Feature Extraction for Financial Markets’ Transactions

Karimi, Afshin | 2022

15 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 55809 (19)
  4. University: Sharif University of Technology
  5. Department: Computer Engineering
  6. Advisor(s): Manzuri Shalmani, Mohammad Taghi
  7. Abstract:
  8. The use of machine learning and deep learning tools to predict the future behavior of trends in massive data requires the extraction and creation of the eigenvector for the chosen model in the problem. It should be noted that simply by increasing the number of features, it cannot be expected that the learning model will have a higher efficiency. Rather, the quality and importance of the features in the field under study should be carefully considered. Topics such as data redundancy, data correlation, the amount of information in the data, distorted data, outliers, etc. are important steps in improving the dataset and creating a feature vector for training the learning model. In the realm of modeling the behavior of financial time series, in recent years, researchers have paid much attention to introducing new features by using unsupervised learning methods on financial data. For example, the automatic labeling of changes in the financial markets into two states, the growing market and the stagnant market, using different models of deep recurrent networks, is one of the efforts made in this field. In this research, In addition to studying the previous methods for feature engineering for the problem of predicting the stock market, after creating the feature vector with the most information an effort is made to create a model for using machine learning tools in the prediction process. The stock price events of Apple, Microsoft, Amazon, Tesla, and Google, which are the top stocks in the American s&p500 stock market, after using long short-term memory neural network, gated recurrent unit neural network, linear regression, XGBoost and Bayesian optimization on The XGBoost model has been performed and the results of the research will be compared using the Mean Absolute Error, Root Mean Squared Error and Mean Absolute Percentage Error. For each of the methods, the prediction of the price trend was shown by drawing a figure. as a result the use of the Bayesian optimization technique compared to other models significantly improved the performance of the XGBoost model and reached evaluation metrics of RMSE: 0.25, MAE: 0.12, and MAPE: 0.02
  9. Keywords:
  10. Deep Learning ; Financial Market ; Time Series ; Feature Engineering ; Feature Extraction

 Digital Object List