Comparison of deep learning and traditional methods in stock index forecasting: An application in G7 and E7 countries

The purpose of this study is to predict selected stock market indices of developed (G7) and emerging (E7) countries and to comparatively analyze the performance of traditional and deep learning-based models. In this regard, prediction success was evaluated in markets with different country dynamics using ARIMA from traditional statistical methods and LSTM model from deep learning-based models. In the modeling process, the closing prices of the relevant stock market indices were used as the dependent variable, while ten technical indicators widely accepted in the literature (MA, WMA, MACD, CCI, RSI, Stochastic %K, Stochastic %D, Williams' %R, Momentum, A/D) and three macroeconomic indicators (Gold ounce price, Brent oil price, and DXY) were used as independent variables. The dataset used in the study consists of daily data for the period 2006-2022, and the relevant data were obtained from investing.com and yahoo.finance. The prediction performance of the models was evaluated using MAE, MSE, RMSE, and MAPE error metrics for each country. The analysis results generally revealed that the LSTM model produces lower prediction errors compared to the ARIMA model and exhibits higher prediction performance, particularly in high-volatility markets, due to its ability to capture long-term dependencies. However, it was observed that the ARIMA model exhibited competitive performance compared to the LSTM model in certain markets. The ARIMA model demonstrated comparable accuracy performance to the LSTM model in the MOEX and Shanghai Composite indices. This finding reveals that structural differences between countries and dataset characteristics affect prediction performance. The results highlight the importance of model selection in financial forecasting studies and emphasize the advantages of deep learning approaches in high-volatility markets.