QUANTA - LAB HMM REGIME DETECTION

Two-state Hidden Markov Model for market regime detection based on Hamilton (1989) Markov-Switching framework.
Methodology:

Full Baum-Welch EM algorithm in log-space for numerical stability
Real-time Hamilton filtering (no lookahead) for trading use
Kim smoothing for historical analysis
Multiple random restarts to avoid local optima

Regime Classification:

Mean-based: R1 = Bearish (lower μ), R2 = Bullish (higher μ)
Volatility-based: R1 = Calm (lower σ), R2 = Turbulent (higher σ)

Key Features:

TRADING vs ANALYSIS mode (filtered vs smoothed probabilities)
Gaussian assumption diagnostics (kurtosis, skewness, outliers)
Data Quality Score (0-100)
Regime Certainty Index (RCI)
Mean separation t-statistic
Expected regime duration and ergodic probabilities
Degenerate model detection

Dashboard Includes:

Filtered probabilities (real-time, safe for trading)
Emission parameters (μ₁, μ₂, σ₁, σ₂)
Transition matrix (p₁₁, p₂₂)
Model fit metrics (LogL, AIC, BIC)

Critical Warnings:

Smoothed ≠ Real-time (smoothed uses future info)
Gaussian assumption: fat tails not captured
K=2 regimes only — may oversimplify dynamics
NOT for high-frequency (minimum 1H timeframe)
Validate with Python hmmlearn / R / MATLAB

References: Hamilton (1989) — Econometrica