مجلة العلوم الإدارية والمالية

Forecasting Coca-cola Stock Prices: A Comparative Evaluation Of Arima And Machine Learning Models

Auteurs : Rouaba Mohammed .

Résumé

This study evaluates and compares the out-of-sample forecasting performance of ARIMA, Support Vector Regression with a radial basis function kernel (SVR-RBF), Random Forest, and XGBoost models for Coca-Cola (KO) daily closing stock prices. The empirical analysis uses 1,155 daily observations covering the period from January 3, 2022, to August 11, 2026. The dataset is divided chronologically into 80% training and 20% final testing samples, with Time Series Split validation applied exclusively within the training period. ADF and KPSS tests indicate that the closing-price series is non-stationary in levels and becomes stationary after first differencing, supporting an ARIMA specification with (d=1). The machine-learning models forecast next-day logarithmic returns using lagged prices, returns, technical indicators, volatility measures, and trading-volume variables. Forecasting performance is assessed using MAE, RMSE, MAPE, SMAPE, R², and directional accuracy, with the Diebold-Mariano test used for statistical comparison. The results show that ARIMA achieves the lowest MAE (0.7196), RMSE (0.9815), MAPE (0.9245%), and SMAPE (0.9269%), as well as the highest R² (0.9624). Overall, ARIMA outperforms the three machine-learning models in price-level forecasting.

Mots clés

Coca-Cola ; Stock Price Forecasting ; Time Series Forecasting ; ARIMA ; SVR-RBF ; Random Forest ; XGBoost