The FORECAST Procedure

For forecasting by using exponential smoothing methods or Winters method, consider using the ESM procedure. The models that can be selected by the PROC FORECAST options METHOD=EXPO, METHOD=WINTERS, and METHOD=ADDWINTERS are provided by PROC ESM, which also provides additional forecasting methods that PROC FORECAST does not support. Unlike PROC FORECAST, the ESM procedure optimizes the smoothing weights for the forecasting model based on the data. Also unlike PROC FORECAST, the ESM procedure can automatically select the form of exponential smoothing model that is most appropriate for your data. For information about forecasting with PROC ESM, see Chapter 14: The ESM Procedure.

For forecasting by using time trend models with autoregressive errors, consider using the AUTOREG procedure. The models that can be selected by the PROC FORECAST options METHOD=STEPAR and TREND= can be fit and forecast using PROC AUTOREG, which also allows the inclusion of additional predictor variables in the forecasting model. For information about PROC AUTOREG, see Chapter 8: The AUTOREG Procedure.

For forecasting by using more general and sophisticated time series models, consider using the UCM procedure, which fits and forecasts unobserved components models that are not available in PROC FORECAST. Using UCM models, you can fit and forecast much more complex data patterns than you can by using the simple methods that PROC FORECAST provides. Unlike PROC FORECAST, the UCM procedure can also model and forecast the effect of independent predictor variables. For information about PROC UCM, see Chapter 34: The UCM Procedure.

For forecasting by using ARIMA models and the Box-Jenkins methodology, consider using the ARIMA procedure. PROC ARIMA identifies, fits, and forecasts general autoregressive integrated moving average models, optionally incorporating transfer function models for the effects of independent predictor variables. (As a special case, you can use seasonal ARMA models for forecasting seasonal series for which the Winters and additive Winters methods might be used.) PROC ARIMA also provides features for automatically identifying the specific ARIMA model that is most appropriate for the data. ARIMA and ARIMAX models are not available in PROC FORECAST. For information about PROC ARIMA, see Chapter 7: The ARIMA Procedure.

For forecasting multivariate time series, where two or more related variables need to be forecast jointly, consider using the VARMAX procedure or the SSM procedure. PROC VARMAX fits and forecasts vector autoregressive moving average models, optionally incorporating multivariate transfer function models for the effects of independent predictor variables. For information about PROC VARMAX, see Chapter 35: The VARMAX Procedure. PROC SSM fits and forecasts general linear state space models. The general state space model encompasses most of the other forecasting models that are mentioned in this section, and it enables generalizations that can model time series data patterns of almost any type and complexity. For information about PROC SSM, see Chapter 27: The SSM Procedure.

For forecasting both the future expectation and future volatility or risk, consider using the AUTOREG procedure or the VARMAX procedure. PROC AUTOREG can fit and forecast many types of GARCH models of time-varying volatility, while also fitting and forecasting future expected values of the dependent variable. PROC VARMAX supports multivariate GARCH models. For information about PROC AUTOREG, see Chapter 8: The AUTOREG Procedure. For information about PROC VARMAX, see Chapter 35: The VARMAX Procedure.

The stepwise autoregressive method is used by default. This method combines time trend regression with an autoregressive model and uses a stepwise method to select the lags to use for the autoregressive process.

The exponential smoothing method produces a time trend forecast. However, in fitting the trend, the parameters are allowed to change gradually over time, and earlier observations are given exponentially declining weights. Single, double, and triple exponential smoothing are supported, depending on whether no trend, linear trend, or quadratic trend, respectively, is specified. Holt two-parameter linear exponential smoothing is supported as a special case of the Holt-Winters method without seasons.

The Winters method (also called Holt-Winters) combines a time trend with multiplicative seasonal factors to account for regular seasonal fluctuations in a series. Like the exponential smoothing method, the Winters method allows the parameters to change gradually over time, with earlier observations given exponentially declining weights. You can also specify the additive version of the Winters method, which uses additive instead of multiplicative seasonal factors. When seasonal factors are omitted, the Winters method reduces to the Holt two-parameter version of double exponential smoothing.