Makine öğrenmesi ile Türk hisse senetleri piyasasında getiri tahmini

In this study I compare machine learning methods for predicting the stock returns of individual Turkish stocks listed in the Istanbul Stock Exchange (Borsa Istanbul). As the main machine learning model I use the Instrumented Principal Component Analysis (IPCA) and as a benchmark model I use Fama-French Factor Model. The IPCA model generates the stock-level expected returns based on observable stock-level and firm-level characteristics and latent common factors estimated within the model. Within the model stock-level characteristics determine the factor betas, namely the covariances of stock returns with the latent common factors. I estimate versions of the benchmark Fama-French models between 3 to 5 factors. The versions of the IPCA models use between 3 and 6 factors and use 10 characteristics. The sample covers all stocks in the XUTUM Index and the sample period includes forecasts between 2010 and 2022. Using a panel data of 252 firms listed in the Borsa Istanbul XUTUM Index and I analyze the comparative performance of the IPCA and Fama-French models. More specifically, I look at the in sample and out of sample performances of the models by comparing the realized and predicted series of returns for each individual stock. I find that the IPCA model significantly outperforms the Fama-French model by obtaining significantly higher out of sample R-squared levels and correlation of return forecasts and realized returns. The performance difference between Fama-French and IPCA models is more pronounced in the Turkish stock market compared to results of (Kelly, Pruitt and Su 2018) for the US stock market. Therefore, my results imply that the use of asset pricing models based on machine learning techniques may provide better results in emerging stock markets.