Data & drivers
Driver discovery: what actually moves the number
A driver is a series that improves the forecast when you are allowed to see it — lagged, tested, and still available at 6am. Correlation on a heatmap is not discovery. Discovery is a backtest with the candidate in and out.
Updated Aug 26, 2026·8 min read
Correlation is not a driver
A heatmap will light up anything that moved with the target last quarter, including the target’s own echo and a coincidental campaign. A driver is a series that improves the forecast when you are allowed to see it — lagged, tested out of sample, and still available at 6am. Discovery is a backtest with the candidate in and out. Everything else is interior design.
The lag is the product
Same-day temperature for same-day load is a nowcast input, and you only have a weather forecast at origin. Promo flags you already approved are known ahead. Fuel cost that shows up in shipping rates three weeks later is a path. If you cannot say the delay in one sentence, you do not have a driver. You have a correlation.
- Known-ahead: calendars, approved prices, booked events.
- Leading: something you will actually have, delayed relative to the target.
- Leakage: anything that arrives with the official number, or is a restatement of it.
Indirect paths
Fuel → freight → landed cost → price → demand. A graph that only allows one hop will miss it. A graph that allows everything will overfit a chain of accidents. You want scored multi-step paths with the rejects left visible, so a person can say “that hop is nonsense” without pretending the search never happened.
Scored, lagged, and left visible
Predict.ai searches the workspace for signals that lead your target, scores lag and strength, and leaves the rejects on the relationship graph. You can argue with it. You cannot quietly drop a driver that failed and keep the one that flattered a training window.
That is the gap with a SageMaker notebook that “includes weather.” The notebook does not keep a museum of what it tried. A goal does. When the path jumps, the analyst can say which driver moved — or that nothing did, and you should look at the pipe.
FAQ
- How do you know a driver is real?
- It reduces error on a walk-forward backtest when added to the target’s own lags, and you can say in one sentence why the delay exists. If it only fits the training window, it is a coincidence.
- Is driver discovery the same as causality?
- No. A useful driver can be a leading proxy, not a cause. Causality is a stronger claim. Forecasts need drivers that will be known at origin. Save the causal debate for the intervention you actually control.
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
Forecasting with covariates
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.
Guide
Energy load forecasting
Load, generation, and price live on a tight clock. Weather is a covariate you only get as a forecast. A skinny band at hour 36 is swagger. Operators plan reserves on a high quantile, not on a pretty MAE.
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.