HMM Regime Pass Node

Hidden Markov Model Regime Detection — Series Input

AdvancedRegimePass

Overview

The HMM Regime Pass Node applies a Hidden Markov Model to a series input to classify market regimes. It infers latent states (e.g., trending up, trending down, ranging) from observed return patterns — assigning each bar a regime label from 0 to numStates−1.

Unlike threshold-based regime filters, HMM learns the statistical properties of each regime from the data and accounts for the probability of transitioning between regimes. As a Pass node it operates on any numeric series, enabling regime classification of derived indicator streams or custom price transformations.

Algorithm

1. Compute rolling returns over the last `period` bars

2. Fit Gaussian HMM with `numStates` hidden states using Baum-Welch EM

3. Decode most-likely state sequence using Viterbi algorithm

4. Output: state label [0, numStates−1] for each bar

Output: integer labels [0, numStates−1]. Warm-up: period bars.

State 0 typically maps to lowest-mean regime; state ordering is data-driven.

Parameters

Parameter	Default	Description
period	100	Rolling window of bars used to fit the HMM at each step
numStates	3	Number of hidden market regimes to detect (2–5 recommended)

Inputs & Outputs

Slot	Direction	Type	Description
input	Input	{ values, timestamps }	Any numeric series for regime classification
values	Output	(number \| null)[]	Regime labels [0, numStates−1]; nulls during warm-up
timestamps	Output	number[]	Unix timestamps aligned to input

Use Cases

Regime-Conditional Strategy Switching

Map state labels to strategy modes — e.g., state 0 = ranging (activate mean-reversion), state 1 = low-volatility trend (activate trend-following), state 2 = high-volatility (reduce position size or step aside).

Volatility Regime Detection

Feed a volatility series (ATR, realized volatility) into HMM Pass to identify low, medium, and high volatility regimes — enabling dynamic position sizing that scales down exposure in detected high-volatility states.

Regime Persistence Filter

Use regime state persistence (how many bars the current state has been active) as a confidence filter — only act on signals that have been in a consistent regime for multiple bars, filtering one-bar false positives.

Tips & Best Practices

State Label Ordering Is Not Fixed

HMM state labels (0, 1, 2) are assigned arbitrarily by the algorithm — state 0 may not always be "bearish." Post-process state labels by comparing each state's mean return to assign semantic meaning (bullish/bearish/neutral).

Large Period Required

HMM fitting requires sufficient data — the default period=100 is a minimum for 3 states. For more states (4–5), use period=150–200. Too few bars per state leads to unstable, frequently shifting regime estimates.

Computationally Intensive

HMM fitting (Baum-Welch EM) runs at each bar update — this is expensive compared to most indicators. If performance is a concern, increase the period to reduce re-fitting frequency or use this on higher timeframes only.