stock filters in v1.4.0 are already effectively capturing the risk-off timing across different sleeves, leaving little room for the aggregate method to provide any real additional benefit. | ||
| Beyond the marginal gains, the aggregate approach failed my validation process in two specific areas: | ||
| First, the system failed to converge. When I iterated the volatility calculations, the monthly returns fluctuated between 1.61%, 1.68%, and 1.50%. Because the leverage is tied to account equity and the equity is determined by that leverage, the system creates a feedback loop that makes the results unstable. In contrast, the decentralized version manages volatility within each individual stream, which is structurally stable and avoids this oscillation entirely. | ||
| Second, the performance was inconsistent across different market regimes. While the aggregate method performed better during the weaker market period of 2015 to 2020 (+0.71% vs +0.61%), the decentralized version outperformed it during the stronger 2021 to 2026 period (+2.55% vs +2.45%). The 6% gain seen in the full-period aggregate test was essentially a false positive driven by a specific, older market environment that does not reflect current conditions. | ||
| Ultimately, this exercise confirmed that my original decision to reject the aggregate approach was correct not just for reasons of complexity, but for its lack of structural robustness. The decentralized volatility-targeting used in v1.4.0 remains the superior design. I will not be making any changes to the current system. |
How this connects
This verification builds on earlier ones (what failed before and what I tried this time, comparisons between approaches).