Data-Driven Weather Forecasting in Singapore Using Markov Chains and Naïve Bayes

Abstract

Singapore’s tropical climate poses challenges for weather forecasting. This paper proposes a hybrid model combining first-order Markov Chains with Gaussian Naïve Bayes. Markov Chains capture temporal dependencies through transition probabilities, serving as dynamic priors for Naïve Bayes classification incorporating temperature, humidity, dew point, visibility, wind direction, cloud cover, and a monsoon phase index. Using daily data from Singapore (2020-2025) with an 80/20 split, the hybrid model achieved 84.3% (two-state) and 81.2% (four-state) accuracy, outperforming the standalone Markov Chain $(79.8 \% / 74.6 \%)$. Comparison against Logistic Regression, Random Forest, and XGBoost shows competitive performance with lower computational cost. Per-class evaluation using precision, recall, and F1-score reveals limitations for rare events. The approach provides a lightweight, interpretable framework for tropical weather forecasting.