ERM Regime Filter — Empirical Regime Model Methodology
The Empirical Regime Model (ERM) is AIBROKER's quantitative market regime filter. Definition, inputs, regime states, backtest evidence, and how it overlays daily momentum rankings.
The Empirical Regime Model (ERM) is AIBROKER's rules-based market regime classifier. It assigns each trading day to a discrete equity-market regime — broadly risk-on, neutral, or defensive — using observed market data: trend, breadth, cross-sectional dispersion, and realized volatility. ERM does not forecast market direction. It labels the environment so downstream momentum rankings and position-sizing rules can adjust aggression appropriately.
Why use a regime filter at all?
Momentum, value, and quality factors all have time-varying base rates of edge. A 12-month momentum ranking is positive on average over multi-decade samples but suffers severe drawdowns during regime transitions — most famously in early 2009 and in March 2020. A regime filter is the simplest way to acknowledge that one number per year ("average return") hides the fact that the factor's edge is concentrated in specific environments.
How ERM works (high level)
- Inputs: rolling trend strength of the broad index, realized cross-sectional dispersion, breadth indicators (advance/decline ratios, percent of names above their long moving average), and realized volatility regimes.
- Aggregation: each input is normalized to a robust z-score and combined into a single regime probability.
- Bucketing: the probability is mapped to discrete regime states — typically risk-on / neutral / defensive — using thresholds calibrated on the historical sample.
- Per-universe calibration: ERM is run independently per universe (S&P 500, NASDAQ 100, Nikkei 225, FTSE 100, EURO STOXX 50, ...) because regime dynamics differ meaningfully across regions.
Backtest evidence — momentum with and without ERM
Comparing pure 12M/6M/1M composite momentum to the same ranking with ERM-conditioned exposure shows the expected pattern: similar long-run CAGR, materially reduced max drawdown, and a higher Sharpe ratio. The biggest improvements come from cutting exposure during transition months (when momentum returns are most fragile), not from broad market timing.
How ERM is different from VIX-based filters
A VIX threshold is a single-variable proxy that tells you the option market expects volatility; it is highly correlated with already-observed drawdowns. ERM uses several non-options inputs (trend, breadth, dispersion) so it can flag emerging defensive regimes before realized volatility spikes — and ignore short, transient VIX pops in otherwise constructive trends.
Further reading
- Learn: Regime detection — how quants know when the market has changed
- Learn: The momentum premium — what the data actually shows
- Companion methodology: Composite momentum score
- Jegadeesh & Titman (1993) — original momentum factor research.
- Asness, Moskowitz & Pedersen (2013) — momentum and value everywhere.
Frequently asked questions
What is a market regime filter?
A market regime filter is a classifier that labels each trading day with a market environment — typically risk-on, neutral, or defensive — using observed market data. It is used as an overlay on quantitative strategies to adjust aggression based on the type of environment the strategy is operating in.
How is ERM different from a VIX threshold?
VIX thresholds use a single options-derived input and tend to flag regimes only after realized volatility has already moved. ERM uses several non-options inputs — trend, breadth, dispersion, realized volatility — so it can flag emerging defensive regimes earlier and ignore short transient VIX pops.
Does ERM forecast market direction?
No. ERM does not predict whether the next day, week, or month will be up or down. It labels the regime so downstream strategies can adjust aggression. The forecast question — 'what will the return be?' — is left to the underlying momentum ranking.
Is ERM the same as a trend-following filter?
ERM includes trend as one of its inputs, but it is broader than a single moving-average rule. The classifier also looks at breadth, cross-sectional dispersion, and realized volatility, which often diverge from the index trend during regime transitions.