- regime switches after the sticky prior and dwell rule
- 243 → 14
- defensive days still caught
- 106 / 106
- continuous risk scalars driving exposure
- 3
The question
Regime models are easy to fit and hard to trade. A classifier that flips state every few days produces a portfolio that churns, and every flip costs spread and impact. We wanted a regime signal stable enough to act on, a risk layer that scales exposure smoothly instead of switching it on and off, and an execution layer honest enough to show what the targets would cost.
Approach
- Regime layer. Compared three classifier families: a macro/half-life model, a feature-rule model, and a Gaussian HMM. A sticky transition prior plus an asymmetric three-day dwell rule cut regime switches from 243 to 14 while still catching all 106 defensive days.
- Risk layer. Three continuous scalars set exposure: portfolio covariance, strategy health, and execution conditions. Separating them lets the platform tell a strategic de-risking from a temporary one, so exposure moves smoothly instead of in steps.
- Execution layer. A one-minute parent/child order simulator with ADV caps, spread impact, partial fills, and synchronized legs for pair trades.
- Paper trading. Risk-adjusted targets are sent to Alpaca paper trading, and broker orders, fills, and positions are reconciled against the simulated execution.
This was an industry-sponsored capstone. The code lives in a private repository managed by the program, so this page describes the design and headline results only.