Changepoint Detection in Multivariate Climate Time Series: Performance Assessment of Hybrid PELT+RF Against Baseline PELT

Abstract

The research studied the performance of Hybrid Pruned Exact Linear Time and Random Forest Proximity Anomaly Scores (PELT+RF) against Baseline PELT in accurately detecting change points in Climate time series Data using simulation. The research adopts a Monte Carlo simulation framework to develop, implement and evaluate a hybrid change-point detection technique that combines the Pruned Exact Linear Time (PELT) algorithm with machine learning anomaly detection method (RF). The hybrid approach Pruned Exact Linear Time + Random Forest Proximity Anomaly Scores (PELT+RF) is compared against baseline PELT using simulated multivariate climate datasets. Across small, moderate, and large sample sizes, the same directional patterns persist- temperature and humidity increase while rainfall decreases with more breaks. However, larger samples make the regime shifts more distinct and less noisy. This finding underscores that robust detection methods must perform well not only in large datasets but also in small samples, where noisy signals make breaks harder to capture. Enhanced detection algorithms are therefore vital for early warning in short observational records or regional climate data with limited length. PELT+RF showed slightly stronger precision and F1-scores in small-sample, high-change scenarios, making it the better performer in noisy and data-limited environments. On balance, PELT+RF emerged as a strong hybrid in practical, small-sample climate contexts, where data scarcity and noise are common. Its incremental improvements in precision and F1 are particularly valuable for regional climate monitoring and early-warning systems. This Study carried out a Performance Assessment on the developed, implemented and evaluated Hybrid change-point detection framework which integrates the Pruned Exact Linear Time (PELT) algorithm with machine learning anomaly detection method (Random Forest proximity Based Anomaly Scores) for robust identification of structural breaks in multivariate climate time series against Baseline PELT Algorithm of which PELT+RF emerged a strong Hybrid. For instance, at four change points with n=70n=70 , PELT+RF achieves a precision of 0.1981 compared to 0.1919 for PELT, alongside a higher F1- score (0.3302 vs. 0.3217). Relative to all other alternatives for Change Point Detection, the Hybrid PELT+RF approach balanced computational efficiency, interpretability, and robustness. It retains PELT's exact segmentation while using ML scorers to capture multivariate, nonlinear anomalies.