Forecasting

Forecast daily demand per shop.

One model per product and per shop, learning from that shop's own weather, its own public holidays and the events on its own street. We publish no accuracy figure: the four-week diagnostic measures it per product and per shop, on your own history.

See forecast in demoSee pricing →

How does BakeWise forecast a single shop's day?

It fits one model per product and per shop. Prophet carries the calendar and the season; an XGBoost model learns the residual it leaves behind, from that shop's own weather, its own public holidays and the events on its own street. The four-week diagnostic measures the error on your own history.

Calculate. Alert. Propose. Doesn't decide. Every prediction is traceable — you'll see why the model lowered rustic bread on Tuesday before approving it.

What does this cost today?

  1. Monday morning production meetingsProduction is estimated every morning on a spreadsheet only one person understands. Each 30-minute meeting multiplies by 7 shops.
  2. 8-15% unmeasured wasteEach bakery bakes "almost the same". That difference costs you thousands every month and nobody in your organization is quantifying it.
  3. Lost sales from stockoutsPastries sell out Saturday noon. The customer goes to the competition. You have no data to know how many.

Before / After

And what changes on an ordinary Monday?

María's Monday — before and with BakeWise.

TodayWith BakeWise
Sunday 9pm: I WhatsApp 7 shop managers.Friday 5pm: weekly plan generated and approved.
I tally by hand in Excel.Automatic export to each shop.
I cross fingers with Tuesday's weather.Proactive alert if Monday's weather forecast changes.

Which numbers can you check?

  • 90 days

    blind backtest window

  • 60+

    Model variables

  • 24h

    Retraining cadence

  • 180 days

    Forecast horizon

What is underneath?

  1. Prophet + an XGBoost residual modelTrend, seasonality, holidays, weather, points of interest and local events. Sixty-plus declared signals; 15–40 are kept per model, chosen by that product's sales velocity. Nightly retrain.
  2. DynamicRulesEngineLearns weather impact from your data. No hardcoded percentages.
  3. Scenario plannerWhat-if: "What if it rains Saturday?" — recalculates in seconds. Chain+.
  4. Online learning + driftModel adjusts with each sale. Alerts if distribution shifts.
  5. One model per shop, not one model splitEach shop gets its own trained model per product, fitted on that shop's history, its own weather station and the points of interest around it. The city-centre shop and the industrial-estate shop stop sharing a curve.

What else gets asked before deciding?

  • What if I don't have 6 months of data?

    We start with a baseline (moving average + generic seasonality) and improve weekly. Accuracy climbs as history accumulates; the four-week diagnostic measures where you actually are, per product and per shop, rather than us quoting a figure.

  • How does it learn from weather?

    DynamicRulesEngine correlates each AEMET condition with your historical sales, product by product. Calibrated per shop. We never use hardcoded percentages.

  • Can I manually adjust the forecast?

    Yes. Each plan is editable before sending to the shop. The adjustment feeds back into the model as feedback for next week.

  • How often are models retrained?

    Nightly. Drift monitoring detects distribution shifts in ≤48h and triggers extra retrains if needed.

  • Are forecasts per shop or aggregated?

    Per shop per product. In Group we additionally aggregate to negotiate with suppliers using network-wide data.

See forecast in demo

See forecast in demo