Arnon Bareket, Bazil Pârv

Abstract: This paper offers an in-depth exploration into stock index forecasting, placing a particular focus on an array of time series decomposition techniques. Our investigation involved both non-recursive and recursive decomposition methodologies. Initial trials with non-recursive methods, such as Seasonal Decomposition using moving averages (SMA) and Seasonal-Trend Decomposition using LOESS (STL), yielded varying degrees of success, with some models like SMA showing a relatively lesser performance. The superior results were achieved through recursive decomposition, especially with the Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) followed by STL, termed as the C-STL approach. Within our specified analysis period, this approach involved decomposing indices into 8 or 9 CEEMDAN sub-components, based on the specific index, with the STL method applied subsequently to each sub-component. Predictive results for these subcomponents were generated using a Support Vector Regression (SVR) model complemented with appropriate data transformations. Overall findings highlighted the recursive models, particularly the C-STL approach, as significantly outperforming non-recursive counterparts. This underscores the potential of recursive methods as a promising avenue for enhancing the precision of stock index forecasting.

Keywords: Stock Index Forecasting; Time Series Decomposition; Seasonal Decomposition; Seasonal-Trend Decomposition Using LOESS (STL); Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN); Recursive Methodology.

Date Published: December 1, 2023
DOI: 10.11159/jmids.2023.006

View Article