Market risk is one of the most critical risks for banks and portfolio managers. According to Basel criteria, Value at Risk (VaR) calculations should be conducted at regular intervals. VaR calculations can be performed using various methods, and the approaches and variables added to the model can vary significantly. Developments in machine learning and deep learning methods have increased the diversity of VaR calculations, enabling the construction of more accurate and complex models.
In this study, a portfolio was created using the stocks of 4 major banks in BIST30 (AKBNK, GARAN, ISCTR, YKBNK) with the help of Monte Carlo simulation and Random Forest. Calculations were made for 126 periods with a 10-day interval using 5 years of daily data. Predictions were made for the last 4 periods using 3 different Value at Risk (VaR) methods (historical, parametric, and Monte Carlo). Independent variables such as VIX (fear index), USD/TL, Gold/TL, and Brent/TL were used. The suitability of the variables was tested with machine learning regularization methods, including Ridge, Lasso, and Elastic Net regression models. Random Forest was again used to measure the impact of independent variables on stocks' weights. For each VaR model, stock weight distributions were determined for the last 4 periods, and the realized VaR results were compared.
As a result of the findings, the parametric VaR method provided the best result for the first period, while the historical VaR method provided the closest result for the other three periods. When comparing the findings with the actual results, it was observed that the findings were more optimistic, and even the closest results did not come within 30% of the actual value. The reason for the difference being greater than expected could be attributed to the fact that the value of bank stocks has been below their value in the last two years and the sharp movements in the stock market in the selected last 4 periods, independent of individual stocks.
