🚕 Taxi Demand Forecasting

📌 Project Overview

"Chetenkoe Taxi" company collects historical data on taxi orders at airports. To attract more drivers during peak hours, it's essential to forecast the number of taxi orders for the next hour. This project develops a time series forecasting model to predict taxi demand with high accuracy.

🎯 Objective

Build a model that accurately predicts hourly taxi order demand with:

• RMSE ≤ 48 on the test set
• Ability to forecast demand for multiple future hours
• Timely identification of peak demand periods to optimize driver allocation

📊 Dataset Description

• Source: taxi.csv - historical taxi order data
• Time period: March 1, 2018 to August 31, 2018 (spring-summer season)
• Original frequency: Every 10 minutes
• Resampled frequency: Hourly (after processing)
• Key feature:
  • num_orders - number of taxi orders

The dataset contains 26,496 entries with no missing values or anomalies. The time series shows:

• Mild upward trend
• Daily seasonality patterns
• No yearly seasonality (limited to spring-summer period)

🔍 Methodology

Data Preparation

• Resampled original 10-minute data to hourly frequency
• Conducted stationarity analysis using Augmented Dickey-Fuller test (confirmed stationarity)
• Created time-based features:
  • Hour of day
  • Day of week
  • Weekend indicator
  • Lag features (up to 248 hours)
  • Rolling mean (window sizes: 19, 24, 29 hours)

Modeling Approach

Implemented custom time series forecasting framework with:

• TimeSeriesSplit cross-validation
• Feature engineering pipeline
• Two candidate models with hyperparameter optimization:
  • Ridge Regression (L2-regularized linear model)
  • LGBMRegressor (LightGBM gradient boosting)

Model Selection

• Tested multiple combinations of lag features and window sizes
• Evaluated models using RMSE metric
• Selected best model based on cross-validation performance

📈 Results

Best Model: Ridge Regression with optimal parameters

Performance Metrics:

• Cross-validation RMSE: 25
• Test set RMSE: 33.79 (well below the 48 threshold)
• Successfully captures daily demand patterns and peak hours

Practical Application:

• Model can forecast demand for multiple future hours
• Example prediction for September 1, 2018:
  • First hour: ~279 orders
  • Full day forecast shows clear daily pattern with morning and evening peaks

The solution enables "Chetenkoe Taxi" to proactively allocate drivers based on predicted demand, improving service quality during peak hours while optimizing operational costs.