Outlier-Detecting&Alarming Model Of The Internet Company's Operational Index Based On Statistical Method

With the rise of e-commerce websites,more and more people are shopping online.And these enterprises collect enormous user behaviors data on daily basis with the help of advanced computer science technology.To operational team in these enterprises,troubles brought by the data explosion are always affecting their decision-making.In this thesis,time-series-model,mean-range method and some other statistical methods are used to detect outliers and monitor indices.Unlike traditional detecting method in financial fraud and clinical abnormality,methods in this thesis mainly focus on whether the indices enterprise are different from average level,once any deviations occurs,alarm will automatically be set.In the thesis,detection method in outliers in monistic normal distribution and box-plot in non-parameter statistics and local abnormality factor detection method in unsupervised learning are firstly used in abnormality detection in training data set.Then we use mean value of data or growth trend in the same period except abnormal outliers to replace these outliers.We creatively use paired T-test to recognize holiday pattern and divide the data into holiday data and non-holiday data.To non-holiday data,we combine Seasonal product ARIMA model and Triple Exponential Smoothing method to predict the range of one-forward-step,and combine former result with mean-range quality control charts to confirm the range of prediction.To holiday data,we first get the diff of real data and Seasonal product ARIMA model and then use Fourier spectral analysis for fitting the series.Finally,we use the signal function to combine the holiday data model and non-holiday model and get the prediction model in the whole period.If one index falls outside of range of normality,outcome of abnormality will be displayed.In empirical analysis,we mainly analyze convert-ratio in one OTA company.The paired T-test shows that the index has obvious holiday-effect in 12 days before the National Day till 6 days after it and the pattern has a clear U-shape during these days.We use the abnormality detection method mentioned above and find outliers in 5 training data and interpolate replacement which are approved effectively and quite close the real business data.We then build Seasonal product ARIMA model and Triple Exponential Smoothing method.Judging from out,sample MSE,we finally chose Seasonal product ARIMA model and combine it with mean-range quality control chart to get non-holiday data prediction model.The out-sample MSE is 7.78E-8.And we chose Two-Cycle Fourier Spectral Analysis Model after overlooking the data of the National Day.Finally,we use the signal function to combine the models to get the whole monitor model.The method diffused from the thesis enables the abnormality recognition ratio get close to 100%and effectively solve the problems in practical work.