#52 — Markov D1 regime filter

Hypothesis — what it tests

Whether a causal daily Markov regime-transition signal carries next-day directional information, and whether gating counter-trend retrace entries by regime alignment improves the edge.

Description

Inspired by a 'How Quants Actually Trade' transcript, this built an empirical 3-state (Bull/Bear/Sideways) daily Markov transition matrix per pair and used the signal P(Bull|state) − P(Bear|state) to gate the post-shock retrace strategy. The Phase-1 IC study found that ALL information coefficients are negative — a Bull regime predicts a negative next-day return — reinforcing the project's recurring counter-trend finding (consistent with the CSI factor study). Phase 3 applied the filter to retrace entries (an upshock SHORT requires Bull regime, a downshock LONG requires Bear), causally using the day D−1 signal, and the best config (mw=10, mt=0.002, sig_thr=0.20) lifted the portfolio from +61.9 to +70.4 p/d and walk-forward from 7/12 to 9/12 at mc_p=0.0000 while trading only 73% as often. It was baked directly into the live retrace service on account 009, with a paper no-filter variant (retrace_nofilter) run as a live A/B; over an early 3.5-day window the filtered live leg roughly broke even while the unfiltered paper leg lost, directionally consistent with the backtested uplift.

Key result

Indicators

Algorithms

Code