Akademik Veri Yönetim Sistemi
DATAMACLEA'26 Econometric Research Association 3rd International Data Analytics and Machine Learning Conference, Ankara, Türkiye, 11 - 13 Mayıs 2026, ss.85-86, (Özet Bildiri)
This study aims to forecast the daily closing prices of the Dow Jones Islamic Market Turkey Index (DJIMTR), a pivotal benchmark for Islamic financial markets, by synthesizing modern financial machine learning and deep learning techniques. The research utilizes a comprehensive dataset consisting of 3,156 daily observations spanning from 02.01.2014 to 31.03.2026.
Methodologically, the study addresses the critical "memory loss" dilemma inherent in traditional stationarity transformations by employing Fractional Differentiation (FracDiff) with a coefficient of d=0.50. While the original series was confirmed to be non-stationary through the Augmented Dickey-Fuller (ADF) test, the application of the FracDiff technique successfully achieved stationarity while simultaneously preserving the long-term memory of the series. This memory-augmented data was then fed into a Long Short-Term Memory (LSTM) neural network architecture utilizing a sixty day sliding window approach.
The experimental results indicate a significant improvement in model convergence, with the Mean Squared Error (MSE) decreasing by 83.01% during the training phase. On the test set, the model achieved a remarkably high R-Squared (R2) score, demonstrating exceptional explanatory power. Furthermore, to filter market noise and enhance signal reliability, a forecasting threshold of 1.5% was implemented, ensuring the model generates only high-conviction trading signals. The empirical findings suggest that the integration of fractional differentiation significantly enhances the predictive accuracy and economic meaningfulness of LSTM-based financial models.