The Funds Inflow And Outflow Prediction Model Based On Feature Extraction And The Two-step Regression Algorithm

Financial companies have millions of service members,so a company's financial business will inevitably involves a lot of funds inflows and outflows every day.Facing such a huge financial data,it is very important to accurately forecast the inflow and outflow of funds,while ensuring the minimum funds liquidity risk and satisfying the daily business operation.However,the financial funds inflow and outflow always changes with influence of political,economic,major events etc.It makes the changes of funds inflow and outflow unstable and contains lots of noise,which bring the difficulties of prediction.Based on the funds flow of the financial platform users' data,this dissertation aims to construct an accurate and effective predicting model of inflow and outflow,in order to get close to the real value and facilitate the management of funds.The main contents and achievements of this dissertation are as follows:1.In this dissertation,with lacking of the initial features of the funds flow prediction,we use the feature extraction methods to dig out the relevant features.And then we take the feature selection strategy to obtain the optimal feature subset.The method is mainly from the time,the user,the interest rate three different aspects to construct the features related to the target value,and then uses the Pierre correlation coefficient method to do the preliminary screening the most relevant features.Later,we delete less relevant features and redundant features to get the final optimal feature subset.The experiment shows that these feature subset get different results when training on different algorithms.Finally,the 12 column features for purchase value(the 10 column features for redeem value)performs best.So this feature subset can be applied to the following algorithms in the next step.2.To solve the problem of unstable and multi-noise in the datasets,we propose a two-step regression algorithm to train and predict.The algorithm includes single-step prediction of unknown features and BP neural networks,which means it uses grey prediction or time series algorithm to predict unknown features in future period.And then we use BP neural network to train all features to get final results.Meanwhile,we use ensemble learning method,which includes gradient boosting decision tree algorithm based on Adaboost and random forest algorithm based on bagging to compare.Showing as the results,the two-step features prediction method performs better than original methods.Even it performs better than ensemble algorithms with suitable parameters.3.We use one-hot encoding on discrete type features in financial datasets,and then apply factorization machine algorithm to predict.When the feature subsets become sparse,the factorization machine can well express the interaction between variables,which means to add the secondary cross features based on the original feature variables.In this way,the FM algorithm can describe the datasets better.Besides,the factorization machine can also deal with datasets effective with low model complexity.Experiments show that the factorization machine algorithm improves the prediction accuracy of the funds inflow and outflow.