A Comparative Assessment of Frequentist Forecasting Models: Evidence from the S&P 500 Pharmaceuticals Index

Abstract

This paper compares three forecasting methods, the autoregressive integrated moving average (ARIMA), generalized autoregressive conditional heteroscedasticity (GARCH), and neural network autoregression (NNAR) methods, using the S&P 500 Pharmaceuticals Index. The objective is to identify the most accurate model based on the mean average forecasting error (MAFE). The results consistently show the NNAR model to outperform ARIMA and GARCH and to exhibit a significantly lower MAFE. The existing literature presents conflicting findings on forecasting model accuracy for stock indexes. While studies have explored various models, no universally applicable model exists. Therefore, a comparative analysis is crucial. The methodology includes data collection and cleaning, exploratory analysis, and model building. The daily closing prices of pharmaceutical stocks from the S&P 500 serve as the dataset. The exploratory analysis reveals an upward trend and increasing heteroscedasticity in the pharmaceuticals index, with the unit root tests confirming non-stationarity. To address this, the dataset has been transformed into stationary returns using logarithmic and differencing techniques. Model building involves splitting the dataset into training and test sets. The training set determines the best-fit models for each method. The models are then compared using MAFE on the test set, with the model possessing the lowest MAFE being considered the best. The findings provide insights into model accuracy for pharmaceutical industry indexes, aiding investor predictions, with the comparative analysis emphasizing tailored forecasting models for specific indexes and datasets.