APP Download Trend Prediction And Prediction Model Evaluation

Nowadays,smart phones have become an indispensable communication tool in our daily life,and the competition of APPs in smart phones has become extremely fierce.At present,most of the research is based on the download of the APP.Some statistical methods are used to show the rules of the downloads,and analyze the market factors that influence the downloads.This paper aims to utilize the time series ARIMA model,residual autoregressive model and neural network to predict APP downloads.This article has collected the daily downloads of one APP in the mobile APP market from January 1,2017 to March 7,2018.By using sas software to analyze all the data,then we established ARIMA model and forecast future five period.We use the BIC rule to choose the best ARIMA model,when the parameter estimation and residual white noise have passed the statistical tests,the model we choosed has statistical significance.Then the scatter plot shows that the fitted values are close to the real data.Moreover,most of the data are within 90% confidence interval,which indicates that the ARIMA model fits the data well.However,the prediction error of the model is farely large,with an average error rate22% for a future five period prediction;Then,the residual autoregression model was performed,and the DW test value was 0.1288,indicating that the residuals had obvious self-correlation.After the model was established,the average error was predicted to be 7%.Finally,by using MATLAB software to model and predict the neural network model,the average error of the next five period is 5%.On the whole,the fitting effect of residual autoregressive model and neural network model is better than ARIMA model in terms of the sample data we take.