Systematic Regime Trading

September 2008. The S&P 500 lost 30% in three weeks.

Most systematic strategies were fully invested when the crash hit. This system detected the regime shift 6 days earlier. By September 19, allocation had dropped from 85% to 12%.

-29%

Market (Sep-Dec 2008)

-4.3%

Strategy (Sep-Dec 2008)

-55%

Buy & Hold Max DD

-28%

Strategy Max DD

Performance

Hover for daily returns. Click legend to toggle strategies. Drag to zoom into any period.

Drawdown from peak. Buy-and-hold hit -55% in 2008. The regime strategy cut that in half while delivering higher returns (10.2% vs 9.6% CAGR).

Strategy	CAGR	Vol	Sharpe	Max DD	Turnover
Regime Momentum	10.2%	9.3%	0.70	-27.7%	5.7x
Binary Regime	7.5%	12.5%	0.49	-35.8%	10.6x
Vol-Managed	6.6%	10.5%	0.47	-23.7%	1.7x
SMA 200	7.0%	11.7%	0.47	-21.9%	6.0x
Buy & Hold (SPY)	9.6%	20.1%	0.46	-55.2%	0.0x

Signal vs Market Drawdown

Blue: equity allocation. Red: market drawdown. The strategy reduces exposure before and during every major drawdown.

Architecture

190 Stocks → 6 Features → 5 Models → Ensemble → Signal → Backtest

HMM (5-state Gaussian, remapped to 3): transition persistence
GARCH(1,1) Student-t: conditional volatility clustering
K-Means: feature-space clustering
Gaussian Mixture Model: Bayesian posterior probabilities
Markov-Switching (Hamilton 1989): regime-dependent variance

Walk-forward validation, expanding window, quarterly refit. 59 windows, 3,771 out-of-sample predictions. SPY benchmark. 1-day execution lag. Asymmetric confirmation filters.

Walk-forward loop

while train_end_idx + step_days <= n_total:
    train = feature_df.iloc[:train_end_idx]
    test = feature_df.iloc[train_end_idx:test_end_idx]

    probs = {
        "hmm": fit_predict_hmm(train, test),
        "garch": fit_predict_garch(train, test),
        "kmeans": fit_predict_kmeans(train, test),
        "gmm": fit_predict_gmm(train, test),
        "ms": fit_predict_markov_switching(train, test),
    }
    combined = ensemble.combine(probs)
    train_end_idx += 63  # quarterly

Regime Prediction

Realized volatility over the next 5, 10, and 20 days by predicted regime. Calm: 12%. Turbulent: 22%. Clear separation across all horizons.

SPY with regime shading. Bottom: ensemble probability decomposition. Turbulent regimes concentrate around 2008, 2011, 2015, and early 2020.

Performance by Regime

Calm markets: Sharpe 1.10, 81% of total return from 52% of trading days.

Try It: Cost Simulator

Drag the slider to see how transaction costs affect strategy performance. Institutional equity costs are typically 5-10 bps.

Transaction Cost: 7 bps

0.70

Sharpe

10.2%

CAGR

-28%

Max DD

31 bps

Breakeven

Through the 2008 Crisis

$10,000 invested in September 2007, right before the worst financial crisis in modern history. Watch what happens.

$10,000

Strategy

$10,000

SPY Buy & Hold

Crude Oil Extension

WTI crude oil regime detection. Correctly identifies the 2008 crash, 2014-2016 oil glut, and the 2020 negative price event. Cross-asset correlation with equity VIX: 0.25.

Ensemble Diagnostics

Pairwise agreement between five models. Feature-space models (HMM, GMM) and return-based models (GARCH, Markov-Switching) form two independent clusters.

Cumulative excess return vs SPY. Alpha concentrates during crisis periods. Fama-French R-squared of 0.20 confirms low factor exposure.