Real-Time Demand Forecasting
Geospatial Demand Prediction at 15-Minute Resolution
Captured from the live deployment, not a mockup
01 · The problem
The dataset spans exactly two days, so the real problem is not a bigger model but a historical-prior feature that cannot leak the label it predicts. Decoded geohashes to coordinates and built a per-location per-time-slot demand profile from the one fully labelled day.
02 · How it works
- 01
Forecast demand in 15-minute intervals across 1,249 geohashes from 77,299 rows, combining spatial indexing with cyclical encoding of time-of-day and day-of-week.
- 02
Chose regularisation over capacity because the leak-free training set is only 7,872 rows, validated with 5-fold cross-validation and published the metrics file rather than an unverifiable leaderboard number.
- 03
Forecast demand across 1,249 geohashes at 15-minute resolution, engineering a historical-prior feature that cannot leak its own label from a dataset spanning only two days.
03 · What it cost, and what it returned
Validated with 5-fold cross-validation and published a reproducible metrics file instead of a leaderboard screenshot, tuning the ensemble with Optuna.