The Problem With One Layer
After the Inverted Shield discovery, I had a working strategy. Sharpe 2.31. Good enough to trade. Not good enough to be proud of.
The issue wasn't the signal. The Inverted Shield, properly filtered for the both-types sub-regime, was producing real edge. The issue was that the strategy was deploying that signal regardless of the macro environment. In calm, low-stress markets, it performed well. During macro stress periods — sustained drawdowns, extended trend regimes, structural market dislocations — it bled.
One regime classifier isn't enough. MES doesn't live in a single behavioral state. It has layers: the micro-level pattern (what's happening right now in price structure) and the macro-level context (where are we in the broader market stress cycle). Treating them the same was costing me roughly two full Sharpe points.
The fix was a second layer of intelligence sitting above the first.
Drawdown-From-Peak: The Macro Stress Signal
The meta-controller is built around one primary variable: drawdown from peak in the strategy's equity curve.
Not the market's drawdown. The strategy's own drawdown from its recent equity peak. This is the macro stress signal because it captures something simple and important: when the strategy is in drawdown, the market conditions that produced the edge have temporarily shifted. The regime classifier at the micro level might still be firing signals, but the macro context says those signals are less reliable right now.
The meta-controller uses the drawdown-from-peak reading to classify the current environment into macro stress states. Each stress state maps to a different portfolio of active strategies. In low-stress states, the system is aggressive — more strategies active, more capital deployed. In high-stress states, the system contracts — fewer strategies, smaller exposure, tighter filters.
This is not a stop-loss. It's a regime-adaptive portfolio selector.
A stop-loss pauses all trading at a fixed dollar threshold. The meta-controller dynamically adjusts which strategies are running based on a continuous stress reading. It doesn't turn the system off. It changes the system's posture.
The Four-Strategy Architecture
PCS ultimately resolved into four distinct strategies, each optimized for a specific combination of macro and micro conditions:
The meta-controller routes capital across these four strategies in real time. The regime multiplexer (the micro layer) handles the signal detection within each strategy. Together, they form a two-layer adaptive system where each layer is doing a job the other layer can't do.
Monday Dominance
Breaking down momentum trades by day of week, Monday wasn't slightly better — it was dominant. Monday profit factor: 1.919. Next best day: 1.115. That spread doesn't happen by accident. It likely reflects weekend gap dynamics and institutional positioning resets. PCS treats Monday as its own regime.
The OR Logic Principle
This is the discovery that nearly killed a good system through well-intentioned engineering.
When multiple regime detectors were producing signals, my first instinct was to combine them: weight them, average them, build a composite score. That was wrong.
Weighted-summing uncorrelated detectors dilutes each one's edge. Two detectors catching different market phenomena produce independent catches — they see different things. A composite score turns two sharp signals into one mushy one that catches neither phenomenon reliably.
The correct combination is OR logic:
If Detector A fires OR Detector B fires, respond. Don't average. Don't weight. Let each detector do its specific job.
The practical impact: switching from weighted combination to OR logic was responsible for a meaningful portion of the Sharpe improvement. Each strategy in PCS has its own dedicated regime classifier. The meta-controller switches between them — it doesn't average across them.
Sharpe 2.31 → 4.29
The performance improvement from adding the meta-controller layer was not incremental. It was transformative.
| Configuration | Sharpe | Calmar | Max DD |
|---|---|---|---|
| Inverted Shield alone | ~1.4 | — | — |
| + Sub-regime filter | ~1.9 | — | — |
| + OR Logic architecture | 2.31 | — | — |
| + Meta-controller | 4.29 | 4.13 | $459 |
Each layer added real edge. The meta-controller roughly doubled the Sharpe ratio on its own. That's what happens when you stop fighting macro stress and start routing around it.
Walk-Forward Validation
Every configuration in that table was validated using rolling walk-forward out-of-sample testing across 5.5 years of MES data. No parameter is allowed to "know" about the test window it's being evaluated on. Each window trains on prior data only. All years were profitable.
The Sharpe and Calmar numbers above are OOS, not backtest artifacts. The final configuration has never been fitted to the full dataset. That's the only kind of confidence worth having.