LSTM Networks and ARIMA Models for Financial Time Series Prediction

In this study, we evaluate the forecasting effectiveness of the classical ARIMA model and the deep learning-based LSTM model in the financial domain. An investment portfolio comprising equal shares of gold, the S&P 500, and 2-year U.S, is constructed to predict future trends, thereby accounting for market factors and risk-free interest rates. Various historical data are used to train the model and forecast the value of protfolio respectively. For the ARIMAmodel,predictions are made by segmenting the model into three groups based on the time span of the training data. The LSTM model utilizes 80% of the data as the training set and 20% as the test set. Furthermore, by employing diverse initial states for parallel training and averaging, errors are reasonably reduced. Key indicators, such as the portfolios expected annual returns, daily logarithmic returns, volatility, and value at risk (VaR), are calculated.The findings suggest that both the forecasting models and the constructed portfolio are effective. Future research could focus on using prediction models to optimize and dynamically adjust portfolios, thereby enhancing returns