Time Series Forecasting Project.

Sweet Lift Taxi wanted to predict peak demand hours near the airport to strategically attract more drivers during busy times.
The goal was to build a model capable of forecasting hourly ride requests with precision.

  • CLIENT:

    Sweet Lift Taxi

  • DATE:

    May 30, 2025

  • PROJECT:

    GitHub

Data Preparation and Exploration

The data was first arranged in chronological order and checked for duplicates or missing values. Since the dataset provided ride counts in 10-minute intervals, I resampled the data by hour using a sum aggregation to reflect the total number of orders per hour. To better understand daily trends, I applied a 6-hour rolling mean, effectively dividing each day into four time segments representing different levels of activity.

A 15-day sample window was then visualized to identify recurring patterns in demand. Through seasonal decomposition, I separated the data into trend, seasonality, and residual components. The analysis revealed a steady upward trend beginning around August and clear daily peaks and low-demand hours, though long-term seasonality diminished over time.

Feature Engineering and Modeling

To enhance predictive power, I engineered several time-based features:

  • Month, day, and day of the week to capture temporal effects
  • Lag features (up to 4 previous hours) to incorporate short-term dependencies
  • Rolling mean of 4 to smooth recent fluctuations

Findings

These features were standardized using a pipeline with StandardScaler and then fed into multiple regression models. After hyperparameter tuning and cross-validation, XGBoost outperformed all other models with an RMSE of 24.61, providing precise and computationally efficient predictions for hourly ride demand.

Full time series of ride orders showing the natural flow of demand over several months. While fluctuations are frequent, a gradual increase in activity becomes noticeable as time progresses.
Full time series of ride orders showing the natural flow of demand over several months. While fluctuations are frequent, a gradual increase in activity becomes noticeable as time progresses.
Hourly aggregation of orders reveals the true operational rhythm of demand. Converting 10-minute intervals into hourly sums highlights daily peaks and valleys essential for forecasting.
Hourly aggregation of orders reveals the true operational rhythm of demand. Converting 10-minute intervals into hourly sums highlights daily peaks and valleys essential for forecasting.
Fifteen randomly selected days smoothed with a 6-hour rolling mean. The pattern shows clear high-demand and low-demand windows throughout the day, confirming the cyclic nature of taxi requests.
Fifteen randomly selected days smoothed with a 6-hour rolling mean. The pattern shows clear high-demand and low-demand windows throughout the day, confirming the cyclic nature of taxi requests.
Time series decomposed into trend, seasonality, and residual components. The upward trend becomes evident from August, while daily seasonality diminishes over longer periods, reflecting both short-term cycles and long-term growth.
Time series decomposed into trend, seasonality, and residual components. The upward trend becomes evident from August, while daily seasonality diminishes over longer periods, reflecting both short-term cycles and long-term growth.