Making long-term forecasts at Lyft
At Lyft, like many other companies, we need to make accurate short and long-term forecasts. Some of the metrics that we need to accurately predict are number of driver hours provided by drivers in different regions — i.e our supply side of the business — and also number of rides taken by riders in different regions, i.e. our demand. We have several internal tools that we use to make forecasts.
One of the tools that we use to make long-term forecasts of up to 52 weeks is a cohort based model. In this blog, we are going to briefly explain how our cohort based model works. We call specific users who completed their first ride or provide their first driving hour in a specific region and in a specific week, a cohort.
For instance, all the passengers who used their Lyft app for the first time in the 2nd week of 2018 are considered in a cohort. It is not hard to assume that there are differences among different cohorts. As an example, the sub-population of a college town who start using the app in September are different from an East Coast city where a subset of the population start using the app in a specific week of a cold winter.