Comparing ARIMA, Neural Networks, and ARIMA-GARCH for NVIDIA Forecasts

Summary

This study compares ARIMA, a multilayer perceptron, an LSTM, and an ARIMA-GARCH combination for next-day NVIDIA share-price forecasts. It describes ARIMA as a model for linear time-series structure, neural networks as ways to capture nonlinear patterns, and GARCH as a method for representing changing volatility alongside ARIMA’s mean dynamics. The data came from Yahoo Finance and covered April 2019 through April 2024.

The evaluation used MAE, RMSE, and R-squared. The reported results are mixed: expanding-window ARIMA performed slightly better on MAE, RMSE, and R-squared in one comparison, while the summary identifies ARIMA-GARCH as strongest for RMSE. The article also notes that ARIMA-GARCH had higher MAE, associated with overprediction near the test period’s end. It provides model configurations and windowing approaches, but the presented material omits the metric values, detailed data split, and full equations. The single-stock, historical comparison does not establish that any model will generalize to other assets or support profitable trading.

Key ideas

• The study compares ARIMA and ARIMA-GARCH statistical models with MLP and LSTM neural networks for next-day NVIDIA price forecasts.
• ARIMA models time-series mean behavior, while GARCH adds a model of changing conditional volatility.
• ARIMA used expanding and rolling windows, with information criteria used for parameter selection in the expanding-window approach.
• The reported performance ranking is inconsistent: expanding-window ARIMA is described as slightly better across several metrics, while ARIMA-GARCH is highlighted for RMSE.
• The article reports higher ARIMA-GARCH MAE and a tendency to overpredict near the test period’s end.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.