Regime-routed experts don't beat always-on

Abstract

A mixture-of-experts (MoE) router conditions which sub-models trade on an observed market state. We test the strongest version available to us: route the validated Phase 1.7 per-coin directional specialists through a freshly fit, strictly-causal three-state market-regime detector. Under an honest, leakage-controlled evaluation, the routed book underperforms an always-on book of the same experts by −4.36 Sharpe on the well-powered discovery window and is mildly negative on the walk-forward window. The regime-conditional skill that exists in-sample does not persist out-of-sample, and routing additionally throws away the diversification that makes the 175-expert always-on book strong. This replicates two earlier regime-detector findings on a much stronger expert set and with three balanced regimes — so the failure is not a regime-coverage artefact. We did not build the live router.

1 · The idea, and why it's tempting

By the end of the Phase 1.7 funnel we hold a roster of per-coin specialists: (architecture, coin) pairs whose directional signal is sign-consistent across three out-of-sample regimes and clears its noise floor. The natural hypothesis is that these specialists are conditionally good — a coin's short specialist should fire in a risk-off regime, a long specialist in a risk-on grind. If true, a regime router would lift the book by activating the right experts at the right time and silencing the rest. This is the classic mixture-of-experts gating argument.

The hypothesis is testable, and the honest test is unkind to it.

2 · The detector we built

A Gaussian mixture model over ten causal, market-wide hourly features — BTC realised volatility at 24h and 168h, cross-sectional mean realised vol and return dispersion, average coin-to-market correlation, BTC trend at 24h and 168h, funding mean and dispersion, and open-interest velocity. Every feature is backward-looking. The standardiser and the GMM are fit on the training window only (≤ 2025-12-31, 11,361 bars), then frozen and applied by argmax posterior to all later bars — strictly causal across both out-of-sample windows. The number of states is chosen by a knee rule over BIC (raw min-BIC degenerately fragments into thin near-duplicate mid-vol states); the knee selects k = 3.

3 · The decisive numbers

The book is an equal-weight portfolio of the per-coin specialist directional trades: honest rolling-median centring, sign-corrected, non-overlapping 24-hour holds, 4 bps per leg, phase-averaged over 24 entry offsets. We fit the static regime → specialist mapping on a train half and evaluate it on a disjoint validate half. Lift = routed book Sharpe − always-on book Sharpe on the same bars.

4 · The "adaptive" variant is a trap, not a green light

A more flexible router — a causal expanding-window gate that activates an expert once its cumulative signed return in the current regime has been positive — looks spectacular: +6.20 Sharpe on discovery, +5.86 on walk-forward. It is not deployable evidence, for two reasons.

• Every plausibility flag fires. Both adaptive books trip EXTREME_SR (Sharpe > 5) and UNUSUALLY_CLEAN (per-bar mean/std > 0.3 — e.g. +47 bps/bar on discovery against +13.6 for always-on). A clean, high mean is the signature of a selection / survivorship artefact: a gate that keeps recently-winning experts manufactures a smooth equity curve regardless of whether any regime structure exists.
• The regime label barely contributes. A regime-agnostic control — the same "keep what's winning" gate, ignoring the regime entirely — already captures +4.97 / +4.02 Sharpe. The incremental regime contribution is only +1.23 / +1.84 Sharpe, sitting on top of a flagged, selection-biased base. Most of the apparent adaptive edge is generic momentum plus diversification, not regime alpha.

5 · This is the third time

Two earlier internal experiments reached the same verdict with weaker tooling: a rule-baseline regime MoE and a hidden-Markov regime detector both failed to beat an always-on baseline. The standard objection to those was regime coverage — one of them evaluated on a window that was 89 percent a single regime. This run removes that objection: three balanced, well-populated regimes over both out-of-sample windows, routing the much stronger Phase 1.7 IC specialists. The conclusion holds. Market-wide regime routing of a per-coin specialist book does not generalise.

6 · What would change the conclusion

A negative result is a verdict on a specific construction, not on the whole idea. We would revisit regime conditioning if:

• the detector were tied to a per-coin or per-cluster state rather than a single market-wide label — specialists may switch on a coin-specific regime the market index can't see;
• we had a longer out-of-sample with more independent regime cycles (the current windows are ~2.5 months each, and thin per-regime samples already flag low-sample warnings);
• the object being routed were the cross-sectional ensemble — which is genuinely deployable — rather than the per-coin specialist book.

None of these were pursued here. The gate's only job was to decide whether to build the live router, and the answer is no. The capital stays in the always-on book and the cross-sectional ensemble, both of which we can defend.