Back to Search
Start Over
Methods for forecasting financial time series
- Publication Year :
- 2022
- Publisher :
- СанкÑ-ÐеÑеÑбÑÑгÑкий полиÑÐµÑ Ð½Ð¸ÑеÑкий ÑнивеÑÑиÑÐµÑ ÐеÑÑа Ðеликого, 2022.
-
Abstract
- Тема вÑпÑÑкной квалиÑикаÑионной ÑабоÑÑ: «ÐеÑÐ¾Ð´Ñ Ð¿ÑогнозиÑÐ¾Ð²Ð°Ð½Ð¸Ñ ÑинанÑовÑÑ Ð²ÑеменнÑÑ ÑÑдов». Рданной ÑабоÑе ÑеÑаеÑÑÑ Ð¿ÑÐ¸ÐºÐ»Ð°Ð´Ð½Ð°Ñ Ð·Ð°Ð´Ð°Ñа пÑогнозиÑÐ¾Ð²Ð°Ð½Ð¸Ñ ÑинанÑового вÑеменного ÑÑда ÑиÑÑовой валÑÑÑ. РаÑÑмаÑÑиваÑÑÑÑ ÐºÐ»Ð°ÑÑиÑеÑÐºÐ°Ñ Ð¼Ð¾Ð´ÐµÐ»Ñ Ð°Ð½Ð°Ð»Ð¸Ð·Ð° и пÑогнозиÑÐ¾Ð²Ð°Ð½Ð¸Ñ ARIMA и Ð¼Ð¾Ð´ÐµÐ»Ñ Ð³Ð»Ñбокого обÑÑÐµÐ½Ð¸Ñ LSTM аÑÑ Ð¸ÑекÑÑÑÑ ÑекÑÑÑенÑной нейÑонной ÑеÑи. ÐадаÑи, ÑеÑаемÑе в Ñ Ð¾Ð´Ðµ иÑÑледованиÑ: ÐÑоанализиÑоваÑÑ ÑÑÑеÑÑвÑÑÑие алгоÑиÑÐ¼Ñ Ð¸ меÑодÑ, иÑполÑзÑемÑе Ð´Ð»Ñ Ð¿ÑогнозиÑÐ¾Ð²Ð°Ð½Ð¸Ñ ÑинанÑовÑÑ Ð²ÑеменнÑÑ ÑÑдов. ÐÑбÑаÑÑ Ð¾Ð¿ÑималÑнÑе меÑÐ¾Ð´Ñ Ð¿ÑогнозиÑÐ¾Ð²Ð°Ð½Ð¸Ñ ÑинанÑовÑÑ Ð²ÑеменнÑÑ ÑÑдов. ÐодгоÑовиÑÑ Ð¸ÑÑ Ð¾Ð´Ð½Ñе даннÑе под ÑÑÐµÐ±Ð¾Ð²Ð°Ð½Ð¸Ñ Ð²ÑбÑанного алгоÑиÑма. СделаÑÑ ÑеализаÑÐ¸Ñ Ð²ÑбÑаннÑÑ Ð¼Ð¾Ð´ÐµÐ»ÐµÐ¹. ÐоÑÑÑоиÑÑ Ð¿Ñогноз ÑазвиÑÐ¸Ñ Ð¸ÑÑ Ð¾Ð´Ð½Ð¾Ð³Ð¾ ÑÑда на 7 дней. ÐÑовеÑÑи анализ ÑезÑлÑÑаÑов пÑогнозиÑÐ¾Ð²Ð°Ð½Ð¸Ñ Ð¸ ÑделаÑÑ Ð²Ñвод о возможноÑÑи пÑÐ¸Ð¼ÐµÐ½ÐµÐ½Ð¸Ñ Ð¼Ð¾Ð´ÐµÐ»ÐµÐ¹ на пÑакÑике. Ð ÑабоÑе пÑедÑÑавлен Ð¾Ð±Ð·Ð¾Ñ Ð¾Ð±Ñего клаÑÑа моделей, коÑоÑÑе могÑÑ Ð±ÑÑÑ Ð¸ÑполÑÐ·Ð¾Ð²Ð°Ð½Ñ Ð´Ð»Ñ Ð¾Ð¿Ð¸ÑÐ°Ð½Ð¸Ñ ÑÑдов и поÑÑÑÐ¾ÐµÐ½Ð¸Ñ Ð¿Ñогноза, а Ñакже алгоÑиÑм маÑинного обÑÑениÑ. Ð ÑезÑлÑÑаÑе бÑл пÑоанализиÑован иÑÑ Ð¾Ð´Ð½Ñй вÑеменной ÑÑд ÑÐµÐ½Ñ Ð·Ð°ÐºÑÑÑÐ¸Ñ ÐºÑипÑовалÑÑÑ Ethereum. ÐоÑÑÑоено две модели: ARIMA и LSTM. РполÑÑен пÑогноз ÑазвиÑÐ¸Ñ ÑÑда на 7 дней Ð¾Ñ ÐºÐ°Ð¶Ð´Ð¾Ð¹ модели. ÐÑл пÑоведен ÑÑавниÑелÑнÑй анализ ÑезÑлÑÑаÑов моделей, коÑоÑÑй показал, ÑÑо нейÑÐ¾Ð½Ð½Ð°Ñ ÑеÑÑ Ñ Ð·Ð°Ð´Ð°Ñей ÑпÑавилаÑÑ Ð»ÑÑÑе нежели ÑÑаÑиÑÑиÑеÑÐºÐ°Ñ Ð¼Ð¾Ð´ÐµÐ»Ñ.     Ðи одна из моделей в полной меÑе не Ñмогла пÑедÑказаÑÑ Ð²ÑбÑоÑÑ Ð·Ð½Ð°Ñений ÑÑда, но пÑи ÑÑом ÑенденÑÐ¸Ñ ÑазвиÑÐ¸Ñ ÑÑда и конеÑнÑе знаÑÐµÐ½Ð¸Ñ Ð±Ñли опÑÐµÐ´ÐµÐ»ÐµÐ½Ñ Ð´Ð¾ÑÑаÑоÑно ÑоÑно. СÑеднÑÑ Ð°Ð±ÑолÑÑÐ½Ð°Ñ Ð¾Ñибка пÑогнозов модели ARIMA на ÑеÑÑовой вÑбоÑке ÑоÑÑавила 80,87. СÑеднÑÑ Ð°Ð±ÑолÑÑÐ½Ð°Ñ Ð¾Ñибка пÑогнозов модели LSTM по ÑÑÐ°Ð²Ð½ÐµÐ½Ð¸Ñ Ñ ÑеÑÑовой вÑбоÑкой ÑоÑÑавила 75,27. ÐоÑÑÑиÑÐ¸ÐµÐ½Ñ Ð´ÐµÑеÑминаÑии Ñавен 0,856, ÑÑо говоÑÐ¸Ñ Ð¾ доÑÑаÑоÑно вÑÑокой ÑоÑноÑÑи пÑедÑказаний модели.<br />The theme of the final qualification work: "Methods for forecasting financial time series". In this paper, the applied problem of forecasting the financial time series of digital currency is solved. The classical ARIMA analysis and forecasting model and the LSTM deep learning model of the recurrent neural network architecture are considered. Tasks to be solved in the course of the study: Analyze existing algorithms and methods used for forecasting financial time series. Choose the best methods for forecasting financial time series. Prepare the initial data for the requirements of the selected algorithm. Make the implementation of the selected models. Build a forecast for the development of the initial series for 7 days. Analyze the results of forecasting and draw a conclusion about the possibility of applying the models in practice. The paper presents an overview of the general class of models that can be used to describe series and build a forecast, as well as a machine learning algorithm. As a result, the initial time series of the closing price of the Ethereum cryptocurrency was analyzed. Two models have been built: ARIMA and LSTM. And a forecast for the development of a series for 7 days was obtained from each model. A comparative analysis of the results of the models was carried out, which showed that the neural network coped with the task better than the statistical model. None of the models was able to fully predict the outliers in the values of the series, but the trend in the development of the series and the final values were determined quite accurately. The average absolute error of the ARIMA model forecasts on the test sample was 80,87. The average absolute forecast error of the LSTM model compared to the test sample was 75,27. The coefficient of determination is 0,856, which indicates a fairly high accuracy of the model's predictions.
Details
- Language :
- Russian
- Database :
- OpenAIRE
- Accession number :
- edsair.doi...........b261bbd0e64c393efa8ec34697ac6c47
- Full Text :
- https://doi.org/10.18720/spbpu/3/2022/vr/vr22-2778