6SIGMA工具ARIMAForecasting.ppt

The Autocorrelations show a significant value only for the first lag and not much structure after that. The Partial Autocorrelations appear to rapidly diminish with lag, notice that there is more structure than in the Autocorrelations after the first lag. So the MA(1) model is a reasonable candidate for this series. The trend of the data doesn’t show any obvious relationships between mean and variance with time. We can assume “stationarity” Modified Box-Pierce Chi-Square Statistic where k=lag and m=# of parameter estimates Original Series A slowly decreasing pattern in the autocorrelations for sequential lags is an indication of a non-stationary mean. Series with a stationary mean will have an autocorrelation pattern that decreases rapidly. Differencing creates a new series by subtracting each value by it’s immediate predecessor: Wt = Yt – Yt-1 Wt+1 = Yt+1 – Yt With a nonlinear trend with time, the data can be stabilized by repeatedly “differencing” After applying differencing twice, the trend against time has evened out, but there might be a relationship to variation with time that also has to be dealt with in order to create a stationary series. Original Sales Time Series Transforming the data to suppress the changes in variation, and then applying the differencing, creates a series that appears to be “stationary” Note that for verification of stationary time series, the analyst will look for an autocorrelation pattern that decreases rapidly with lag. The raw data shows both nonstationary behavior with the mean and a cyclical component. Note that the autocorrelation values seem to rise every fourth lag. In order to eliminate the seasonality, transform the data by differencing on the fourth lag. It’s good to check the autocorrelation pattern to confirm that the series has been stabilized into “stati