Application Of Time Series Analysis And Machine Learning In Fund Flow Forecasting

With the advent of the big data era,the Internet financial industry represented by Ant Group is developing rapidly in China.The Yu'E Bao business under the Ant Group has become one of the most famous cash management tools.The huge user base has brought lots of fund flows,which has increased the requirements for the management pressure and risk control capabilities.How to accurately predict the capital flows is the focus of business development in the context of Internet finance.Based on the real purchase and redemption fund flow data of Yu'e Bao,this article aims to compare and analyze the performance of various forecasting methods on the fund flow forecasting.The main research contents and results of this paper are as follows:1.Use time series analysis methods to predict fund flow time series data,including seasonal ARIMA model and Prophet model.In order to make the modeling more refined,the forecast targets are split before the modeling,and finally the results are summarized and compared.2.Perform feature engineering for fund flow time series data for machine learning regression methods to forecast the fund flow,and compare with the results of time series analysis methods.In the feature engineering step,combined with the real fund flow business,feature mining and feature screening are carried out from the aspects of date,sequence and auxiliary information.Based on the selected feature subsets,multiple linear regression and machine learning models are used for prediction,including random forest,GBDT,XGBoost and Light GBM models.3.The results shows that,in the time series methods,the Prophet model forecasts better than the seasonal ARIMA model,especially in the purchase capital flow scenario,RMSE and MAPE are reduced by about 40%;the overall forecasting performance of the machine learning regression methods is better than the time series method,especially in the redemption scenario,the RMSE and MAPE of each model are reduced by about 30%;in the machine learning regression methods,the combination of feature engineering and the XGBoost model has the best predictive ability and is most suitable for the task of fund flow prediction.