CEEMD-based Multivariate Financial Time Series Forecasting using a Temporal Fusion Transformer

Time series analysis, a crucial tool for understanding trends and patterns over time, has been applied in diverse domains, such as finance, healthcare, manufacturing, telecommunications, economics, and energy. Among the various applications, time series forecasting is of significant interest. This paper presents a machine learning framework for time series forecasting employing a complete ensemble empirical mode decomposition (CEEMD) -based method in conjunction with a temporal fusion transformer (TFT) model. This study aimed to investigate a machine learning technique that utilizes multivariate features derived from CEEMD-decomposed time series data, leveraging a TFT predictor for multi-horizon forecasting. The application of this framework is demonstrated using a financial time series dataset, specifically the 14-day horizon forecast of the S&P 500 index. The results showcase the effectiveness of the proposed approach, revealing a substantial improvement (up to 33.72^%) compared to similar TFT models utilizing closing prices, classical EMD, and EEMD approaches. This machine learning framework has the potential to provide insights into forecasting financial data and its broader applicability across diverse fields of time series forecasting.