Forecast On Paper Downloads Based On ARIMA Model And Neural Network

Nowadays China has NO.1 quantity of paper publishing in the world, and with the development of Internet technology, commercial database and the increase of the open source database, the spread of paper is also speeding up. With the good and bad are intermingled of literature how to more scientific evaluation of its quality,It has become a hot issue in the literature metrology that how to scientificly evaluate its quality among the amount of spotty literatures. Previous evaluation index has many disadvantages, the improvement of paper download detection system makes paper downloads become one of the most important standards of paper evaluation index. Scholars mainly studied the thesis downloads index should be included in the evaluation of paper quality system, the relationship between paper downloads and the number of the paper citations, and the law of some paper downloads with some statistical methods. This paper will predict paper downloads on the basis of the time series and neural network in the stochastics.This paper collected thesis downloads of SQL server data in the Biology plate of PLOS from June 2003 to May 2014, and then after data preprocessing, the data from November 2003 to May 2014 is chosen to analyze. With the help of SQL server software for data preprocessing and EXCEL software for descriptive statistics data, all scatterplots are roughly the same. It proves that the data has general representative; This paper uses SAS software to analyze time series of data and thenmakes ARIMA model and prediction. It also uses BIC criterion to select the optimal estimation model. And both parameter estimation and residual white noise test pass the statistical test, with statistical significance. It comes out five-month forecast data and the average prediction error is 19.43%. Predicting scatterplot shows predictive value and true value are closely the same. They are both in the 90% confidence interval;With the help of matlab software, it can make BP neural network modeling and forecasting for the data. Prediction value and true value are very close. And the evaluation of BP neural network prediction error is only 4.32%. It proves that the prediction result is reasonable. In the end, it concludes that the BP neural network prediction result is better than time series prediction results.