ARIMA vs. Machine Learning in Portfolio Return Forecasting: A Comparative Study Integrating GARCH-Based Volatility Estimation and Value-at-Risk Applications

Ruiheng Chen

2025

Abstract

This study aims to compare the application effects of traditional econometric models and machine learning models in portfolio return prediction and risk management, and selects Apple's daily return as sample data. First, the Augmented Dickey-Fuller test is used to confirm the data stationarity. The optimal ARIMA model is constructed under the AIC and BIC criteria, and its in-sample return is predicted. In order to further characterize the return volatility characteristics, ARCH-LM test and residual square ACF analysis are performed on the ARIMA model residuals, and then the GARCH model is established to obtain the in-sample volatility forecast. Based on this, an LSTM model with 25-order lag as input is constructed, and the model is trained using the full sample data to generate the in-sample forecast of the return. Finally, under the premise of controlling the confidence level to 95% and uniformly using the GARCH volatility forecast results, the Value at Risk (VaR) is calculated using the normal distribution assumption, and the VaR of the ARIMA, LSTM models and real return data are compared and analysed. The research results show that the LSTM model is more sensitive to the ARIMA model under extreme market volatility conditions, but both have the limitation of underestimating extreme risks, which provides a direction for the introduction of methods such as heavy-tailed distribution or extreme value theory in the future.

Paper Citation

in Harvard Style

Chen R. (2025). ARIMA vs. Machine Learning in Portfolio Return Forecasting: A Comparative Study Integrating GARCH-Based Volatility Estimation and Value-at-Risk Applications. In Proceedings of the 2nd International Conference on Innovations in Applied Mathematics, Physics, and Astronomy - Volume 1: IAMPA; ISBN 978-989-758-774-0, SciTePress, pages 419-428. DOI: 10.5220/0013827000004708

in Bibtex Style

@conference{iampa25,
author={Ruiheng Chen},
title={ARIMA vs. Machine Learning in Portfolio Return Forecasting: A Comparative Study Integrating GARCH-Based Volatility Estimation and Value-at-Risk Applications},
booktitle={Proceedings of the 2nd International Conference on Innovations in Applied Mathematics, Physics, and Astronomy - Volume 1: IAMPA},
year={2025},
pages={419-428},
publisher={SciTePress},
organization={INSTICC},
doi={10.5220/0013827000004708},
isbn={978-989-758-774-0},
}

in EndNote Style

TY - CONF

JO - Proceedings of the 2nd International Conference on Innovations in Applied Mathematics, Physics, and Astronomy - Volume 1: IAMPA
TI - ARIMA vs. Machine Learning in Portfolio Return Forecasting: A Comparative Study Integrating GARCH-Based Volatility Estimation and Value-at-Risk Applications
SN - 978-989-758-774-0
AU - Chen R.
PY - 2025
SP - 419
EP - 428
DO - 10.5220/0013827000004708
PB - SciTePress