Data & drivers
Forecasting with covariates: weather, promotions, and prices
A covariate is an extra series the model is allowed to see — a promo flag, a temperature, a price. Used well, it explains swings the target's past cannot. Used as a dump of every column you have, it just overfits last quarter.
Updated Aug 17, 2026·9 min read
What a covariate is
A covariate is an extra series the model is allowed to see besides the target. Price. A promo flag. Temperature. Spend. Economists say exogenous. Same job: explain a move that the target’s own past cannot.
Used well, a covariate turns a mystery spike into a known lever. Used as a dump of every column in the warehouse, it memorizes last quarter. Add them one at a time. Keep the ones that move the backtest. Retire the rest without a funeral.
Target + a covariate on the same clock
Must be known ahead
This is the whole contract. If you will not have the value at forecast origin, you may not train as if you will. Tomorrow’s actual temperature is cheating. This morning’s weather forecast is fair. A promo you have not approved is not a feature; it is a scenario. Run it as a what-if, or leave it out.
The usual suspects
- Calendar: holidays, payday, school terms. Cheap and honest.
- Price and promo depth — if you can timestamp them.
- Weather: for melt, freeze, energy, outdoor demand.
- Marketing: only if the spend is dated, not a monthly lump.
Promotions are the classic mess. A clean flag with depth and coverage helps. A “campaign” field that also includes clearance, stockouts, and a CEO visit will teach superstition. Start with one well-defined lift driver.
Too many inputs
Each extra series is a chance to overfit an accident. Correlated weather products, five flavors of “holiday,” lagged everything — the model will look brilliant in-sample and confused in June. A tournament that includes a univariate baseline will keep you honest. If the kitchen-sink model cannot beat last week plus a calendar, you added noise.
FAQ
- What is a covariate in forecasting?
- An extra time series the model can use besides the target. Price, promo, holiday, temperature, marketing spend. Some people say exogenous variable. Same job: explain a move the lag of the target cannot.
- Can I use tomorrow's weather in today's forecast?
- You can use a weather forecast, not the observation you will only have tomorrow. If the covariate is not available at prediction time, it is leakage dressed up as a feature.
- Do promotions always help?
- A clean promo flag with depth and coverage helps. A messy 'campaign' field that also includes clearance and stockouts will teach the model superstition. Start with one well-defined lift driver.
Keep going
Guide
Finding leading indicators
A leading indicator moves before the number you care about. Search queries before orders, bookings before departures, parts receipts before output. The lag is the product; correlation on the same day is just a mirror.
Guide
Feature engineering for time series
Most of the work is turning a clock and a few raw series into lags, rolling stats, and calendar flags the model can chew. Do that on a schedule, with no future leaking into the past, or the backtest is theater.
Guide
What is demand forecasting?
Demand forecasting estimates how much of a product or service people will want in a coming period, so you can buy, staff, and price before the rush — not after it has already walked in.
Use case
Retail demand forecasting
Anticipate demand by product, store, channel, and region before buying or allocation decisions are locked.
Use case
Energy load forecasting
Forecast demand by interval, feeder, zone, and customer class with weather-driven uncertainty.