Forecasting · Hospitality operations
We made a forecast less clever and far more useful.
The frustration
An early forecasting engine hit around 11% MAPE. Engineering leaders were close to awe. A regional manager and a store user were completely confused.
The pattern
Mistaking model quality for product quality. A forecast can be mathematically respectable and operationally useless at the same time.
The solution
A rebuild from a plain piece of paper, putting the user's journey at the front and layering the engineering underneath, aiming for around 85 to 88% accuracy people could understand and act on.
The outcome
The best compliment it can earn is “Yeah, that's about right.” It resembles the world the manager is standing in, so they trust it enough to change a rota.
The lesson
“Never let the sophistication of what you built matter more than the person who has to use it.”
Read the full story in the bookBy the numbers
Around 11% MAPE
Original model error
85 to 88%
Rebuilt target accuracy
Paper
Starting point of the rebuild