Project 01 · SPY Latent-State Allocation Framework

A walk-forward SPY allocation study built around a latent-state core model and modular risk overlays. The core estimates broad recurring market environments from low-dimensional market behavior, while volatility and trend overlays adjust sizing in a transparent and interpretable way.
Walk-forward
Weekly rebalance
Core + overlays

Framework Notes

Core model: a latent-state model estimates which broad market environment is most likely active using only SPY’s own historical behavior.

Exposure logic: the core sets baseline exposure by regime, while volatility and trend are applied afterward as modular overlays rather than being forced into the latent-state classification itself.

Research goal: test whether regime-aware exposure combined with lightweight risk controls can improve out-of-sample risk-adjusted performance relative to static and one-rule alternatives.

Core

Estimate persistent market environments using a latent-state model on low-dimensional features, then map those environments to a baseline exposure ladder.

Overlays

Apply volatility and trend as post-classification sizing controls. This keeps state estimation separate from risk control and makes the framework easier to interpret.

Evaluation

Focus on out-of-sample risk-adjusted behavior, drawdown control, and exposure efficiency rather than relying on raw return alone.

Core and Overlay Variants

The regime shading is drawn from the final combined overlay variant (Core + Vol + Trend Overlay).

Return vs Drawdown Tradeoff

This chart compares return against max drawdown magnitude. The drawdown number is displayed as a positive magnitude for readability, so lower is better.

Balanced Performance Radar

This radar provides a compact multi-metric comparison across OOS Sharpe, OOS Sortino, OOS CAGR, OOS max drawdown magnitude, and average exposure.
Conclusion

The main comparison here is not just raw return, but whether any regime-aware variant improved the return / risk tradeoff out of sample. That is why the two retained visuals focus on CAGR versus drawdown and a more holistic balanced radar view.

In this run, the strongest model by OOS Sharpe is Core + Vol + Trend Overlay at 0.7795. The strongest by OOS CAGR is Latent-State Core at 10.44%. The shallowest OOS drawdown belongs to Core + Vol + Trend Overlay at 14.89% max drawdown magnitude.