Abstract:

With the recent development in technology, the use of deep learning methods in many fields has gained momentum. One of the most popular of these areas is financial markets. The estimations and analyzes performed on financial data, where many inputs are effective, have great importance on the decision-making mechanisms of investors and institutional structures. In the study, Borsa Istanbul 100 (BIST100) index is tried to be estimated with a deep learning network. In addition, the effect of the use of statistical dimension reduction methods such as Factor Analysis (FA), Principal Component Analysis (PCA), Independent Component Analysis (ICA) in the data preprocessing stage on Long Short Term Memory (LSTM) deep learning network performance is investigated. The data set used during the experiments; is prepared based on daily historical data and technical analysis information of the BIST100 index. In the data preprocessing stage, the models created from the statistical dimension reduction methods added to the deep learning network were compared over the R2 and RMSE criteria while trying to predict the price at the end of 5-days. During these processes, parameters other than deep learning hyper-parameters, which are thought to affect the performance of technical indicators and forecasting models, were tried to be improved. Accordingly, the PCA+LSTM hybrid model outperformed the hybrid models created by the other, dimensional reduction methods and achieved more competitive results. At the same time, it was observed that the PCA+LSTM hybrid model improved the results of the LSTM model alone by 4.60% and 13.35% for R2 and RMSE, respectively.

Keywords: