Earnings Forecast Based On Recurrent Neural Network And Online Comments Mining

Many traditional industries are beginning to change in the direction of modern management with the development of computer technology.Companies are also introducing new tools as new management style,such as the use of ERP systems,marketing through the Internet and so on.Companies and users generate large amounts of multidimensional data,the data itself can reflect the companies' status and provide valuable information for the development of the companies.Quoted companies usually have a longer duration of financial data which can be used for earnings forecast with traditional statistical methods,but its difficult to do earnings forecast for medium and small size companies as they generally lack of sufficient financial data.This paper developed a model focuses on the restaurants which can be seen as small and medium sized companies earnings forecast problem.The restaurants are more likely to encounter problems such as bad management due to the the catering industry characteristics.Earnings forecast not only can help manage but also can be combined with risk investment for risk assessment.This paper use two kind of data for research,one of which is about 2000 restaurants,monthly earnings data for one,another is comments from the online reviews site.Because monthly earnings are not suitable for traditional time series models,this paper developed a new earnings forecast model named Maxout-LSTM based on LSTM(Long Short-Term Memory)neural network and Maxout activation function.The experiment proved that Maxout-LSTM outperform than SVR and GM.Beside the traditional financial data,this paper use the futures from comments mining.Futures from comments can be divided into four categories:basic rank feature,comments topic feature,comments sentiment feature and the rank correction feature.The experiment show that features from comments enhanced the performance of Maxout-LSTM earnings forecast model.