Escolar Documentos
Profissional Documentos
Cultura Documentos
R Boojhawon
Identifying the orders (p,d,q) through use of ACF/PACF of differenced plots. But be careful unnecessary differencing may introduce correlation Numerical estimation of model parameters (& CIs) by using Yule-Walker equations or method of MLE or method of LSE. Residual Diagnostics: Graphs of Residual ACF/PACF, QQ, histograms, Ljung-Box-Pierce Q statistics to indicate white noise else we need to resimulate using other parameters. Do predictions
TSA By Assoc. Prof. R Boojhawon, UoM 3
Goal
To emphasize plotting methods that are appropriate and useful for finding patterns that will lead to suitable models for our time series data.
A Model-Building Strategy
We will develop a multistep model-building strategy
espoused so well by Box and Jenkins (1976). There are three main steps in the process, each of which may be used several times:
model specification (or identification) model fitting, and model diagnostics
parsimony
The quarterly earnings per share for 19601980 of the U.S. company Johnson & Johnson, are saved in the file named JJ. Plot the time series and also the logarithm of the series. Argue that we should transform by logs to model this series. The series is clearly not stationary. Take first differences and plot that series. Does stationarity now seem reasonable? Calculate and graph the sample ACF of the first differences. Interpret the results. Display the plot of seasonal differences and the first differences. Interpret the plot. Recall that for quarterly data, a season is of length 4. Graph and interpret the sample ACF of seasonal differences with the first differences.
Fit the model ARIMA(0,1,1)(0,1,1)4, and assess the significance of the estimated
Case Problem 1
coefficients. Perform all of the diagnostic tests on the residuals. Calculate and plot forecasts for the next two years of the series. Be sure to include forecast limits.
TSA By Assoc. Prof. R Boojhawon, UoM 6
Step 2: The series is clearly not stationary. Take first differences of the log data and plot that series. Does stationarity now seem reasonable?
We do not expect stationary series to have less variability in the middle of the series as this one does but we might entertain a stationary model and see where it leads us.
TSA By Assoc. Prof. R Boojhawon, UoM 8
Step 3: Calculate and graph the sample ACF of the first differences of ln(data). Interpret the results.
Strongest autocorrelations are at the seasonal lags of 4, 8, 12, and 16. Clearly, we need to address the seasonality in this series. Also graph suggests MA effect as well.
TSA By Assoc. Prof. R Boojhawon, UoM 9
10
Step 4: Display the plot of seasonal differences and the first differences. Interpret the plot. Recall that for quarterly data, a season is of length 4.
The various quarters seem to be quite randomly distributed among high, middle, and low values (e.g we see 4 up, mid,down), so that most of the seasonality is accounted for in the seasonal difference.
TSA By Assoc. Prof. R Boojhawon, UoM
11
Step 5: Graph and interpret the sample ACF of seasonal differences with the first differences.
They only significant autocorrelations are at lags 1 and 7. Lag 4 (the quarterly lag) is nearly significant. (2 out of 20 = 0.1 can be considered as non-significant here)
TSA By Assoc. Prof. R Boojhawon, UoM 12
Step 6: Fit the model ARIMA(0,1,1)(0,1,1)4, and assess the significance of the estimated coefficients.
Both the seasonal and nonseasonal ma parameters are significant in this model since due to small p-values.
TSA By Assoc. Prof. R Boojhawon, UoM 13
No inadequacies with the model. There is little autocorrelation in the residuals/Independence due to p-values from LjungBox being large (large p-values) and hence cannot reject Ho: error is IID
TSA By Assoc. Prof. R Boojhawon, UoM 14
15
Histogram
16
Normal QQ plot
17
Normally Test
18
Step 8: Calculate and plot forecasts for the next two years of the series. Be sure to include forecast limits.
Forecasts follow the general pattern of seasonality and trend Forecast limits give a good indication of the confidence in these forecasts.
TSA By Assoc. Prof. R Boojhawon, UoM 19
20
21
Conclusion
We may extend our models using higher parameters
but it is good practice to use the simplest one (least parameters: Principle of Parsimony) which satisfies the model assumptions.
22