A Stock Price Prediction Method Based On Meta-Learning And Variational Mode Decomposition

With the rapid development of China’s stock market and the increase of per capita disposable income,people are increasingly enthusiastic about stock speculation,and the number of investors in China’s securities market has exceeded 200 million.As a classical problem at the intersection of finance and computer science,stock price prediction has been attracting attention from stock investors and scholars.Correct prediction results can provide investors with valuable guidance and thus reduce the risk of investment,so stock price prediction is an important research with broad application prospects.However,stock market is complex,and stock price data is characterized by high noise and nonlinearity,so the research of stock price prediction is very challenging.Among many prediction models,decomposition-based hybrid models have shown superior performance.However,the existing decomposition-based hybrid models generally suffer from data leakage,which can lead to a look-ahead bias in the models.In addition,the distribution of stock price data changes over time,resulting in concept drift,but the traditional forecasting methods cannot handle this phenomenon well.In this thesis,we propose a new stock price forecasting method,VML,which combines variational mode decomposition(VMD),model-agnostic meta-learning(MAML)algorithm and long short-term memory(LSTM)network.The VML model first uses sliding windows to slice the stock price series to obtain multiple window series,and then uses variational mode decomposition to decompose the window series to obtain multiple subseries.After that,the decomposed subseries are divided into multiple tasks.The MAML algorithm trains initial parameters with good generalization ability for the LSTM network on the source task set.The LSTM network with initial parameters is fine-tuned on the support set of each target task to better fit its data distribution,then makes predictions on the query set of the target task.Finally,the model merges the prediction results of the subseries to obtain the final predicted stock price.In this thesis,extensive experiments are conducted on stock datasets from the Chinese Stock Market and the American Stock Market,and the results show that the constructed VML model improves the prediction accuracy.In the proposed VML model,the window series is decomposed instead of the entire stock series,eliminating the look-ahead bias.In addition,this thesis introduces meta-learning into the stock price prediction problem,which provides a new idea for stock price prediction.This enables the model to be dynamically fine-tuned according to the latest data distribution,thus reducing the impact of concept drift on the prediction accuracy of the model.What’s more,in order to use the meta-learning on the stock price prediction problem,this thesis defines a way to divide the stock price data into tasks.