Forecasting financial markets remains challenging due to nonlinearity and volatility. This study evaluates the performance of two recurrent neural network architectures—Long Short-Term Memory (LSTM) and Gated Recurrent Units (GRU)—for predicting five financial assets (Bitcoin, crude oil, gold, Google stock, and the S&P 500) using daily data from 2021 to 2023. Models are trained using Adam and RMSprop optimizers and assessed under Mean Squared Error (MSE) and Mean Absolute Percentage Error (MAPE) loss functions. The results show that predictive performance varies across assets and evaluation metrics. GRU with Adam performs well for assets with moderate volatility, while GRU with RMSprop and MSE is more effective for smoother price dynamics. In contrast, all models face difficulties in capturing extreme fluctuations in highly volatile markets such as Bitcoin. The findings indicate that the effectiveness of RNN-based models is conditional on asset characteristics, volatility regimes, and training configurations, highlighting the importance of model selection tailored to specific market conditions.
Moghadam et al. (Fri,) studied this question.