Clustered Bootstrap Similarity Learning Framework: Time Series Forecasting

This study introduces the Clustered Bootstrapped Similarity Learning (CBSL) framework, a novel approach for time series forecasting that integrates deep feature extraction, similarity-based clustering, and bootstrapped ensemble learning to address challenges in capturing complex temporal dependencies and improving predictive accuracy. The proposed framework combines CNN with dilated convolutions and LSTM units for feature extraction. It then performs clustering on the extracted features using fast data time warping (fastDTW) and K-Means, followed by bootstrapped ensemble learning for prediction. The resulting model, CBSL, enhances forecasting stability and generalisation, and the current study is limited to univariate time series. We evaluated CBSL on six benchmark datasets, where it demonstrated strong and competitive performance outperformed state-of-the-art models such as TCN, XGBoost, and N-BEATS, particularly in long-term forecasting horizons. CBSL achieved the lowest mean square error (MSE) and symmetric mean absolute error (sMAPE) across all prediction windows, ranging from 96 to 720 steps. Specifically, we achieved the mean MSE reduction of 18%, 22%, 41%, 40% and 80% compared to the benchmark models including TCN, XGBoost, N-BEATS, Informer and ARIMA respectively. Similarly, for the mean sMAPE, respective reductions of 13%, 16%, 22%, 25% and 53% were achieved for the three benchmark models. Notably, the framework performance gains are more prominent for longer horizons of 336 and 720 steps, where an overall decrease of 38% and 26% in MSE and sMAPE are obtained respectively. Statistical significance test, measured using the Diebold-Mariano (DM) test, confirms CBSL's superior predictive accuracy compared to the benchmarks.