Forecasting against a clock
Beer and most drinks have a shelf life, and production has a lead time measured in weeks. Brew too much and it ages out; too little and you miss the sales window in a category where availability drives trial.
That combination makes forecast error expensive in both directions, which is exactly the situation where a proper quantile approach beats an average plus a buffer.
Where weather actually helps
Drinks demand is genuinely weather sensitive, particularly in on-trade and for lighter styles. Unlike many categories where weather adds little beyond seasonality, here it is a real driver.
The constraint is the forecast horizon. Weather is forecastable a few days out; production lead times are weeks. So weather helps with short-term allocation and distribution rather than with brewing decisions.
- Brewing decisions - use seasonal norms, not weather forecasts
- Distribution and allocation - weather within a few days is genuinely useful
- On-trade replenishment - short horizon, high weather sensitivity
- Event-driven demand - sporting fixtures, festivals, bank holidays, known well ahead
Events are knowable and often unused
Major sporting fixtures, local festivals and bank holidays drive substantial demand changes and are known months in advance. They frequently do not appear in the forecasting data at all.
Building an events calendar with location and expected scale is straightforward and usually improves forecasts more than any modelling change. It is also reusable across planning, sales and marketing.
The brewhouse data nobody analyses
Modern brewing plant records a great deal: temperatures, gravities, timings, pressures, yields. It is used for process control and quality assurance and rarely analysed across batches.
| Question | What the data could show |
|---|---|
| Why does yield vary between batches? | Process or ingredient factors correlating with loss |
| Which batches have shorter stability? | Process signatures preceding quality issues |
| Is a vessel underperforming? | Systematic differences by equipment |
| Does raw material lot affect outcome? | Supplier and lot correlation with results |
Yield variation is usually the most immediately valuable. Small consistent losses across many batches accumulate, and the cause is often visible in data already being recorded.
Small producers, small data
A craft producer brewing a few hundred batches a year has limited data for batch-level analysis, and that is a real constraint on what is possible.
Demand forecasting still works because sales data has many more observations than batches. Process analysis needs more patience - which is an argument for recording well now so the analysis is possible in two years.
Weather helps you decide what to deliver on Friday. It cannot help you decide what to brew in March.