Katkat
Back

How to forecast bike dock demand

Two SigmaData Scientist · Technical Screen
EasyData ScientistTechnical ScreenMachine Learning

Last updated

This problem tests your ability to design a predictive pipeline for spatio-temporal demand, focusing on feature extraction from time-series data and robust validation strategies.

Problem Statement

You are tasked with building a model to predict the hourly demand for bike-sharing docks at a specific station in a major city. The goal is to provide a forecast for the next 24 hours to help the operations team decide when to move bikes between stations.

You have access to the following data:

  • hourly_logs: A table containing timestamp, station_id, num_bikes_available, and num_docks_available for the past 2 years.
  • weather_data: Historical hourly temperature, precipitation, and wind speed.
  • calendar_data: Indicators for weekends, holidays, and local events (concerts, games).

Requirements

  1. Define the target variable: How would you structure the supervised learning problem given that the output is a continuous integer (number of bikes needed)?
  2. Feature Engineering: Propose 5 features that capture seasonality, trend, and external influence on bike usage.
  3. Validation: Explain how you would perform cross-validation for this time-series data to avoid data leakage.
  4. Model Choice: Choose a baseline model and a more complex model, justifying your selection.

Constraints

  • The prediction must be granular (per station, per hour).
  • Avoid "look-ahead" bias where future information accidentally leaks into training samples.
  • The model must be robust to sudden spikes or drops in usage (e.g., sudden storms or station maintenance).

Hints

Hint 1

Consider transforming the raw counts into a change-in-demand variable or a delta to account for station capacity limits.

Hint 2

Use a sliding-window or 'rolling' cross-validation approach where the training set grows over time instead of standard K-fold cross-validation.

Hint 3

Think about lag features: historical demand at the same hour on the previous day or the previous week is often highly predictive.

Follow-up questions

  • How would you handle cold-start issues for a newly installed bike station that has no historical data?
  • If the model predicts demand correctly but the bikes cannot be moved fast enough due to traffic, how would you adjust the business objective?
  • How would you monitor the model performance in production to detect concept drift caused by seasonal changes or new infrastructure?