Financial Heterogeneous Information Acquisition And Prediction Ananlysis Using Multi-level Model

With the development of the big data and Internet technology, the traditional method of finance has been unable to satisfy users’ personalized, diversified, automation needs, heterogeneous information processing and data mining techniques to obtain useful information are becoming more and more important. To help financial users make investment and show users comprehensive and accurate financial information, we need to obtain relevant information from many sources, analyze them to provide valuable advice to users. The thesis aims at financial users interested problem, extract the important features to investors from heterogeneous information from the user’s perspective, analyze the characteristics, then use a multi-level model to predict the financial data for users to help them make strategies and investment.The main contents of this paper are shown below:Ontology adaption method to extract features from PDF based on rules: includes information extraction from the convertible bonds listing announcement and IPO PDF documents. Firstly converse different PDF documents. Secondly create ontology rules warehouse for different financial products. Change weight of some rule adaptively. For the new document, we choose the rule which has high weight firstly and get the feature according to pattern matching methods. Finally, make extracted information normative and get comprehensive and accurate features.Multiple heterogeneous information acquisition: the acquisition of heterogeneous information includes three parts, real-time financial data, financial data in PDF and relevant historical data from web. Use socket transportation, pattern matching and regular expression to get data from each source respectively, then make data validation with the ontology knowledge verification, third-party verification, manual verification, cross validation.Prediction analysis of financial data: predicting the feature of convertible bonds and closed-end funds according to establishing multi-level model. In the first level, use three single models, trend estimation model, SVR model, Neural Network model to predict financial data; In the second level, use the predicted result as the input of the neural network to get fused results; In the third level, improve the second level through genetic algorithm to select the optimal and appropriate weights and thresholds to obtain better results. Finally, the multi-level model completed the prediction of the convertible bond and closed-end funds with higher accurate rate.