When a SARIMAX model includes exogenous regressors, what must you supply to forecast?
answer
- the model predicts y, not x
- an h-step forecast needs h rows of something
- some drivers are known, some must be guessed
- the interval treats supplied inputs as certain
basics
~20 sFuture values of every exogenous regressor, for every period you forecast. A SARIMAX forecast is conditional on those inputs, so calendar or planned values are safe while values you must forecast yourself add error the intervals never show.
solid answer
~50 sSARIMAX models the series as a regression on external drivers with ARIMA errors, so it never forecasts those drivers for you. To predict twelve weeks ahead with a promotion flag and a price column, you must hand it twelve future promotion flags and twelve future prices. Some drivers are genuinely known ahead — holidays, a signed promotion calendar, a price you control — and those are the ideal case, turning the model into a scenario tool: run it under two promotion plans and compare. Others, like a competitor's price or a macro index, have to be forecast themselves, which stacks one model's error on another. Worse, the reported prediction intervals treat the supplied x as certain, so they understate real uncertainty. The usual fix is to enter the driver at a lag longer than the horizon so the values you need are already observed.
go deeper
Remember the core fact: external regressors have to be provided for every future period you want a forecast for. The model predicts the target only, never the inputs it is given.
Explain the regression-with-ARIMA-errors structure and sort drivers into calendar-known, business-decided and genuinely unknown, saying what each case implies for the forecast you can honestly produce.
Demonstrate the interval caveat unprompted, and offer a concrete fix such as lagging the driver beyond the horizon. Mentioning collinearity between a promotion flag and price shows you have fitted these in anger.
Frame it as a decision-support question: argue when the right deliverable is a scenario comparison rather than a point forecast, and set the standard for how conditional forecasts are communicated to stakeholders.
## SARIMAX conditions on x; it does not predict x The X in SARIMAX is a block of exogenous regressors — external columns like a promotion flag, a price, a holiday indicator, a temperature. The standard formulation is a regression with ARIMA errors: `y_t = beta' x_t + n_t`, where `n_t` follows a SARIMA process. The systematic part is ordinary regression on the drivers; the SARIMA machinery is applied to what the regression leaves over. Because `x_t` sits on the right-hand side, producing `y` for a future period requires `x` for that same future period. The model has no opinion about where those values come from. ### Three kinds of driver, three different situations **Known by construction.** Holiday and calendar indicators, day-of-week flags, harmonic terms for a long cycle, a school-term dummy. These are deterministic functions of the date, so the future block is exact and free. This is the case where exogenous regressors are unambiguously worth adding. **Chosen by the business.** A promotion calendar, a planned price change, a marketing spend commitment. These are known if a decision has been made, and if it has not, that is a feature rather than a problem: the model becomes a scenario engine. Forecast under "no promotion", forecast under "two-week promotion in March", and present the pair. The comparison, not the point forecast, is what the model is really for. **Genuinely unknown.** A competitor's price, an index, a weather variable beyond the reliable forecast window. Here you must predict the predictor. That stacks errors: your `x` forecast is wrong by some amount, the coefficient multiplies that error, and it lands in your `y` forecast on top of the SARIMA error. Sometimes it is still worth it; often it is not, and a model without the driver forecasts better despite fitting worse in-sample. ### The interval problem This is the point that separates candidates. A SARIMAX prediction interval is conditional on the supplied `x`: it propagates the uncertainty of the ARIMA errors and the estimated coefficients, but it treats the future regressor values as facts. If you fed in a forecast of `x`, the interval is too narrow, and by an amount nothing in the output reveals. If a stakeholder reads that band as the full range of outcomes, they are being misled by construction. Either say so explicitly, or widen the picture by re-running under a range of `x` paths and reporting the envelope. ### Making the future values available The cleanest structural fix is **lagging the driver**. If you enter `x_{t-8}` instead of `x_t` and forecast at most eight periods ahead, every regressor value you need has already been observed. You lose contemporaneous effects, but you gain a forecast that does not depend on a second forecast. The lag choice should be defended by the mechanism — how long a driver takes to move the outcome — not chosen to fit. A second option is replacing the unknown driver with a deterministic proxy. If a macro index mostly encodes a slow trend, a trend term or harmonic regressors reproduce most of the value with no future-value problem at all. ### Other practical cautions A promotion flag and a price column are usually badly collinear, because price drops during promotions. Both coefficients then get large standard errors and unstable signs, and the model may forecast acceptably while telling you nothing trustworthy about either driver. Differencing must be applied consistently: if the model differences `y`, the same differencing has to reach `x`, or the regression relates quantities on different scales. Formulations differ in whether they do this for you, so it is worth being explicit about which quantity a coefficient is attached to — a level effect and an effect-on-changes are very different claims. And the coefficients are associations from observational history. A promotion coefficient estimated on a schedule chosen by a merchandising team reflects when they chose to promote as much as what promoting does. Treating it as the causal effect of a promotion you have not yet run is a leap the model does not license. ### The short answer to give Name the requirement first — future values of every regressor for every forecast period — then classify the drivers as known, decided, or unknown, then raise the interval caveat. That sequence is what an interviewer is listening for.
- How do you handle a driver whose future values you cannot know?Three routes. Enter it at a lag longer than the forecast horizon so the values you need are already observed; replace it with a deterministic proxy such as a trend or calendar term; or accept the unknown and produce scenario forecasts over a set of assumed paths, reporting the envelope rather than one line.
- Do SARIMAX prediction intervals account for uncertainty in a forecasted regressor?No. They are conditional on the regressor values you supply, propagating only the ARIMA innovation variance and coefficient uncertainty. If those inputs were themselves forecast, the interval is narrower than the truth by an amount the output never shows, so the caveat has to be stated by you.
- Why is including a promotion flag and a price column together often problematic?Because price typically falls during promotions, the two columns are strongly collinear. Coefficients then have large standard errors and can flip sign with small data changes. Forecasts may still be fine, but attributing effect to one driver rather than the other is not defensible from that fit.
It is like a route estimate that depends on tomorrow's traffic: the model turns traffic into arrival time, but somebody still has to tell it what tomorrow's traffic will be.
saying these in an interview costs you the question
- Assumes the model forecasts its own regressors
- Presents intervals as full uncertainty when x was forecast
- Adds a driver only measurable after the fact
- Reads a SARIMAX coefficient as a causal effect
- Ignores that differencing must reach the regressors too