← Research logMachine Learning
PARTIAL · risk gate
#

BOCPD (Adams–MacKay) flags structural breaks in the return-generating process in near-real-time; cutting risk at a detected changepoint avoids regime-transition losses.

Recursively update the posterior over the current run length $r_t$:

$$ P(r_t\mid x_{1:t}) \propto \sum_{r_{t-1}} P(r_t\mid r_{t-1})\,P(x_t\mid r_{t-1})\,P(r_{t-1}\mid x_{1:t-1}) $$

BOCPD with a Gaussian observation model and hazard $1/\lambda$ on volatility-standardized returns; de-risk when changepoint probability spikes.

Detection lag2–6 bars
Drawdown reduction (as gate)−22%
Standalone signalnone
PARTIAL EDGE
Useful as a de-risking gate — it catches volatility-regime breaks a few bars in and trims drawdowns ~20% — but it generates no directional edge and lags by construction. Kept as a portfolio risk overlay.
Changepoint detectors are insurance, not alpha. Their value is cutting exposure into a regime break, accepting that "near-real-time" still means a few bars late.

We publish the failures too.

This is one of 100+ documented hypotheses. Browse the full lab notebook, or see the strategies that survived into production.