A forecast that turned into a product

Sugared + Bronzed runs sugaring and spray tan studios in cities across the country. When Colby and I started working with them through PRIORA, they needed a forecast they could actually plan around. Staffing, budgets and leases all hinged on how many services each studio would do in a given month, and what they had to go on was a spreadsheet and third-party site reports they were paying tens of thousands of dollars a year for.

The problem

A forecast for a business like this is mostly a forecast of operating conditions. A studio that was short-staffed in March served fewer clients than wanted to come in, a promotion pulls demand forward from the following weeks, and the weather has a say as well. A model trained on what actually happened, without accounting for any of that, learns the staffing gaps and the promo calendar as if they were demand.

What we built

One model across every location, trained on each studio’s daily service volume since it opened, with staffing, promotions and weather as inputs. To get at underlying demand, it re-predicts history as if every studio had been well staffed with no promotions running, and planned promotions and known openings are then layered back on top. Daily predictions roll up to any month or quarter, which sounds trivial until you remember that some Februaries have five Thursdays and Thursday is the busiest day of the week.

Alongside the forecast, we brought the CFO’s Excel operating model into the warehouse and reconciled the two until they matched exactly, so finance and operations were finally working from the same numbers.

What changed

Any model can be made to fit the past, so how well a forecast matches history tells you very little. The honest test is to freeze it and wait. The forecast we froze on January 1 landed within about 1.4% of the company’s total for the year, and 42 of 48 locations came in within 10% of their own number. It was about as accurate twelve months out as it was one month out, and it replaced the benchmarking tool they had been paying for.

Where it went next

To help them decide where to open next, we built a map that scored every neighborhood in a metro on how many people looked like their best customers, with their own studios and their competitors on top. It worked better than the site reports they had been buying, so we turned it into Hexmark.