مجلة الدراسات الإقتصادية والمالية

A Comparative Study Of Traditional Statistical Models And Deep Learning In Stock Price Prediction: Apple Inc. As A Case Study (arima) Vs( Lstm)

Auteurs : Bouzid Sarah .

Résumé

Abstract: This study aims to evaluate the effectiveness of two contrasting models in predicting the daily closing prices of Nvidia stock. The Box-Genghis methodology was employed, specifically the ARIMA model as a representative of classical statistical methods, and the LSTM (Long Short-Term Memory) model was used as a representative of artificial intelligence and deep learning techniques. The models were trained on historical data spanning 2516 trading days, from January 01/01/ 2014 to December 31/12/2024, to test the validity of the study’s hypotheses, which focused on the predictive power of the time series, the extent to which AI outperforms traditional models, and the stability of each, in addition to testing the extent to which neural networks improve performance through learning. The ARIMA model relied on step-by-step forecasting, while a deep neural network was built for the LSTM model using 20 training epochs. The study found that there is a strong statistical relationship between historical data and predictive ability using both models, and that the ARIMA model demonstrated superior performance with remarkable accuracy, recording an average absolute error (MAE) of 0.68 compared to 1.44 for the LSTM model. This superiority confirms ARIMA’s high capability, demonstrating that traditional statistical models are capable of tracking stable trends in the very short term, while the LSTM showed a deviation resulting from the need for finer parameter tuning and larger datasets, suggesting that this shortcoming is not due to the model’s limitations but rather to the narrow scope of training.

Mots clés

Forecasting and forecasting models ; Neural Networks ; The ARIMA Model ; The LSTM Model.