On the Adequacy of Simpler Models: Multilayer Perceptrons Outperform LSTMs in Short-Term S&P 500 Index Prediction
16th International "Artemis" Scientific Research Congress, Bucuresti, Romanya, 28 - 30 Ağustos 2026, (Tam Metin Bildiri)
- Yayın Türü: Bildiri / Tam Metin Bildiri
- Basıldığı Şehir: Bucuresti
- Basıldığı Ülke: Romanya
- Erciyes Üniversitesi Adresli: Evet
Özet
While Long Short-Term Memory (LSTM) networks have become the dominant standard for stock market forecasting, their high architectural complexity may be redundant for short-term predictions relying solely on close history. This study challenges the necessity of complex recurrent architectures by proposing an optimized Multilayer Perceptron (MLP) model utilizing a restricted five-day lookback window to predict the next-day closing value of the S\&P 500 index. Evaluated within an identical and robust experimental framework over a 6-year dataset, both models underwent systematic hyperparameter tuning via Grid Search and were tested across 30 independent runs. The empirical results demonstrate that the simpler MLP architecture consistently outperforms the LSTM across all evaluated metrics, achieving a significantly lower average Mean Squared Error (MSE) of 2.32e-4 compared to 3.79e-4 for the LSTM. These findings suggest that for short-horizon financial forecasting, lower architectural complexity not only mitigates overfitting risks but also provides a more robust and computationally efficient alternative for modern trading systems.