Implementation of AR(1) model from ARIMA Procedure

kybowma — Tue, 21 Nov 2017 17:17:58 GMT

I posted earlier, however it seems that I accidently deleted or it was removed. If this post/question is considered "taboo" or is posted in the wrong community, sorry and please let me know.

My objective is to implement a model which was scored with the PROC ARIMA procedure in SAS. Working with SAS Tech support I was able to get a more simple explanation of the backend equation that the ARIMA procedure uses. Here was the correspondence:

y_t= y_t-1 + C + w1x1_t + w2x2_t+ ϕ ((y_t-1 - y_t-2) - w1x1_t-1- w2x2_t-1 ) + a_t

The random error term at time t, a_t, takes on its expected value of 0 and effectively drops out of the equation. In the forecast horizon, when actual values of the lags of y on the right-hand-side of the equation are no longer available, the corresponding predicted value is used in place of the lagged y values.

*** Update 1 ***

Through trial and error I noticed that the above (if I implemented it right) produced a static difference for type='Actuals'. I noticed the difference being -C*ϕ, thus the below code accommodates for this and is now replicating the ARIMA procedure model fit.

New equation:

y_t= y_t-1 + C + w1x1_t + w2x2_t+ ϕ ((y_t-1 - y_t-2) - w1x1_t-1- w2x2_t-1 ) + a_t - Cϕ

Here is the parameter estimates which I believe the model to need:

Which I believe to imply from the equation above that C=MU=-15.59089, ϕ=AR1,1=-0.57206, w1=NUM1=-0.0249 and w2=NUM2=0.05201.

Can someone assist me with calculating the values in the type='Forecast' portion of the below data? My assumption is that this is close, however it is still off. For the type='Actuals' I match to within extreme rounding. Any help would be greatly appreciated.

data arima_data;
    format date date9.;
    length type $ 8;
    do date='01JAN2000'd to intnx('month',today(),-12,'B');
        date=intnx('month',date,0,'B');
        type='Actual';
        y=ranuni(28269)*1000;
        x1=ranuni(28123)*1000;
        x2=ranuni(28722)*1000;
        output;
        date=intnx('month',date,0,'E');
    end;
    /* Out of time. */
    do date=date to today();
        date=intnx('month',date,0,'B');
        type='Forecast';
        y=.;
        x1=ranuni(28123)*1000;
        x2=ranuni(28722)*1000;
        output;
        date=intnx('month',date,0,'E');
    end;
run;

proc arima data=arima_data;
    /* Difference model with 2 x terms */;
    identify var=y(1) crosscorr=(x1 x2) noprint;
    /* Including 1 lag. */
    estimate input=(x1 x2) p=1 noest noprint
        ar=-0.57206
        mu=-15.59089
        initval=(-0.02439 x1 0.05201 x2)
    ;
    forecast interval=month id=date out=arima_forecast lead=12 noprint;
run;quit;

data merge_arima_forecast;
    merge arima_data (in=mas)
          arima_forecast (in=fcst keep=date forecast residual)
    ;
    by date;
run;

data implementation;
    set merge_arima_forecast;
    by date;

    /* From SAS Tech Support:
    y(t)=y(t-1) + C + beta1*x1(t) + beta2*x2(t) + AR(1)((y(t-1) - y(t-2)) - beta1*x1(t-1) - beta2*x2(t-1)) + a(t)
    Assumed, C=Mu (-15.59089), AR(1)=Auto Regressive Parameter (-0.57206), beta1 (-0.02439) and beta2 (0.5201)
    a(t) is the error term which is essentially zero and drops out of the model.
    */

    y_lag=lag(y);
    y_lag2=lag2(y);
    x1_lag=lag(x1);
    x2_lag=lag(x2);

    c=-15.59089;
    ar=-0.57206;
    beta1=-0.02439;
    beta2=0.05201;

    /* Actuals */
    *if type='Actual' then imp_forecast=sum(y_lag,c,beta1*x1,beta2*x2,ar*((y_lag-y_lag2)-beta1*x1_lag-beta2*x2_lag));
    /* Updated with -c*ar, matches the ARIMA fit. */
    if type='Actual' then imp_forecast=sum(y_lag,c,beta1*x1,beta2*x2,ar*((y_lag-y_lag2)-beta1*x1_lag-beta2*x2_lag),-c*ar);
    /* Forecast */
    lag_imp_forecast=lag(imp_forecast);
    lag2_imp_forecast=lag2(imp_forecast);
    if type='Forecast' then imp_forecast=sum(lag_imp_forecast,c,beta1*x1,beta2*x2,ar*((lag_imp_forecast-coalesce(y_lag2,lag2_imp_forecast))-beta1*x1_lag-beta2*x2_lag),-c*ar);

    arima_forecast_diff=forecast-imp_forecast;

run;

Re: Implementation of AR(1) model from ARIMA Procedure

Ksharp — Wed, 22 Nov 2017 12:56:29 GMT

Post it at Forecast forum, since it is a time series analysis question.

Re: Implementation of AR(1) model from ARIMA Procedure

kybowma — Wed, 22 Nov 2017 13:11:26 GMT

Hi Ksharp, I posted there originally, however it was flagged as spam by SAS. SAS has sense marked it as a legible question.

https://communities.sas.com/t5/SAS-Forecasting-and-Econometrics/Implementation-of-AR-1-Model-from-ARIMA-Procedure/m-p/415175#M2869

I appreciate your feedback!

topic Re: Implementation of AR(1) model from ARIMA Procedure in Statistical Procedures

Implementation of AR(1) model from ARIMA Procedure

Re: Implementation of AR(1) model from ARIMA Procedure

Re: Implementation of AR(1) model from ARIMA Procedure