How to forecast a busy night

Most forecasting advice comes down to one instruction: look at the same day last year and last month, average them, adjust for the trend. That is sound, and it is also why so many forecasts miss the nights that matter most.

Why history misses the big nights

A sales history model assumes next Tuesday resembles previous Tuesdays. It is right most of the time, which is exactly the problem: it is confidently wrong on the handful of nights where being wrong is expensive. The arena two blocks away does not run on a weekly cycle. It runs on a touring schedule and a fixture list, and neither appears anywhere in your till data.

The nights your model misses are the nights with the most upside and the most downside. Miss one in the optimistic direction and you send two people home before the rush. Miss one in the other direction and you pay four people to watch an empty room.

Forecast in two parts

The fix is to stop asking one model to do two jobs. A night is a baseline plus an event effect, and the two are worked out separately.

1. The baseline, from your own data

   What this day of the week normally does at this time of year, in this weather, at this point in the month. Sales history is genuinely good at this, and nothing else can do it, because it is specific to your room. Build it from transaction counts rather than revenue if your prices have moved.

2. The event effect, from what is on nearby

   What is happening within walking distance, how many people it draws, and when they arrive and leave. This cannot come from your history at all. It comes from a calendar, and the calendar exists weeks before the night does.

3. Your capture rate, from the two together

   The link between the two. Take nights you already remember, look up what was on, and compare the crowd figure to what you actually did over baseline. Three or four points give you a ratio specific to your door, your street and your offer.

Treat every crowd figure as a range

A single number invites false confidence. Every crowd estimate should carry a range and a reason, and the reason should tell you how much weight to give it.

• Counted. Someone published a real registration number. Tight range, plan against the middle.
• Modelled from capacity. Seats in the room multiplied by how well it is selling. A wider range, and worth checking how the show is selling closer to the date.
• Typical for the category. No capacity and no sales signal, so the figure is what this kind of event usually draws. The widest range of the three, and the one to treat as a prompt to look rather than a number to order against.

Pay attention to timing, not just volume

Two events with identical crowd figures can do opposite things to your night. A matinee finishing at five gives you a long early rush and a normal evening. An evening show starting at eight empties your room for two hours and returns it at eleven. Volume tells you how many; the schedule tells you when, and when is what the rota actually needs.

Review it weekly

Once a week, put the forecast next to what happened. You are not looking for the model to be right; you are looking for the direction it is wrong in. If you are consistently over on sports and under on theatre, that is a capture rate to split by category rather than a reason to stop forecasting.

Gigwake covers the half your sales history cannot see: every event within 5 miles, 7 days ahead, ranked by the crowd it brings, with a range on each one and what that range rests on.