Causal Forecasting at Lyft (Part 1)

DJ Rich, Sameer Manek28 Jun 2022

Problem

Lyft runs separate teams for driver bonuses, rider coupons, pricing and driver activation. Each team sets inputs such as a weekly spending budget, and the company needs to choose them well and to see their effect on rides, sessions and revenue. Past decisions heavily confound the data, so a model fitted only to correlations gives wrong answers about what happens when a budget changes. It cannot support planning.

Example from the article: on synthetic data, the driver incentive budget fell over the year while driver hours rose. A correlational model learns that a higher budget means fewer driver hours.

Solution in brief

Write the business as a directed acyclic graph (DAG) in which every variable is predicted from the variables that cause it, and let each team fit its own node. Where an experiment measured how a decision changes an outcome, force the model to follow that curve, and fit the remaining variation to history. The joined graph forecasts metrics for any plan and can be optimized over the policy variables.

Questions this case answers
  • How to forecast outcomes under alternative planning decisions when past decisions confound the history and a purely correlational model gives the wrong direction of effect
  • How to combine independently built causal models of linked business metrics into one model that supports scenario planning and optimization
  • How to forecast business metrics under different budget decisions when historical data is confounded by past decisions?
  • How to make a forecasting model respect the effect of a budget that an experiment measured?

Tools and methods

  • Causal DAG — states which variables cause which; small graphs from different teams join into one
  • Experimentally measured cost curve — fixes how a policy variable changes its outcome
  • Residual model — a smoothly evolving moving average that fits what the cost curve leaves unexplained
  • Constrained optimization of the plan — maximizes a function of forecast metrics over policy variables
Continue reading — it's free

Google or a one-time email code. No password.