Demand forecasting for bars: baselines that beat gut feel
By Jack Hawkins · Updated
Bar demand forecasting is predicting next week's hourly sales well enough to buy, roster and prep against, and a useful baseline to compare with a manager's judgement is straightforward: average the same weekday across the last eight weeks, adjust for the two or three known events in the coming week, and hold the residual error with a buffer policy instead of a prediction. Venues that forecast badly aren't bad at guessing; they're forecasting the wrong thing, weekly, at the wrong grain.
The honest frame first: bar demand is noisy, weather- and event-driven at the daily level, and anyone selling you a model that predicts Saturday to the pound is selling the confidence, not the accuracy. The goal isn't a perfect number; it's being less wrong than a spreadsheet-sized version of yourself, on purpose, in the places where wrongness costs money: the keg order, the rota, and the batched juice that browns on Tuesday.
A baseline you can measure against alternatives
For each weekday, calculate mean net revenue over the last eight matching weekdays as a baseline to test. The mean retains large observations, while a median can be a useful robust comparison when unusual events dominate; neither is automatically best. Log the forecast before the outcome and compare its error with a seasonal-naive forecast and the manager's estimate. This creates a measured performance record rather than a promised accuracy band.
Four adjustments earn their keep because they're visible a week out, which is the only lead time a bar has:
- Calendar: bank holidays, school holidays, the local gig or football at 19:45, the street festival that will either bring crowds or close roads. A five-minute check against a calendar with these marked beats any algorithm.
- Weather, at the threshold: not "sunny" but the specific weather your room responds to. A venue can have a temperature breakpoint where trade shifts outside (up for terraces, down for basements — know yours), and rain on a Friday is a different event than rain on a Tuesday.
- Your own events: a takeover may move demand relative to the baseline; estimate the effect and record the assumption, but the nights either side often undershoot it. Forecast the whole week, not the event night.
- The trend drift: rising eight-week Friday means suggest a possible trend, but prices, events and availability can also explain it. Compare a trend adjustment out of sample at the daily or venue level relevant to the decision.
What the forecast is actually for
Three decisions, nothing else, and each wants a different error tolerance:
The rota wants the shape more than the level, which is why the hourly overlay method and a ±20% forecast coexist fine: you're placing two shifts, not counting pounds.
The order wants the level plus a buffer policy: choose a product-level buffer using demand variability, lead time, shelf life and the cost of a stockout; a one-standard-deviation allowance is only one scenario, not a universal service policy, and let the stock cycle catch what the buffer misses. Dead stock is a forecast error with a shelf; the buffer price is a fortnight of extra case beer instead of a lost Friday.
The prep wants yesterday's number more than next week's: batched juice, cordials and garnish prep sized to the next three days' baselines, and a rule about the batch that didn't sell that nobody wrote down and everyone breaks.
The forecasting graveyard
Things that look like signal and aren't: last Saturday (n=1, the venue's noisiest day), competitor closures read as demand shifts (they're, until they're not), social-media buzz as a weekly driver, and "the model said" arguments in venues where the model is one Excel sheet and the manager's hunch is three years of standing there. The hunch is a forecast too, an unmeasured one; the eight-week mean's advantage is that it keeps score, and after a quarter of logged forecasts-versus-actual, a venue can honestly say which forecaster it has, spreadsheet or human, per weekday. That sentence alone justifies the five minutes a week.
Where the process goes next depends on the input quality and the decision. Better Bartender's weekly review workspace helps connect recorded sales, contribution and the next action; the venue pilot explores permitted reporting access separately. Forecast automation should be evaluated against a documented baseline before claiming that it improves ordering or staffing. Keep events and local knowledge as explicit inputs, not unexplained adjustments made after seeing the outcome.
Score the forecast before adding complexity
Keep a forecast log with issue date, trading period, prediction, actual outcome and a short note for exceptional events. Compare an eight-week same-weekday mean with a seasonal-naive baseline and a manager estimate on the same future dates. Mean absolute error in pounds is easy to interpret; a percentage metric needs a rule for zero or tiny actual sales. For ordering, add on-hand stock, open purchase orders, delivery lead time, pack sizes and product shelf life. For staffing, compare hourly shape as well as daily totals. A weekly revenue forecast alone cannot specify how many limes to prep or which hour needs another bartender.
FAQ
How do bars forecast demand? Same-weekday means over 6-8 weeks as the baseline, plus visible-a-week-out adjustments (calendar, weather thresholds, own events, trend). Measure the error honestly; use it for rota, ordering and prep, and hold the residual with a buffer policy.
How accurate should a bar forecast be? Measure accuracy at your venue on held-out periods; this guide does not establish a universal ±15–25% error range. Weekly accuracy supports a budget, but hourly staffing and item-level orders need forecasts at those finer grains. Check accuracy separately at each decision level; a good weekly total can conceal a badly timed peak.
Does weather actually affect bar sales? Yes, at venue-specific breakpoints: terrace venues gain at the first genuinely warm evenings and lose to extremes, basement venues invert, and rain hits arrival-heavy evening trade harder than destination trade. Log temperature against covers for one season and your breakpoint stops being folklore.
Can AI forecast my bar's sales? As an amplifier of the above, yes: compare a model, a seasonal-naive forecast and a manager forecast on the same held-out periods before concluding which works best at your venue, so the practical 2026 setup is computed baselines plus human overlays. The venues winning are those logging forecasts versus actuals, regardless of who's forecasting.
Method and assumptions
Error-band ranges and adjustment priorities are practitioner estimates labelled indicative, not published studies; the eight-week same-weekday baseline is a seasonal-average forecasting method adapted to weekly bar seasonality. Worked figures illustrative. Method: data standards.
Last reviewed 19 October 2026.