Is TAR the same as SETAR?

SETAR is the case where the switching variable comes from lagged values of the series itself.

Is switching always smooth?

No. Hard threshold models can have abrupt changes; smooth-transition models use a different transition mechanism.

Does a better training fit prove better forecasting?

No. Additional flexibility can fit historical variation that does not recur.

Threshold Autoregressive

Definition

A threshold autoregressive, or TAR, model is a time-series model whose autoregressive rule changes according to a threshold variable. Within each regime, the next value can depend linearly on earlier values, but switching between regimes makes the overall relationship nonlinear. When a lagged value of the same series determines the regime, the model is commonly called self-exciting threshold autoregressive, or SETAR.

What it means

An ordinary linear autoregression uses one set of coefficients across the observations being modeled. A TAR model allows different dynamics in different conditions. For example, a series may return toward its usual level differently after a large deviation than after a small one.
The threshold defines which rule applies; it does not by itself explain why the dynamics differ. A fitted threshold is a modeling result, not automatic evidence of a causal boundary in the world.
Although the supplied category is Deep learning, TAR is a statistical time-series model and does not inherently require a neural network.

How it works

For a two-regime example, define a threshold variable z and threshold r. When z is at or below r, use one intercept and set of lag coefficients; when z exceeds r, use another. Add an innovation term to represent variation not explained by the conditional rule.
The threshold variable, lag order, delay and boundary convention must be specified. Regimes can have different coefficients and potentially different noise behavior. A SETAR choice uses an earlier value of the modeled series as z; broader threshold formulations can use another variable.
For fixed threshold assignments, coefficients can be fitted through conditional least squares. Searching for a threshold adds a model-selection step. Implementations can restrict the search so each regime contains enough observations.

Three practical examples

An operations analyst investigates whether an equipment-load series behaves differently above a specified load level. The comparison asks whether a threshold model improves forecasting over a single autoregression.
A research exercise models deviations in daily demand using different lag relationships for positive and negative deviations. The analyst checks whether both regimes have enough observations to estimate their coefficients.
A classroom simulation generates a series from two known rules. Students fit a threshold model and examine how errors in the estimated threshold affect predictions near the switching boundary.

Think of it this way

Imagine a driving simulator that uses one response rule below a speed boundary and another above it. Each rule may be simple, yet the combined response changes when the boundary is crossed.
The analogy explains piecewise behavior, not a claim that every real process contains a sharp switch. If the actual transition is gradual, imposing a hard threshold may be an unhelpful approximation.

Worked example

Consider an original SETAR teaching model with threshold variable equal to yesterday's value. If y yesterday is at or below 10, today's conditional mean is 2 + 0.5 × y yesterday. If it is above 10, today's conditional mean is 1 + 0.8 × y yesterday. Assume the innovation has conditional mean zero.
For yesterday's value 8, the forecast is 2 + 0.5 × 8 = 6. At exactly 10, the lower rule applies and gives 7. At 12, the upper rule applies and gives 1 + 0.8 × 12 = 10.6.
Just above the boundary, at 10.1, the forecast becomes 9.08. The difference between 7 at 10 and 9.08 at 10.1 illustrates that this specified conditional-mean function is discontinuous. Threshold models need not join smoothly unless constraints impose that property.
Suppose today's actual value after yesterday's 12 is 11.4. The one-step forecast error is 11.4 − 10.6 = 0.8. For a later horizon, uncertainty about future values can also create uncertainty about which regime will apply; simply substituting forecast means is not generally an exact multi-step conditional expectation.

Illustrative case study

Fictional Cedar Operations models hourly workload. A threshold model appears to fit the historical series better than a single autoregression, but its upper regime contains very few observations.
The analysts inspect the regime counts and discover that the apparent improvement depends heavily on two unusually busy hours. They compare thresholds with adequate support, retain a simpler baseline and evaluate forecasts in chronological order using only information available at each forecasting time.
The team reports overall error and errors within each regime. If the threshold model remains unreliable during high-load periods, a lower average error does not settle the operational decision. The exercise helps distinguish an attractive historical fit from a model that supports the intended planning task.

Common misconceptions and mistakes

A threshold crossing is not necessarily a permanent break at a particular date. The series can move between regimes repeatedly as the threshold variable changes.
The threshold variable need not be the immediately preceding observation. Its definition and delay are part of the specification.
Piecewise linearity does not make the whole model an ordinary linear autoregression. Selecting different rules changes the overall relationship.
An estimated threshold does not authorize changing the underlying process to force a preferred regime. Predictive association and the effects of intervention are different questions.

Business use

Use TAR when a clear forecasting question suggests that one dynamic relationship may be inadequate. State what would count as an operationally useful improvement before searching many thresholds.
Review the number and range of observations in each regime, the availability of the threshold variable at prediction time and the stability of results across evaluation periods. Avoid selecting the final model using future information unavailable during deployment.
Document the boundary convention explicitly. Two implementations that place equality on opposite sides can produce different forecasts exactly at the threshold, even if their displayed coefficients appear identical.

Check your understanding

Which rule applies at 10 in the worked example?

Answer: The lower rule, because its condition includes equality.

What is the forecast after a value of 12?

Answer: 10.6.

Why inspect regime counts?

Answer: A regime with very few observations can yield unstable estimates and misleading apparent improvements.

Related terms

Autoregression
SETAR
regime switching
threshold variable
lag
conditional least squares
smooth-transition autoregression
time-series validation

Learn next

Calculate forecasts on both sides of a chosen threshold. Then compare a threshold model with a single-rule baseline using chronological evaluation, keeping threshold selection inside the training process.

Sources

tsDyn, SETAR documentation: model equation, threshold specification and conditional least-squares estimation. https://search.r-project.org/CRAN/refmans/tsDyn/html/setar.html Nikolay Balov, Bayesian threshold autoregressive models, Stata technical explanation of TAR regimes and SETAR. https://blog.stata.com/2022/05/18/bayesian-threshold-autoregressive-models/ The numerical model and fictional operational case are original teaching illustrations.

Previous: Three-tiered architecture Next: Threshold (for decision trees)