Comparative Analysis of Machine Learning Algorithms for Road Traffic Forecasting in Developing Countries

Abstract

Accurate short-term traffic forecasting is essential for effective urban mobility management, yet developing countries face unique challenges due to sparse, fragmented, and frequently missing data. This systematic review compares four major machine learning families: classical statistical models (ARIMA), shallow machine learning (Support Vector Regression and Random Forest), deep recurrent networks (LSTM and GRU), and spatio-temporal graph models (Graph WaveNet and STGCN) for road traffic prediction. A total of 42 peer-reviewed studies published between 2018 and 2025 were synthesized. Unlike traditional benchmarks that prioritize accuracy alone, this review evaluates models across three operational dimensions: predictive accuracy, robustness to data missingness (up to 40%), and computational efficiency (training time and memory footprint).

Results reveal a fundamental trilemma: graph models achieve the highest accuracy with a mean MAPE of approximately 6.3%, but degrade severely under missing data and require over 45 minutes per training epoch. In contrast, ARIMA is highly robust and fast, yet inaccurate with a MAPE of 11.4%. LSTM and GRU offer the most balanced compromise, with GRU achieving a MAPE of 7.9% in under 19 minutes.

This study provides a context-aware decision framework, recommending GRU with light imputation as the pragmatic choice for low-resource settings, and calls for standardized benchmarks tailored to data-sparse environments.