Can an exogenous predictor be random?

Yes; it may be modeled as externally generated and uncertain.

Can future exogenous values be unknown?

Yes, in which case forecasting or scenario assumptions are needed.

AI Glossary / Time series and signal processing / Intermediate

Exogenous variable

Definition

An exogenous variable is a variable whose value is treated as determined outside the model or system being analyzed. In forecasting, it often refers to an external predictor supplied to the model. In econometrics, exogeneity has more precise definitions concerning the relationship between predictors and model errors, so the intended meaning must be stated.

What it means

A demand model may use temperature or a published holiday calendar without modeling how the weather or calendar is generated. Those inputs are external to that demand model.
However, placing a variable on the right-hand side of a regression does not prove statistical exogeneity. A price, staffing level, or advertising budget may respond to unobserved demand conditions. Such feedback can matter greatly when coefficients are interpreted as effects rather than merely used for prediction.

How it works

Define the model boundary. A variable can be external to one model and explained within a larger model. The label depends on the system and question, not on an intrinsic property of the variable's name.
Distinguish a supplied predictor from a formal assumption. Some software calls its predictor matrix exog as a naming convention. That does not certify independence from errors, absence of confounding, or suitability for a causal analysis.
In a simple regression interpretation, a condition such as E(error|X) = 0 states that the error has zero conditional mean given the predictors. Different time-series settings require different assumptions; strict exogeneity can condition on the full sequence of regressors, which is stronger than a contemporaneous condition.
Check how the predictor is generated. Simultaneous decisions, omitted common causes, measurement error, and feedback from past outcomes can invalidate assumptions needed for a particular estimator or interpretation. A lagged variable is not automatically exogenous under every dynamic model.
Plan future values for forecasting. Some exogenous predictors, such as a known calendar indicator, are available in advance. Others, such as tomorrow's temperature, must be forecast or assigned scenarios. External to the model does not mean known without uncertainty.
Propagate uncertainty where relevant. A forecast interval conditional on a fixed predictor path does not include uncertainty about that path. Combining predictor uncertainty with outcome uncertainty requires assumptions about their joint behavior.
Preserve realistic information timing. A predictor observed after the forecast target period cannot be used as though it were available at the forecast origin. Historical data revisions and delayed publication can also introduce leakage.
Evaluate predictive and causal goals separately. An external predictor may improve forecasts without its coefficient measuring an intervention effect. Conversely, a scientifically important adjustment variable may be needed for interpretation even when its incremental predictive improvement is modest.

Three practical examples

A visitor-demand model uses a holiday indicator whose future values are already known.
An energy forecast uses a separate temperature forecast as an external input, carrying its uncertainty into planning.
An analyst investigates whether promotional spending responds to anticipated demand before interpreting its regression coefficient causally.

Think of it this way

Think of a simulator that receives tomorrow's weather as an input. The simulator explains what happens given that weather; it does not generate the weather itself. Calling the input external does not make the weather forecast exact or prove that every other supplied input is unrelated to hidden influences.

Worked example

Consider an original demand model with known coefficients: Y = 100 + 3T + 40H + error, where T is temperature and H is a holiday indicator. Assume the error has conditional mean zero.
For a known holiday, H = 1. If the forecast mean temperature is 20, the mean demand forecast is 100 + 3 x 20 + 40 = 200. A temperature scenario of 24 instead gives 212. The 12-unit change is the model's scenario sensitivity, not automatically a causal estimate.
Suppose temperature forecast error has standard deviation 2, outcome error has standard deviation 8, and the two errors are independent. With the holiday known and coefficients treated as fixed, total forecast-error variance is 3 squared x 2 squared + 8 squared = 100, giving standard deviation 10.
An interval calculation conditioning on temperature as exactly known would use only the outcome-error standard deviation 8. This illustrates why an external predictor's uncertainty matters. If the errors are correlated or coefficients are estimated, additional terms or uncertainty must be considered.

Illustrative case study

Fictional case: Willow Market adds planned promotions to a sales forecasting model. The forecast improves in a historical evaluation, but managers want to interpret the promotion coefficient as the gain from increasing spending.
Analysts discover that promotions were scheduled when managers expected weak demand using information absent from the dataset. They retain the variable's predictive role while explaining that its coefficient does not automatically identify the intervention effect. A separate design is needed for that question.

Common misconceptions and mistakes

Exogenous does not mean statistically independent of every other predictor. It does not mean the value is fixed, known in advance, or measured perfectly. Not every explanatory variable satisfies a formal exogeneity assumption. A software argument called exog is not a diagnostic result. Good forecast accuracy cannot by itself establish that a coefficient has a causal interpretation.

Business use

External predictors can improve planning when their values are available or forecastable and their relationships remain useful. Maintain ownership and release schedules for each input, document scenario assumptions, and monitor whether input forecasts degrade. When a business wants to change an input deliberately, assess whether the forecasting relationship can support that intervention question.

Check your understanding

In the example, which future predictor is assumed known?

Answer: The holiday indicator.

Why is total forecast-error standard deviation 10 rather than 8?

Answer: Temperature uncertainty adds variance under the stated independence assumption.

Does a 12-unit scenario difference prove a causal effect?

Answer: No.

Related terms

Confounding
Endogenous variable
explanatory variable
regression error
strict exogeneity
dynamic regression
predictor uncertainty
scenario forecast

Learn next

List which predictors are known, forecasted, or decided at each forecast origin. Then distinguish the assumptions needed for prediction from those needed to interpret a coefficient as an effect.

Sources

Hyndman-hosted Glossary of Forecasting Terms, Exogenous variable: https://robjhyndman.com/mwh3/FG4.pdf statsmodels, endog and exog terminology and formal-assumption caveat: https://www.statsmodels.org/stable/endog_exog.html Hyndman and Athanasopoulos, Correlation, causation and forecasting: https://otexts.com/fpp3/causality.html

Previous: Existential risk Next: Expand (pyramid operator)