Akademik Veri Yönetim Sistemi
Gazi University Journal of Science, cilt.39, sa.1, ss.499-522, 2026 (ESCI, Scopus, TRDizin)
This paper re-evaluates the forecasting performance of Long Short-Term Memory (LSTM) models against traditional Autoregressive Integrated Moving Average (ARIMA) and Random Walk (RW) models using S&P 500 index data. The LSTM models in the literature, which utilize observed lagged values for their forecasting (static forecasting), are compared to traditional models that use recursive, self-generated predictions for forecasting (dynamic forecasting). To provide comparable statistics, we gather relevant statistics for the static forecasting of ARIMA and RW models. We repeat the exercise for three different cross-validation schemes. The empirical evidence presented here suggests that the static forecasting power of static ARIMA and RW models consistently matches or outperforms that of LSTMs. Moreover, we find that while traditional models remain robust across varying sample sizes, the performance of LSTMs decreases when the training datasets are reduced. The empirical evidence presented here suggests that the reported superiority of LSTM in financial time series forecasting might be due to the forecasting method that is employed rather than a genuine predictive advantage.