How is it different from a plain HMM?

Observations depend on previous observations as well as the hidden state.

How many states should I use?

Start with a small number guided by domain knowledge, and compare fits on held-out data.

No. It is a classical statistical model, though it can be part of larger systems.

AI Glossary / AI and machine learning / Advanced

Autoregressive Hidden Markov Model

Definition

An autoregressive hidden Markov model is a hidden Markov model in which each observation depends not only on the hidden state but also on the previous observations, letting it capture short-term patterns within each state.

What it means

A hidden Markov model, or HMM, describes a system that moves between hidden states, such as calm, busy or faulty, and produces observations whose statistics depend on the state. You see the observations but not the states.

In a plain HMM, observations are assumed independent once the state is known. Many real signals are not: today's temperature depends on yesterday's, and a machine's vibration this second depends on the last.

The autoregressive version adds a link from earlier observations to the current one, so each state has its own short-term dynamics, not just its own average level.

This helps when a state is better described by how the signal evolves than by its level alone, such as a steady oscillation or a slowly drifting pattern.

For a business, it is a specialist tool for time series where regimes shift, such as demand under different conditions or sensor data under different operating modes.

It sits in the family of switching models: the signal keeps its own short-term behavior, but which behavior applies changes over time. That matches many business series better than assuming one fixed pattern.

How it works

The model has a set of hidden states and probabilities of moving from one to another. Each state has a rule for generating the next observation.

In the autoregressive form, that rule uses the previous one or more observations: the next value is predicted from recent values, with parameters that differ by state.

Training estimates the transition probabilities and each state's autoregressive parameters from data, commonly with an iterative method that alternates between guessing the states and refitting the parameters.

Once trained, the model can infer which state the system is most likely in, detect when it switches, and forecast the next values using the current state's dynamics. More lag terms capture richer patterns but need more data.

Choosing the number of states is a judgment call. Too few and the model misses real regimes; too many and it invents patterns. Analysts usually try a small range and compare how well each version explains held-out data.

Three practical examples

Speech and sound: modeling segments where the signal has different short-term structure.
Finance: describing market periods with different volatility and momentum patterns.
Machinery: detecting a switch from normal to degraded operation from vibration data.

Think of it this way

Think of a river that is sometimes calm and sometimes in flood. In each mode, the water level tomorrow depends on today, but in a different way. The model tracks the mode and the day-to-day pattern together.

The limit: you only see water levels, not the mode, so the model infers it. Wrong assumptions about the number of modes lead to confident but misleading stories.

Worked example

spotting regime shifts in daily orders

Suppose a retailer has daily order counts that behave differently in normal weeks and in promotion weeks. In promotions, orders spike and then decay quickly; in normal weeks they move gently around a steady level.

A plain average cannot capture both. An autoregressive HMM with two states lets each have its own dependence on yesterday's orders.

After fitting, the model labels past days as normal or promotion-like, and the retailer checks whether the labels match its real promotion calendar. The data and results here are illustrative.

If the labels match, the model can help flag unplanned surges early; if not, the states may be capturing something else, such as holidays.

Illustrative case study

A fictional energy firm models smart-meter data with two hidden states, working-day and holiday usage, each with its own short-term pattern.

The model forecasts better than one global model on holiday weeks. A reviewer finds, however, that a third state, an outage period, is being absorbed into holiday. The firm adds a state and checks fit against outage records.

The invented lesson: hidden states are model constructs, and checking them against known events keeps them honest.

Common misconceptions and mistakes

The states are directly observed. They are hidden and inferred.
More lags are always better. More lags add parameters and can overfit short histories.
It is a general-purpose forecasting tool. It is specialized and needs reasoned choices about states and lags.

Business use

Consider it only for time series with clear regimes where short-term dependence matters; for most forecasting, simpler models are easier to explain and maintain.

Compare against a plain autoregressive model and a simple benchmark, check inferred states against real events, and document the assumptions.

Check your understanding

What does the autoregressive part add to a hidden Markov model?

Answer: It lets each observation depend on earlier observations, so each state has its own short-term dynamics, not just its own average level.

Related terms

Hidden Markov model
Time series
Forecasting
autoregression
state inference
regime switching

Learn next

AI for Business Owners helps you decide when advanced forecasting is worth it and when a simple method will do: https://aiforbusinessandlife.com/academy/ai-for-business-owners. [C2]

Forecasting

Sources

[S13] Minka, Statistical Learning/Pattern Recognition Glossary. Full URLs are in the source ledger. Verified 11 October 2026.

Previous: Autoregressive flow Next: Autoregressive image model