Last revision: Monday, 12 April 2001.
STAMP 6.0 is a package designed to model and forecast time series. It is based on structural time series models. These models use advanced techniques, such as Kalman filtering, but are set up so as to be easy to use - at the most basic level all that is required is some appreciation of the concepts of trend, seasonal and irregular. The hard work is done by the program, leaving the user free to concentrate on formulating models and using them to make forecasts.
Independent reviews of the earlier STAMP version 5.0
can be found in:
- Review of STAMP 5.0, by Guy Judge, The Economic Journal, 106, July 1996.
-
Structural Time Series Analyser, Stamp 5.0, by Guy Judge, CHEER, 9, June 1995.
- Stamp 5.0: A Review,
by Diebold, Giorgianni, and Inoue, International Journal of Forecasting, 1996.
- Forecasting with STAMP by D.J. Pedregal, O.R. Insight, Vol. 9, No. 3, 29-32, 1996.
System requirements: Windows 2000/NT/98/95.
Click here for more information on STAMP 6.0 (including new features).
Structural time series modelling can be applied to a variety of problems in time series. Macro-economic time series like gross national production, inflation and consumption can be handled effectively, but also financial time series like interest rates and stock market volatility can be modelled using STAMP 6.0. Further, STAMP is used for modelling and forecasting time series in medicine, biology, engineering, marketing and in many other areas.
by Jordi van Kessel.
To get some impression of STAMP 6.0 we present here some illustrations:
Univariate modelling: Energy.in7
Univariate modelling with explanatory variables: Spirit.in7
These illustrations do not show the entire functionality of STAMP 6.0. Multivariate structural time series modelling offers a wide range of possibilities like homogeneity restrictions, VAR models, cointegration / common factors, parameter and variance matrix restrictions, using a control group to monitor effects of laws or changed behaviour. Illustrations on these topics are being developed and will appear soon on this web page.
Start GiveWin, load the Energy dataset by clicking the Open File icon and select Energy.in7, which is usually located in the same directory as STAMP. The energy dataset consists of quarterly consumption of coal, gas and electricity for the 'Other industries' and 'Other final users' in the UK. The letter l, attached to the name of the series, implies that the data are in logarithmic form. In GiveWin, the Tools menu enables you to select Graphics where you can graph various time series. Below are graphs of the log values of these series. These were created by selecting the variables in the Graphics menu. All series have a clear trend and seasonal component. An interesting shift in the level of oiGASl in the early seventies can be detected. Here we will examine the ofuCOALl (log coal comsumption of other final users) series using STAMP 6.0, which is the purple dotted line in the graph below. It has a downward trend and the seasonal component seems to be rather constant over time.
In the Modules menu of GiveWin, you can start STAMP. In the Model menu of STAMP you should select Formulate.... In the Formulate a model dialog, select ofuCOALl as the Y variable in the first window.
Accept the default settings in the other dialogs leading to the estimation of a basic structural time series model. STAMP will produce an estimation report. This report tells us that convergence was VERY STRONG, which is good news. Under 'summary statistics' some elementary statistics like the Box-Ljung Q-statistic, Durbin-Watson statistic are printed. Finally the variances of the selected stochastic components are given. The variance of the seasonal disturbance is estimated to be zero which implies that component is fixed.
After estimating the model you can access the Test menu. In this menu you can evaluate some diagnostics to check the correctness of the model for the selected time series. For a start you can print extra output in the Results window by entering the Further Output dialog and mark the boxes with the topics you're interested in. The exact type of output depends on the model you have formulated. Marking the following boxes gives some interesting additional output. It outputs the standard deviations of disturbances, which are the square roots of the variances already given in the standard output. The coefficients of the final state vector are the estimated state coefficients at the end of the sample and they will be used to make forecasts. At the moment the value of the growth rate is not significantly different from zero as the t-value with its accompanying two-sided probability value shows. Further the Anti-log analysis gives a current growth rate of -2.74% per year. This seems correct according to the graph of the coal consumption shown before. Finally the seasonal analysis gives the percentage of difference with the underlying trend for each season.
It is also possible to inspect your results graphically with the Components Graphics dialog where you can select any option you like. The graph below was created using the default options. If you wish you can edit the appearance of the graph by right-clicking and selecting Edit Graph in GiveWin. There are many possibilities of printing graphs generated by STAMP (see the dialog box). The graph below shows: (i) the ofuCOALl series with its trend; (ii) the seasonal component which is constant over time; (iii) the irregular component which move randomly around zero but with a few large values. In the last quarter of 1984 there was a strike of coal miners and this appears in the irregular. Finally there is an interesting option which produces the seasonally adjusted series. This series can also be saved to the data set by clicking Store when only the box with Seasonally adjusted Y is selected. Now you can name the seasonally adjusted variable. Please don't forget to save the data set if you want to keep this variable in GiveWin.
In the Residuals Graphics dialog you can draw the time series of residuals, its correllogram, its estimated density and more. There is also a dialog Auxiliary Residuals Graphics which produces smoothed estimates of the irregular and level disturbances. These can be useful to detect outliers and structural breaks or distinguish between them.
One of the important tasks of time series modelling is forecasting. STAMP offers two possibilities. The first one is prediction and is accessed by choosing Prediction Graphics. If this option is not accessible then first choose Model and Estimate. Now change the value for Less forecasts to the amount of periods you want to predict. The graphs below show the predictions from 1984 till the end of the sample together with the real values and twice the root mean squared errors. Multi-step prediction means that the predictions are made using the information till the end of 1983 and are not updated with the arrival of new observations contrary to one-step prediction. It is very interesting that the multi-step prediction turns out to predict the first three quarters of 1985 better than the one-step prediction. The reason for this is the strike of the coal miners in quarter three and four of 1984. These quarters may be treated as outliers which is very easy to do in STAMP.
After examining the prediction properties we can do a post-sample prediction (don't forget to re-estimate the model using the full sample), because the model seems to be reasonably good at predicting. Just click Test and Forecasting to enter the following dialog box and select the options you are interested in. In this example, the standard options are selected, 5 years of forecasting and plots of ofuCOAL, its forecast +/- 2 R.M.S.E and a forecast of the trend.
This is just a simple example showing the abilities of STAMP. This package also offers the possibility of entering explanatory variables and multiple dependent variables in the model. Further up to three cycles and an autoregressive term can be specified.
In this example a famous data set will be examined: annual observations of per capita consumption of spirits in the UK, per capita income and the relative price of spirits from 1870 to 1938. This data set is not available in the demo, but it is in the full version. An econometric model may consists of some explanatory variables and a linear or quadratic trend plus an AR(1) disturbance. In a structural time series model the linear trend is replaced by a stochastic trend.
First load the data in GiveWin and then choose Formulate to select spirits as dependent (Y) variable and then select price and income as explanatory variables. It is not necessary to highlight them as X-var in this case. In estimation report are the standard output of this regression, followed by some Additional output: State and regression output and Anti-log analysis. Of particular interest is the estimated coefficients of explanatory variables. As in standard regression these coefficients can be interpreted as elasticities: a one per cent increase in price leads to a fall in spirits consumption of 0.95 per cent.
To detect outliers and structural breaks the auxiliary residuals can help. In this case auxiliary residuals are printed for the irregular and level component. STAMP also prints the periods and values of large residuals in the results window. When a level auxiliary residual is large this indicates a structural break, which seems to be the case for 1909. This can be tested by re-estimating the model with a level intervention in 1909 in the Model Interventions dialog and do a likelihood ratio test. Furthermore you can introduce interventions for outliers such as the ones in 1915 and 1918.
There are several extra possibilities for forecasting when working with explanatory variables to implement the future values of these explanatory variables:
We now compare the state space model with the AR(1) model with a linear trend by setting fixed level and fixed slope in STAMP and marking the Autoregression box. Also the last 8 observations are omitted to be able to compare the predictive value of the models. This model gives a loglikelihood value and Rd^2 similar to the previous model, but the multi-step predictions are worse than the ones for the previous structural time series model.
The following changes from version 5.0 to version 6.0 have been made but are not reported in STAMP 6 book (appendix A2):
The following bugs have been identified in the STAMP program version 6.0:
These problems have been corrected in version 6.01.
The following bugs have been identified in the STAMP program version 6.01:
The current users of the full version STAMP 6.0 can download the
STAMP 6.02 installation program.
A valid STAMP 6.0 licence is required for this upgrade.
Changes from STAMP 6 to STAMP 6.01:
Changes from STAMP 6.01 to STAMP 6.02:
!!! Free Demo version of STAMP 6.0 !!!
You can freely install the PcGive/PcFiml/Stamp Demo and start exploring the capabilities of STAMP 6.0.
Click here for more information on STAMP 6.0 (including new features).
STAMP 6.0 can be ordered via Timberlake Consultants Ltd.
Timberlake Offices can be found in the US and Europe.
