Learning and Predicting Asset Management

Abstract

Instant exchange rates offered to customers are the most critical issues in the banking industry. It is very important for both the bank and the customer that the offers given are at the appropriate level. In this study, approximately 5 months of data were used and estimation models were designed for the estimation of the currency offers given to the customers. The study was conducted over 18 different currencies. In the study, dependent variables were determined as customer segment, instant exchange rate, day information, time information and volatility value. The independent variable is the exchange rate margin. The training was carried out with daily data and using RF, GBM, ANN, DNN and CNN algorithms. Random search algorithm was used to find the hyperparameters of the algorithms and the results of the model training were compared. The models with the lowest error values were selected to be used in the estimation phase. Mean Square Error (MSE) and Mean Absolute Error (MAE) functions were used to measure performance. It has been observed that artificial neural networks and convolutional neural network algorithms reveal better results than other algorithms according to the trainings carried out on three different models. Estimation time for 18 currencies is about 3 seconds.