Forecasting Demand

Most schedules are built from a demand forecast. The forecast is usually fine; what is done with it frequently is not.

What you are forecasting

Transactions, covers, footfall or sales, by hour or by part-day.

Which drive labour requirement through a ratio — covers per server, transactions per till — that is itself an assumption.

The ratio matters more than the forecast and gets far less attention.

What a forecast can do

Capture the weekly pattern, which is strong and stable in both sectors.

Capture seasonality and known events — a bank holiday, a local fixture, a school term.

Give a useful expectation two weeks out, which is the horizon that matters for publication.

What it cannot do

Predict a specific Tuesday to the hour with useful precision.

Anticipate weather beyond a few days, which in hospitality is a large share of the variance.

Know about the roadworks, the closure next door, the coach party.

Which means the last mile of precision is unavailable, and building a rota that depends on it produces the cuts and call-ins that everything else in this collection is about.

The granularity error

Forecasting to the hour and staffing in four-hour blocks means most of the forecast's precision is discarded anyway.

Match the forecast granularity to the scheduling granularity, and spend the effort saved on the ratio instead.

A forecast more precise than your shift structure is a false economy that produces confidence rather than accuracy.

The ratio nobody revisits

Covers per server, transactions per hour, cases per picker.

Usually set years ago, sometimes by a different menu, layout or till system.

Check it against actual outcomes: where the ratio says four people and the shift consistently runs badly with four, the ratio is wrong.

Ask the people working it. They know within ten minutes of a shift starting.

Using it honestly

Staff to the forecast plus a small buffer, not to the forecast exactly.

Publish on the buffered figure, which is what removes the need to cut later.

And measure the forecast error so you know how large the buffer should be, rather than guessing.

The measure

Forecast against actual, by site, by day part.

Systematic bias in one direction is a model problem.

Large scatter with no bias is the irreducible variance, and it tells you the buffer size directly.

Size the buffer from measured error

Rather than guessing it.

Compare forecast against actual, by site and day part, for a quarter.

The scatter is the irreducible variance, and it tells you directly how much slack removes the cuts and call-ins.

A buffer sized this way has a number attached, which is what makes it defensible against a labour-percentage argument.

Connect policy to configuration

The choices in this note can be compared with workforce optimisation software. Enable only the data required for the stated purpose and confirm who can see and change it.

Independent reference

For a thematic point of reference, see the Office for National Statistics. This widely used specialist site offers a useful second reference for the issue.