You are on page 1of 4

Information Sheet on ARIMA Modelling and Forecasting in EViews To specify the sample period to be used for the model

use the smpl command, e.g. smpl 1959:1 1999:4 In the series dialog box we can use the command view/correlogram to see the ACF and PACF, as well as the Ljung and Box Q statistics. When choosing correlogram you are given the option of looking at the ACF of the original series or of the series in first and second differences. To difference a series you can use the series or genr command, e.g. to construct the first difference of a series y, we can type, genr dy = y y(-1) genr dy = d(y) series dy = d(y) To estimate ARIMA models on differenced series you can do one of two things: (1) Construct the differenced series using the genr or series commands and use ARIMA on the differenced series (2) Use the difference operator directly in the estimation specification, e.g. ls(h) d(y) c stuff The latter method should be preferred. The reason being that when forecasting, EViews will provide forecasts of the differenced series using the first method, but will give you the opportunity of forecasting on the original series using the second method. More complicated forms of differencing can also be used, examples being d(y,n) specifies the nth order difference of the series d(y,n,s) specifies the nth order ordinary differencing of the series with a seasonal difference at lag s. The dlog option allows you to consider the differences in the logs of the original series. AR and MA components of the model can be specified using the ar and ma keywords. For example, ls(h) y c ar(1) Would specify an AR(1) model, while ls(h) y c ma(1) Would specify an MA(1) model These can obviously be combined to form an ARMA model, such as ls(h) y c ar(1) ma(1) ma(2) Specifies an ARMA(1,2) process, while ls(h) d(y) c ar(1) ma(1) ma(2) Specifies an ARMA(1,2) process on the first difference of y and is thus an ARIMA(1,1,2) process.

The ar and ma terms need not be included consecutively, that is, we could have, ls(h) y c ar(1) ar(3) ar(7) One way of accounting for seasonality is to include so-called additive seasonal terms, which simple involves including ar and ma terms of the correct order. For example, using quarterly data we may have, ls(h) y c ar(1) ar(4) ma(4) An alternative is to use multiplicative seasonal terms which allow for the interaction between the standard ARMA terms and the seasonal terms. For a multiplicative seasonal autoregressive model on quarterly data we may write ls(h) y c ar(1) sar(4) And, for a multiplicative seasonal moving average model we could write ls(h) y c ma(1) sma(4) Forecasting To forecast we can use the forecast button on the equation toolbar, or select procs/forecast. The forecast box gives a number of options Forecast name name that you give to the forecasted series (make sure this is different to the name of the original series otherwise EViews will overwrite the original series). S.E. you may provide a name for the series to be filled with the forecast standard errors. If no name is included, then the forecast standard errors will not be saved. GARCH allows you to save the forecasts of the conditional variances if a GARCH model is used. Forecasting method: Dynamic Calculates multi-step forecasts starting from the first period in the forecast sample (it uses forecasted values in previous periods to construct forecasts in future periods). The starting point of the forecast sample has an important bearing on the forecasts using dynamic forecasting. Static calculates a sequence of one-step ahead forecasts using actual rather than forecasted values for lagged dependent variables (Note: this method is not possible if you are forecasting beyond the sample period, i.e. for a period for which you dont have data on the dependent variable). Static forecasts are more accurate than dynamic forecasts because for each period they use the actual value of the lagged values in forming the forecast rather than forecasted values of the lagged values. Sample Range you can set the sample to be used for forecasting. EViews sets this to the sample of the workfile. You can set it outside the sample range used in your model, thus allowing you to produce out-of-sample forecasts. Note: To do this you may need to expand the workfile range, using the expand command: expand 1959:1 2001:4

Output you can choose to see the forecast output as a graph or a numerical evaluation or both. If there are data available on the forecasted variable for the forecast sample, Eviews reports a table of statistical results evaluating the forecast. Suppose the forecast sample is j = T+1, T+2,,T+h, and denote the actual and forecasted value in period t as yt and yt , respectively. The reported forecast error statistics are computed as follows: Root Mean Square Error
T +h

t =T +1

(y
t

yt ) / h
2

Mean Absolute Error


Mean Absolute Percentage Error
Theil Inequality Coefficient

T +h t =T +1

yt / h
yt yt /h yt
t

T +h

100

t =T +1

T +h

t =T +1 T +h

(y

yt ) / h
2 T +h t =T +1

t =T +1

yt 2 / h +

2 t

/h

The first two forecast error statistics depend upon the scale of the dependent variable and should be used to compare forecasts for the same series across different models: the smaller the error, the better the forecasting ability of that model according to that criterion. The remaining two statistics are scale invariant. The Theil inequality coefficient always lies between zero and one, where zero indicates a perfect fit. The mean squared forecast can be decomposed as 2 2 ( yt yt ) 2 / h = (( yt / h ) y ) + (s y s y ) + 2(1 + r )s y s y ,
where

/ h, y , s y and s y are the means and standard deviations of yt and y, and r is the correlation between y and y. The proportions are defined as follows:
t

Bias Proportion

Variance Proportion
Covariance Proportion

(( y / h ) y ) (y y ) / h (s s ) (y y )/ h
2 t 2 t t
2 y y t t

2(1 r ) s y s y (y
t

yt ) / h
2

- The bias proportion tells us how far the mean of the forecast is from the mean of the actual series - The variance proportion tells us how far the variation of the forecast is from the variation of the actual series - The covariance proportion measures the remaining unsystematic forecasting errors If the forecast is good, the bias and variance proportions should small, so that most of the bias should be concentrated on the covariance proportions.

You might also like