Must an explanatory model be linear?

No. The relationship can be nonlinear or include interactions.

Can it include past values of the outcome?

Yes; dynamic models can combine those with other predictors.

Does a high-quality forecast prove the model explains the mechanism?

No. Predictive usefulness and mechanistic understanding are distinct.

AI Glossary / Statistics and probability / Intermediate

Explanatory model

Definition

An explanatory model represents an outcome as related to one or more other variables. In forecasting, the term commonly describes a model that uses external predictors, such as weather or workload, to forecast the variable of interest. In broader research, it can describe a model intended to understand relationships or mechanisms. Calling a model explanatory does not itself establish that those relationships are causal.

What it means

A forecast based only on past demand asks how the series has behaved. An explanatory demand model also uses variables such as temperature, calendar effects, or price. These predictors can improve understanding and prediction if their relationships are relevant and sufficiently stable.
The purpose should be explicit. Describing associations, predicting future outcomes, and estimating the consequences of an intervention are different tasks. The same equation may be useful for one and inadequate for another.

How it works

Define the outcome, observation unit, time horizon, and intended decision. Identify candidate predictors using subject knowledge and data availability. A predictor recorded after the outcome may explain history while being unusable for an advance forecast.
Specify a functional relationship. A linear regression might use y_t = β_0 + β_1x_(1,t) + β_2x_(2,t) + ε_t. Nonlinear terms, interactions, and lagged predictors may be needed when a straight additive relationship is unsuitable. The error term represents variation not captured by the specified systematic component.
Estimate parameters using data appropriate to the task. Inspect residual patterns, influential observations, missingness, and relationships between predictors. Highly correlated predictors can make individual coefficients unstable even when combined predictions remain useful.
For time-series data, check trends, seasonality, and serial dependence. Unrelated trending variables can appear strongly associated. A regression with autocorrelated errors may require a time-series error model and an evaluation procedure that respects temporal ordering.
To forecast, supply future predictor values. Some are known, such as calendar indicators. Others must be forecast or specified as scenarios. A prediction conditional on an assumed temperature or workload is not an unconditional forecast that includes uncertainty about those inputs.
Evaluate the model against relevant baselines using data unavailable during fitting. For causal interpretation, additionally justify the design and identification assumptions, such as appropriate control of confounding and a clear intervention. High R², plausible coefficients, or flexible AI algorithms cannot substitute for this justification.
Document units, transformations, coefficient interpretation, and the domain in which predictions are supported. Reassess the relationship when operating conditions change. Extrapolating far beyond the observed predictor combinations can be particularly unreliable.

Three practical examples

A utilities analyst models demand using temperature and calendar variables, supplying weather scenarios for future periods.
A support team forecasts incoming tickets using active customers, product launches, and recent incident indicators.
A research team models an outcome using measured characteristics to describe associations, while keeping causal claims separate from predictive fit.

Think of it this way

Think of a forecast as a recipe whose ingredients are predictor values. Knowing the recipe is not enough: future ingredients must also be known or assumed, and changing an ingredient in the equation does not automatically establish the effect of changing it in reality.

Worked example

Consider an invented daily-demand equation: ŷ = 120 + 4T + 30W, where T is temperature in degrees Celsius and W equals one on a weekday and zero otherwise. These coefficients are assumed for illustration, not estimated from real observations.
At T = 25 on a weekday, predicted demand is 120 + 4 × 25 + 30 = 250 units. At T = 28 on the same weekday, it is 262 units. The model's conditional difference is 12 units.
If the plausible temperature scenario ranges from 23 to 27, holding W = 1 gives predictions from 242 to 258. This 16-unit span is a scenario range, not a statistical prediction interval: it excludes residual variation and coefficient uncertainty.
The intercept describes the model's prediction at T = 0 and W = 0, which may be outside the relevant data range. The temperature coefficient is not automatically a causal effect of manipulating temperature.

Illustrative case study

Fictional case: Summit Services develops an explanatory model for daily support demand. A strong historical predictor turns out to be the number of agents who eventually handled tickets, a quantity determined after demand became known. The team removes this input, uses variables available at planning time, and evaluates on later weeks. This invented case illustrates why a convincing retrospective explanation may fail as a forecast.

Common misconceptions and mistakes

Do not equate an explanatory variable with a proven cause or assume that every predictor is exogenous. Confounding, feedback, and selection can affect coefficients.
Do not evaluate a real-time forecast using future predictor values that would not have been available. That answers a different question.
A scenario forecast is conditional on its assumptions. A narrow interval that ignores uncertain future inputs can understate the uncertainty relevant to a business decision.

Business use

Explanatory models can support scenario planning and reveal which measured conditions are associated with outcomes. They are useful when managers need to see how forecast inputs relate to expected demand or workload. For intervention decisions, such as changing prices, additional causal evidence is needed. Reports should separate associations, conditional scenarios, and justified causal conclusions.

Check your understanding

What input is known in advance in the example?

Answer: The calendar-based weekday indicator.

Why is 242-258 not a prediction interval?

Answer: It only varies an assumed temperature range.

Why might a coefficient change after adding another predictor?

Answer: It represents a conditional relationship within the specified model.

Related terms

Exogenous variable
Confounding
Explanatory variable
regression
predictor
dynamic regression
scenario forecast
causal model

Learn next

Study regression interpretation and predictor availability. Then learn time-series validation and the distinction between forecasting a relationship and identifying an intervention effect.

Time series

Sources

Glossary of Forecasting Terms, “Explanatory model”: https://robjhyndman.com/mwh3/FG4.pdf Hyndman and Athanasopoulos, Forecasting: Principles and Practice, time-series regression: https://otexts.com/fpp3/regression.html

Previous: Explanation-based learning Next: Explicit multi-threading