مهسا توکلی

Predicting financial markets has always been one of the most challenging issues, attracting the attention of many investors and researchers. In this regard, deep learning methods have been used a lot recently. Due to the desired results, such networks are always in development and progress. One of the networks that is being implemented in various fields is capsule network. The first time the classification capsule network was introduced, it was able to attract a lot of attention with its success on MNIST data 1 . In such networks, as in the other ones, the parameters are obtained by minimizing a loss function. In this paper, we first change the classification capsule network to a regression capsule network by modifying the last layer of the network. Then we use different information measures such as Kullnack-Leibler, Lin-Wang and Triangular information measures as a loss function, and compare their results with wellknown models including Artificial Neural Network (ANN), Convolutional Network (CNN) and Long Short-Term Memory (LSTM) as well as common used loss functions such as Mean Square Error (MSE) and Mean Absolute Percentage Error (MAPE). Using appropriate accuracy metrics, it is shown that the capsule network using triangular information measure is well able to predict the price of bitcoin for the medium and long term period including 10, 90 and 180 days.

Stock market forecasting has attracted so many researchers and investors that many studies have been done in this field. These studies have led to the development of many predictive methods, the most widely used of which are machine learning-based methods. In machine learning-based methods, loss function has a key role in determining the model weights. In this study a new loss function is introduced, that has some special features, making the investing in the stock market more accurate and profitable than other popular techniques. To assess its accuracy, a two-stage experiment has been designed using data of Tehran Stock market. In the first part of the experiment, we select the most accurate algorithm among some of the well-known machine learning algorithms based on artificial neural network, ANN, support vector machine, SVM. In the second stage of the experiment, the various popular loss functions are compared with the proposed one. As a result, we introduce a new neural network using a new loss function, which is trained based on genetic algorithm. This network has been shown to be more accurate than other well-known and common networks such as long short-term memory (LSTM) for both train and test data.