Regime Detection

Regime detection addresses one of the fundamental challenges in algorithmic trading: financial markets are non-stationary. A strategy that works in trending regimes may lose money in chop; a strategy that works in low-volatility regimes may blow up in high-volatility regimes. Regime detection aims to identify the current state and adapt accordingly.

Common regime classifications:

**Trend vs Range:** - Trending markets: directional persistence over multiple bars, useful for trend-following strategies - Ranging markets: bouncing between support/resistance, useful for mean-reversion strategies - Common detection: ADX above 25-30 = trending; ADX below 20 = ranging - Alternative: linear regression slope significance; price relative to long moving average

**Volatility Regimes:** - Low-volatility: tight ranges, small candles, mean-reversion dominant - Normal: typical operating environment - High-volatility: wide ranges, large candles, news-driven moves, breakouts dominant - Common detection: realised volatility relative to historical distribution; VIX-equivalents for forex (e.g. FX Volatility Index); GARCH-modelled volatility regimes

**Risk-on / Risk-off:** - Risk-on: positive risk sentiment, commodity currencies and equities rally, safe-havens (USD, JPY, CHF) weaken - Risk-off: negative risk sentiment, opposite movements - Common detection: correlation patterns across asset classes; equity index movements (S&P 500 as risk-sentiment proxy); commodity-currency vs USD relative strength

**Microstructure regimes:** - High-frequency: 'normal' liquidity, narrow spreads, predictable execution - Stressed: wide spreads, fragmented liquidity, dealer caution - Common detection: realised spread; quoted-vs-traded volume ratios; order book depth (if accessible)

Detection methods:

**Indicator-based (simplest):** - Threshold rules on existing indicators (ADX, ATR, Bollinger Band width) - Easy to implement; sometimes too binary - Example: 'trending' if ADX(14) > 25 for 5 consecutive bars; 'ranging' otherwise

**Statistical models:** - Hidden Markov Models (HMMs): assume the market is in one of K hidden states, each with its own emission distribution; estimate state probabilities from observed data - Markov-switching models: parameters switch between regimes; estimation via maximum likelihood - Example: 2-state HMM with 'low-volatility mean-reverting' and 'high-volatility trending' states

**ML classifiers:** - Train a classifier (random forest, neural network) to predict the regime label from features - Labels come from a separate analytical process (e.g. label periods of historical trend persistence as 'trending') - Features: rolling volatility, autocorrelation, technical indicator values, time-of-day, etc.

**Unsupervised clustering:** - K-means or Gaussian mixture models on feature vectors to identify regime clusters without explicit labels - Useful when regime structure is not pre-specified

Applications in trading strategies:

**Parameter switching:** - Maintain different parameter sets for different regimes - Switch parameters when regime change is detected - Example: in trending regime use wide stop and trailing logic; in ranging regime use tight stops and fixed targets

**Strategy activation:** - Run multiple sub-strategies optimised for different regimes - Activate only the strategy(ies) matching the current regime - Example: trend-follower active when ADX > 25; mean-reverter active when ADX < 20; neither active in ambiguous middle

**Position sizing:** - Reduce position sizes during stressed or low-confidence regimes - Increase position sizes during high-confidence regimes - Example: during high-volatility regime, halve typical position size to maintain similar dollar volatility

Challenges: 1. Regime change is gradual; detection lags reality 2. Regime classifications are often subjective — the labels themselves are constructed by the developer 3. Regime detection adds complexity that can itself overfit 4. Past regime distributions may not match future regimes (regime regimes change too!)

For FxRobotEasy: our flagship AI EAs include regime-aware parameter switching, particularly Trendopedia which adjusts trailing-stop tightness based on detected volatility regime.