Comparative evaluation of forecasting models ets, arima, and lstm with preprocessing

application to Ibovespa time series

Authors

  • Nathalia Najla Costa Borges Aeronautics Institute of Technology (ITA)
  • Mauri Aparecido de Oliveira Instituto Tecnológico de Aeronáutica (ITA)

DOI:

https://doi.org/10.36942/reni.v11i2.1638

Keywords:

Time series, Financial forecasting, LSTM, ARIMA, Savitzky–Golay

Abstract

This study investigates the effects of smoothing techniques on the performance of statistical and deep learning models applied to financial time series forecasting. The six largest stocks by weight in the IBOVESPA index were analyzed from January 2023 to October 2025. ARIMA, ETS, and Long Short-Term Memory (LSTM) models were evaluated under three preprocessing conditions: original series, Savitzky–Golay filtering, and Moving Average smoothing. Performance was assessed using RMSE, MAE, MAPE, sMAPE, and the directional accuracy metric (dStat). Results showed that the LSTM model combined with Moving Average preprocessing achieved the best overall performance, obtaining the lowest forecasting errors for all analyzed stocks and reducing sMAPE by approximately 38% compared with statistical models. In contrast, ARIMA and ETS demonstrated greater robustness when applied to non-preprocessed data. The findings suggest that the benefits of smoothing techniques depend on the forecasting paradigm and are particularly relevant for deep learning approaches.

Downloads

Download data is not yet available.

Published

2026-10-05

How to Cite

Costa Borges, N. N., & de Oliveira, M. A. (2026). Comparative evaluation of forecasting models ets, arima, and lstm with preprocessing: application to Ibovespa time series. Journal of Entrepreneurship, Business and Innovation, 11(2), 29–46. https://doi.org/10.36942/reni.v11i2.1638

Issue

Section

Artigos