Problem
Lyft needs forecasts of rider demand and driver supply for every geohash-6 cell, in 5-minute steps for the next hour, refreshed every minute. Dynamic pricing and driver incentives consume them. At this grain the signals are sparse, with many zeros, and show local spikes and dips of a few minutes to about 30 minutes, caused by events like concerts. Scale is 4M geohashes per minute per signal, and a trained model can go stale within days or a dozen minutes.
Example from the article: after a concert ends, demand spikes fast, stays high for 15–20 minutes, then returns to normal over the next 15–20 minutes.
Solution in brief
Choose the model by accuracy against engineering cost. Time-series models refitted every minute beat neural networks refitted hourly at horizons of 5 to 30–45 minutes, on noisy demand, and in hilly or watery regions. Neural networks win beyond 30–45 minutes or when refitted as often as the classical models.
Questions this case answers
- How to choose a forecasting model for thousands of noisy, sparse, spatially correlated series when predictions are needed within minutes and training cost is capped
- How to keep a high-frequency forecast accurate when the underlying signal changes within minutes and heavy models cannot be refitted online
- How to forecast demand and supply per small map cell every few minutes for pricing and driver incentives?
- How to choose between classical time-series models and neural networks for high-frequency spatial forecasts?
Tools and methods
- ARIMA and auto-regression — near-term forecasts, refitted every minute per geohash or per partition of geohashes
- SVD or PCA plus a time-series model — compress thousands of geohash series to a few dozen, forecast, project back
- LSTM and CNN — neural baselines for longer horizons, refitted offline several times a day
- Flink, Kafka, Kinesis, ClickHouse, Airflow — streaming features, delivery, monitoring, daily history refresh
Google or a one-time email code. No password.