Time-Frequency Representation and Filtering of Time Series of Economic Variables as a Method to Improve Their Forecasting

Abstract In an era of dynamic financial markets and digital transformation, the ability to accurately forecast economic variables is paramount for investors, financial analysts, and decision-makers. This research delves into the critical task of improving forecasting accuracy for economic variables, particularly stock prices, through advanced time series filtering algorithms. By enhancing predictive capabilities, the study aims to provide valuable insights for making informed investment decisions and managing financial risks effectively. The methods utilized, including Empirical Mode Decomposition, Fast Fourier Transform, and Wavelet Transform for data filtration, are instrumental in deciphering market trends and optimizing forecasting models. The evaluation of various filtering algorithms alongside predictive models, such as SARIMAX from traditional econometrics, XGBoost from machine learning, and Temporal Fusion Transformer from deep learning will shed light on the most effective strategies for enhancing predictive accuracy in financial time series analysis. Anticipated outcomes suggest that filtering noisy data, like stock prices, is likely to positively impact the predictive capabilities of the models. The research aims to establish which filtering algorithms work best for each model, ultimately contributing to the refinement of forecasting techniques in financial timeseries analysis. Keywords: stock prices forecasting, time series filtering, Empirical Mode Decomposition, Fast Fourier Transform, Wavelet Transform, Temporal Fusion Transformer, XGBoost.