Regime Detection: A 50-Point Win Rate Spread and the System That Learned to Exploit It

Market regime is the single most predictive feature in the Apex signal pipeline. This was not a design assumption. It was discovered empirically after auditing 13,266 usable live trades across a 12-day window. Segmenting by regime classification revealed a 50-percentage-point win rate spread: choppy regime produced 62.1% win rate; trending regime produced 12.3% win rate.

The problem was that the regime classifier in production was not classifying regime at all. The function infer_regime() classified by symbol name, not by market structure. A symbol whose ticker happened to match an internal string pattern would be routed to the choppy bucket; everything else fell to trending. The 62.1% win rate that made regime appear valuable was partly an artifact of which symbols the name-matching logic routed to which bucket.

All five mean-reversion discovery signals produced 0% win rate in the then-current trending crypto environment. These strategies had been tuned for equity sideways markets, where mean reversion is a structurally valid edge. In a trending crypto regime they were anti-patterns, systematically entering against momentum. The system had no mechanism to suppress them.

The corrected regime classifier consumed five data inputs: ADX computed over a 14-bar lookback, price-to-MA200 ratio, 5-day momentum, realized volatility over a 20-day window, and funding rate for crypto symbols sourced from Binance 8-hour settlement across 719 symbols. A rule-based ensemble combined these inputs into one of five regime states: trending_up, trending_down, choppy, range_bound, or high_vol.

The Argus signal schema included both regime and regime_confidence fields in every emitted signal, making regime context a first-class component of the decision payload rather than a lookup that happened at execution time. Every signal was tagged at generation time with the regime in which Argus detected the pattern, and Apex's filtering layer enforced source-regime compatibility rules.

The crypto regime gate was implemented in safety/regime_gate.py, which blocked all disc_ strategies in trending regimes. The gate wrote 19 distinct anti-patterns to the brain's persistent memory, making the suppression durable across restarts.

The train/test methodology for regime validation used a 24-month lookback across the full symbol universe with walk-forward expanding windows. Each fold used 18 months for regime-label generation based on ex-post price trajectory and 6 months for out-of-sample evaluation. The regime label generation used a simple rule: if price moved more than 5% directionally over the next 20 bars, the regime at entry was labeled trending; if price stayed within a 3% band, it was labeled choppy or range_bound based on ADX.

The LTC-BTC contagion signal, confirmed at a Pearson correlation of r=0.52 on 60 days of minute-bar data, was wired as a regime-dependent trigger. When BTC moved more than 1% in either direction, LTC followed within 1 to 2 bars. SOL and LINK were decorrelated at r=-0.10 and r=-0.003 respectively, confirming that contagion was pair-specific rather than a general altcoin property.

After correcting the infer_regime() function to use actual market structure features rather than symbol name matching, regime classification accuracy on the held-out validation window reached 74% versus the ex-post trajectory label. The 50-point win rate spread between choppy and trending regimes remained, confirming that the spread was real and not purely an artifact of the broken classifier.

Supressing all disc_ strategies in trending regimes eliminated an entire category of losing trades. The 19 anti-patterns written to the brain's persistent memory represented the full set of mean-reversion entries that had fired during the April trending crypto environment. Regime key TTL enforcement reduced Redis memory usage for regime data by approximately 40% compared to the pre-fix accumulation state.