A Data-Driven Framework for Short-Term Bus Arrival Time Prediction Integrating AVL, AFC, and Weather Data

Accurate prediction of bus arrival times is vital for improving the reliability, efficiency, and user satisfaction of urban public transportation systems. However, unpredictable service can discourage ridership, undermine system equity, and hinder progress toward sustainable urban mobility. This study presents a comprehensive and data-driven framework for predicting short-term bus arrival times by integrating Automatic Vehicle Location (AVL), Automatic Fare Collection (AFC), and weather data. Focusing on Line 144 in Tabriz, Iran, a high-demand urban corridor with 22 stations and over 11,000 daily passengers, the study utilizes more than 2.7 million operational records collected over one month. The methodology follows the CRISP-DM framework and includes data preprocessing, temporal synchronization, missing data imputation, and feature engineering based on inter-vehicle intervals and environmental context. Three predictive models were developed and compared: Artificial Neural Network (ANN), Autoregressive Integrated Moving Average (ARIMA), and a Chi-squared Automatic Interaction Detection (CHAID)-based Decision Tree. Model training used a 70/30 data split, and performance was evaluated using Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), and Relative Absolute Error (RAE). The Decision Tree model yielded the best performance, achieving an average MAE of 4.1 s and RAE of 5.53%. Variable importance analysis highlighted instantaneous speed and recent travel history as dominant predictors, while weather showed minimal short-term impact. This study presents a novel large-scale validation of a unified AVL–AFC–weather framework and provides practical insights for developing efficient transit prediction systems in emerging urban environments.