Hybrid approach to time series forecasting: a comparative analysis of ARIMA, ARIMA-ANN and ARIMA-LSTM models

This paper explores several models for time series forecasting, with a specific focus on comparing the performance of traditional models and hybrid models that integrate machine learning techniques. Using the dataset of Google, Apple, Facebook, Amazon (GAFA), focusing mainly on Apple’s share price, the paper explores the limitations of traditional statistical models. To overcome these limitations, the paper proposes hybrid forecasting models that integrate the strengths of the AutoRegressive Integreated Moving Average (ARIMA) model, which is effective in capturing linear patterns, with Artificial Neural Networks (ANNs) and Recurrent Neural Networks (RNNs) – in particular Long-Short Term Memory (LSTM) – which are capable of modelling non-linear patterns. The performance of the ARIMA-ANN and ARIMA-LSTM hybrid models is compared with the pure ARIMA model, using accuracy metrics such as MAE and RMSE. The results show that hybrid models generally outperform traditional models, but there is no single universally valid hybrid model, as the characteristics of the time series under consideration must be taken into account.