Research on S&P 500 Forecasting During High Volatility Periods: Comparative Analysis of ARIMA and LSTM Models

Yitong Mu

doi:10.6981/FEM.202512_6(12).0012

Authors

Yitong Mu

DOI:

https://doi.org/10.6981/FEM.202512_6(12).0012

Keywords:

S&P 500; High Volatility; ARIMA; LSTM; Time Series Forecasting.

Abstract

This paper investigates the predictive performance of ARIMA and LSTM models for the S&P 500 index under different market conditions. Using daily data from 1990 to 2024, the study also downloaded VIX and computed ATR as potential volatility indicators; however, the forecasting models were primarily based on closing prices and their log returns. High-volatility days were defined as those with daily returns exceeding 3%. The results show that in the overall sample, LSTM tends to underestimate and smooth the trend, performing similarly to ARIMA; however, during high-volatility periods, ARIMA provides more robust directional forecasts, while LSTM lags behind the real market. The findings suggest that deep learning models are not always superior to classical time-series models, and linear models remain competitive in extreme market conditions.

Downloads

Download data is not yet available.

References

[1] Fan, Z. (2025, May 26). The continued shock of the U.S. debt crisis causes stock-market volatility to resurge. First Financial Daily, A05. [Newspaper; in Chinese]

[2] Bai, J. (2020). Reflections on “anti-fragility” after “black swan” events. Journal of Fujian Jiangxia University, 10(04), 63–69, 85. [in Chinese]

[3] Cui, W. (2025). The impact of stock-market volatility on the corporate bond market. China Science & Technology Investment, (08), 22–24. [in Chinese]

[4] Yang, J. (2025). Intraday cash-flow forecasting for A-share listed firms based on ARIMA: A case study of Nanjing Xinbai. Modern Marketing, (06), 40–42. [in Chinese]

[5] Shao, M., Ren, Z., Zhao, Z., et al. (2018). Application of the ARIMA model in China’s GDP forecasting. Value Engineering, 37(09), 205–207. [in Chinese]

[6] Zhang, Z., & Duan, S. (2023). Post-pandemic economic trends in China and future GDP forecasting. Contemporary Economy, (40), 10–17. [in Chinese]

[7] Wang, C., & Zhou, H. (2024). Application of the ARIMA model in forecasting demand for high-value volatile drugs. Hospital Management Forum, 41(02), 59–63, 88. [in Chinese]

[8] Yang, Q., & Wang, C. (2019). Global stock-index prediction using deep-learning LSTM neural networks. Statistical Research, 36(03), 65–77. [in Chinese]

[9] Xiong, R., Nichols, E. P., & Shen, Y. (2015). Deep learning stock volatility with Google Domestic Trends. arXiv preprint, arXiv:1512.04916.

[10] Shi, R., & Yang, Y. (2025). Assessing and early-warning systemic risk in China’s insurance industry-An analysis based on an Attention-LSTM model. Finance Theory & Practice, 46(02), 26–34. [in Chinese]

[11] Long, M. (2024). Design and implementation of an adaptive validation platform for the interpretability of stock neural networks (Master’s thesis). East China Normal University, Shanghai. [in Chinese]

[12] Ci, B., & Zhang, P. (2022). Financial time-series forecasting based on the ARIMA-LSTM model. Statistics & Decision, 38(11), 145–149. [in Chinese]

[13] Peng, P. (2023). Price-index forecasting for traditional Chinese medicine based on R-language ARIMA and LSTM models. Yangtze River Information & Communications, 36(08), 26–28. [in Chinese]

[14] A. W., & A. V. (2004). Measuring the impact of natural disasters on agricultural markets: An empirical application using intervention analysis. Applied Economics, 36(19), 2177–2186.

[15] Kulkarni, M. S., Bharathi, S. V., Perandana, A., et al. (2025). A quest for context-specific stock price prediction: A comparison between time series, machine learning, and deep learning models. SN Computer Science, 6(4), 335.