The intuition is right, the benefit is smaller
Nobody selling ice cream, garden furniture or hot drinks doubts that weather moves demand. The question for a forecasting project is narrower: does adding weather data make the forecast better than one that already knows the date?
Often it does less than expected, because seasonality already captures most of it. A model that knows it is July has implicitly learned that July is warm. Weather only adds value where it deviates from the seasonal norm.
The forecast-of-a-forecast problem
To use weather in a forecast covering next month, you need next month's weather, which nobody has. You have a weather forecast, and its accuracy decays quickly.
That places a hard ceiling on the useful horizon. Within a few days ahead, weather forecasts are good and the benefit is real. Beyond a week or two, you are adding an uncertain input to an uncertain model and usually making it worse.
- Short horizon, weather-sensitive category - genuine, measurable benefit
- Long horizon - use seasonal norms rather than a weather forecast
- Backtesting - use the forecast that was available then, not the weather that actually happened, or you will badly overstate the benefit
That last point catches many projects. Backtesting with actual historical weather shows an impressive uplift that cannot be reproduced live, because live you only have the forecast.
Deviation from normal is the useful variable
Raw temperature is usually the wrong feature. What moves demand is the difference from what is normal for that place at that time of year - the first warm weekend in spring does more to garden centre trade than an equally warm day in August.
Threshold and change features tend to work better than the raw number: days above a temperature that matters for your category, consecutive dry days, a sharp drop from the previous week.
Test before you buy
Weather data feeds cost money and add a dependency that can fail. Establish the benefit before committing.
- Build the forecast without weather and record its accuracy honestly.
- Add weather using historical forecast data, not observed actuals.
- Compare at the horizon you actually operate at.
- Check the uplift is concentrated where you expected - if a feed improves accuracy on categories with no plausible weather link, suspect a leak.
Within three days, weather is information. Beyond three weeks, it is a guess wrapped in a model.