Comparison of Machine Learning Methods with Classical Statistical Approaches

In this graduate qualification work there is conducted a practical comparison of the efficiency of the various time series analysis models. Twenty classic ML- and DL-models were applied to the different time series for this research. Time series intended for the analysis came from the Yahoo Finance and Tinkoff invest services- in the summary there were used eighteen time series with either daily or per-minute cost of the security papers. For all of the forecasts for the different time intervals there were calculated their own quality metrics. There were used standard metrics like RMSE and MAPE, and moreover specially implemented business metrics: the investment result received from following the model forecasts and the proportion of the correctly predicted directions of the time series movements in the next moment. For the execution of the required calculations there was developed a distributed information system. This system is running in the cloud environment the “Yandex Cloud” in the leased cluster Kubernetes. The initial data is being updated in this system every minute ( ETL-processes in the Airflow are being launched), the security paper cost forecasts are being made in the next moment of time according to the calculated aggregated metrics of the quality predictions for each model. The prediction is being implemented in pods of the cluster Kubernetes, which is automatically scaled according to the current loading. The results display in the web interface made by using the framework Vue.js.