Baikely
TL;DR
- Forecast per product, per weekday: a bakery's week has a shape, and every product has its own.
- Same-weekday averages beat fancy models for most small bakeries, as long as you fix two things: sell-outs and special days.
- Sell-outs hide demand: a product that ran out at 1 p.m. sold less than customers wanted. Estimate the gap.
- Special days need last year, not last week: a holiday is compared with the same holiday a year earlier.
How do you forecast sales for a bakery?
For each product, take its sales on the same weekday over the last few weeks, correct the days it sold out, and replace the forecast for holidays and events with the same day last year. That method is simple enough to run in a spreadsheet and accurate enough to cut both waste and sell-outs.
Why same-weekday averages work
Bakery demand has a strong weekly rhythm. Illustrative croissant sales:
| Mon | Tue | Wed | Thu | Fri | Sat | Sun |
|---|---|---|---|---|---|---|
| 92 | 88 | 95 | 101 | 128 | 176 | 150 |
A forecast that ignores the weekday is wrong almost every day. A forecast that averages the last three Saturdays to predict next Saturday captures most of the pattern with no modeling at all.
How many weeks? Three to four is a good balance: enough to smooth out one odd day, short enough to follow a trend or a season change.
Problem 1: sell-outs make history too low
If croissants sold out at 1:15 p.m. last Saturday, 176 is what you had, not what customers wanted. Averaging that day in drags next Saturday's forecast down, which causes another sell-out. A loop.
The fix: for days a product sold out, estimate the missed sales. If you know what share of a normal Saturday's sales usually happens by 1:15 p.m., divide by it. Say a normal Saturday has made 70% of its croissant sales by 1:15: 176 ÷ 0.70 ≈ 250 implied demand, or 74 missed.
Be conservative. An early sell-out extrapolates from a small part of the day, so cap the estimate (for example, at half of what you sold). Overcorrecting just swaps sell-outs for leftovers.
Problem 2: special days break the pattern
A long weekend, a holiday, a local festival: last week's average knows nothing about them. For those days, forecast from the same special day last year instead, product by product. Holiday demand shifts between products, not just in volume: sandwich loaves might double while office-lunch pastries halve.
Problem 3: closed days and zero days
Leave days the shop was closed out of the average entirely. But keep days where the shop was open and a product sold zero: that is real information.
What about weather?
Weather affects some products, but the effect is specific to your shop and product, and it is noisy. Use it as a manual adjustment where your own history shows a pattern rather than a rule for every product.
From forecast to production
A forecast says what will sell. Production adds what you choose on top: a small buffer for products you never want to run out of, a smaller one for products that do not keep. Our bakery production planning guide covers that step.
How Baikely forecasts
Baikely runs this method for every product every day from your point-of-sale data: the same weekday over the last three weeks, with estimated missed sales added on sell-out days (from your shop's own sales rhythm through the day, capped), and special days planned from last year. You see the suggestion next to last week's sales, sell-outs and leftovers:
Frequently asked questions
How accurate can a bakery sales forecast be?
Accurate enough to cut leftovers and early sell-outs noticeably, not accurate to the unit. Measure it by what you care about: leftovers and sell-out times, not a statistical score.
Do I need machine learning to forecast bakery sales?
Usually not. For a small bakery, a same-weekday average with sell-out and special-day corrections captures most of what matters.
How much history do I need?
A few weeks for weekday patterns. A year for special days, which is why recording sales from day one pays off.