The paper considers the problem of increasing the accuracy of forecasting dynamically changing data using cryptocurrency rates as an example. The aim of the paper is to research and develop machine learning methods used in applied problems for forecasting dynamically changing data. The relevance of the work is due to the fact that existing methods and models do not always allow achieving sufficient accuracy and reliability in forecasting quotes in conditions of high volatility and complexity of cryptocurrency markets. Thus, there is a need for further in-depth research and development of new approaches and methods to solve this problem. The object of the study is financial quotes of cryptocurrencies. The subject of the study is the performance indicators of machine learning methods used to forecast quotes. To assess the quality of model predictions, the Mean Squared Error (MSE) metric was chosen, which helps measure the accuracy of the model by identifying forecast errors. Additional quality indicators were also proposed, such as the correct prediction of maximum and minimum points, which is important for analyzing cryptocurrency price fluctuations. Forecasting models were created using the TensorFlow library and the T4 graphics accelerator. The Adam algorithm was used for optimization, training was performed using the mini-packet technique. The following research methods served as the methodological basis for the work: comparison, description, measurement, scientific abstraction method, as well as analysis and generalization. The conclusion provides the main findings obtained as a result of the study.
Leokhin et al. (Wed,) studied this question.