Gaussian Mixture Models

Gaussian Mixture Models fits a Gaussian mixture model to recent returns and volatility, plotting each fitted regime's probability for the current bar from 0 to 100% in a dedicated pane. Components are re-estimated by expectation-maximization over a rolling 200-observation window and ordered by variance, so the first line is always the calm regime and the last the stressed one. Candles tint by regime, opacity scaling on confidence; a dashboard reports weights, means and EM status.

How to Trade the Gaussian Mixture Models?

• Calm regime dominant: the lowest-variance component leads — the backdrop for quiet-market tactics.
• Stressed regime dominant: the highest-variance component leads — reduce risk; it is not a directional call.
• Middling probabilities: the bar sits between regimes; wait for a confident reading.
• Low-Likelihood Anomaly alert: the bar resembles nothing in the training window — treat labels with suspicion.

Alerts fire on each dominant-regime change; as across the machine learning family, the output describes the fitted window, never the future.

Gaussian Mixture Models Settings

• Components (default 2): 3 adds a middle transition tier.
• Features (default Returns + Volatility): the observation vector.
• Return Source (default close): feeds the log returns.
• Volatility Measure (default Stdev of Returns) with Volatility Length (default 14): the volatility feature and its lookback; or ATR %.
• Training Window (default 200): rolling observations per fit.
• Max EM Steps per Bar (default 10): update cap per bar.
• Update Mode (default Incremental (Warm Start)): or a full refit each bar.
• Probability Smoothing (default 3): EMA on the probabilities.
• Anomaly Percentile (default 5): likelihood cutoff for anomaly flags.
• Fit on Closed Bars (default enabled): confirmed bars only.
• Dashboard & style: placement, gradient, bar coloring, anomaly markers.

Frequently Asked Questions

How do Gaussian Mixture Models differ from Spectral Clustering?

The mixture is a probability model: each bar gets soft membership odds per regime. Spectral Clustering also groups behavior without labels, but commits to one hard label per bar.

Do the probabilities forecast direction?

No. They report how the current bar compares with the recent training window — context for sizing and strategy selection, not a directional call.

Why do readings drift over time?

The window rolls: each new observation replaces the oldest and the mixture is re-estimated, so parameters drift with the market. Variance ordering pins the labels; boundaries travel with the fit.