How can I tell whether a bar demand forecast is useful?
By Better Bartender · Published
Test a forecast on later periods it has not seen, compare it with a simple baseline and measure the cost of getting it wrong. Predicting 100 serves when 120 sell creates a 20-serve error; a convincing explanation does not change that result.
Choose a decision and a unit first
Decide what the prediction will support: preparing a batch, planning a delivery or reviewing likely category demand. Forecast a defined quantity for a named horizon, venue or account group. Recorded serves, litres, revenue and customer visits are different targets. Do not change the target halfway through judging performance.
Write down when the forecast is made and what information is available at that moment. A next-Friday prediction should not use Friday’s eventual weather or confirmed sales as if they were known in advance. Keep closures, stockouts and missing data distinct from ordinary zero demand.
Start with a baseline worth beating
A previous comparable weekday or a rolling average of comparable weekdays gives a practical starting point. It may be imperfect, but it exposes whether a more complex model adds value. Use the same horizon, data availability and evaluation dates for both methods.
For seasonal or event-sensitive trading, retain the operational notes that explain unusual periods. If there is little history for a new drink, show that limitation. Do not manufacture detailed confidence from a handful of observations.
A worked forecast comparison
These are invented results for four held-out periods. Actual serves are 100, 120, 80 and 100. A candidate predicts 90, 110, 100 and 100. Its absolute errors are 10, 10, 20 and zero: a total of 40 serves. Mean absolute error is 40 ÷ 4 = 10 serves per period.
A baseline predicts 100, 100, 100 and 100, producing absolute errors of zero, 20, 20 and zero. It has the same mean absolute error. On this small test, the candidate has not beaten the baseline on that measure.
Weighted absolute percentage error for either is 40 ÷ (100 + 120 + 80 + 100) × 100 = 10%. This definition divides total absolute error by total actual demand. It is undefined when total actual demand is zero and can conceal poor performance on individual items.
Connect error to the operating decision
Under-preparing a popular batch and over-preparing a perishable one have different costs. Inspect bias, large misses and results by venue, item and trading context. A model can improve an overall metric while making one commercially important category worse.
Test across multiple later windows and retain the original forecast issued at each point. Replacing yesterday’s forecast with a retrospectively improved one makes the record unusable. A scenario such as “what if ten accounts adopt the drink?” is arithmetic under assumptions, not a prediction of adoption.
Where predictive intelligence fits at Better Bartender
Predictive engines are part of Better Bartender’s development programme for bars and drinks brands. This guide sets out how to judge that work; no customer forecast-accuracy result is claimed by this illustrative example. Explore our AI-native approach and current reporting boundaries.
Further reading: Hyndman and Athanasopoulos, Forecasting: Principles and Practice — evaluating forecast accuracy, on testing forecasts using data outside the training sample. The numerical example and business checklist above are Better Bartender’s own.