Class RegimeDetector

java.lang.Object

com.quantfinlib.ml.RegimeDetector

public final class RegimeDetector
extends Object

Two-state Gaussian Markov-switching model (hidden Markov model) fitted by Baum-Welch EM with scaled forward-backward recursions (Rabiner 1989, Proc. IEEE 77, 257-286; Hamilton 1989, Econometrica 57, 357-384). Detects calm/turbulent regimes in a return series. State 1 is always the high-volatility regime.

Filtered versus smoothed. The model reports two probability paths. filteredHighVolProbability[t] = P(high vol at t | r_0..r_t) uses only data up to t and is the one a strategy may act on at t (after observing r_t). smoothedHighVolProbability[t] = P(high vol at t | r_0..r_{n−1}) uses the whole sample, including data after t; it is for labelling regimes after the fact. De-levering on the smoothed path in a backtest is look-ahead. Note that the parameters themselves are fitted on the whole sample, so even the filtered path is only out-of-sample when the model was fitted on an earlier window.

EM is local: fit(double[], int) runs from the median-split start plus four seeded random starts and keeps the highest likelihood. Emission densities are handled in log space with a per-period shift, so an extreme return cannot underflow both densities to zero. Inputs must be finite.

Nested Class Summary

Nested Classes

Modifier and Type	Class	Description
static final class	RegimeDetector.RegimeModel

Method Summary

Modifier and Type	Method	Description
static RegimeDetector.RegimeModel	fit(double[] returns, int maxIterations)	Fits with five starts (median split plus four seeded random starts).
static RegimeDetector.RegimeModel	fit(double[] returns, int maxIterations, int starts, long seed)	Fits the regime model by EM from several starts.
static int[]	viterbi(double[] returns, RegimeDetector.RegimeModel model)	Most likely state sequence (Viterbi, in log space) under a fitted model; uses the whole sample, so it is an ex-post labelling.

Methods inherited from class Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Method Details

fit

public static RegimeDetector.RegimeModel fit(double[] returns,
 int maxIterations)

Fits with five starts (median split plus four seeded random starts).

fit

public static RegimeDetector.RegimeModel fit(double[] returns,
 int maxIterations,
 int starts,
 long seed)

Fits the regime model by EM from several starts.

Parameters:

maxIterations - EM iterations per start (0 evaluates the start)

starts - number of starting points, ≥ 1; the first is always the median split, the rest are random

seed - seed for the random starts

viterbi

public static int[] viterbi(double[] returns,
 RegimeDetector.RegimeModel model)

Most likely state sequence (Viterbi, in log space) under a fitted model; uses the whole sample, so it is an ex-post labelling.