It seems to me that it is possible that a forecasting model does very well on one step ahead forecasts (or on any other point forecast) but performs poorly on multistep forecasts (if you average the error for all forecasts on the horizon). For example, a simple exponential smoothing model might do well on the next step ahead forecast because of the local behavior of the time series, but will perform very poorly on a multistep forecast horizon because it produces a flat forecast that ignores any trend or seasonality.

Conversely is seems possible that a model can have better generalizing properties and perform better on average on a multistep forecast horizon, but at the cost of increased error on the first step or the first few steps.

Yet all the forecasting algorithms I've seen use one step ahead point forecasts to estimate goodness of fit.

Am I missing something? Does better one step ahead accuracy always imply better multi-step accuracy?

Or is using one step ahead forecasts for goodness of fit simply a pragmatic solution because trying to optimize for multistep ahead forecasts is not feasible or practical?

| cite | improve this question | | | | |

For real world problems, I've found fitting a unique equation and/or coefficients for each forecast step ahead (direct forecasting) produces the most accurate prediction. It tends to limit the propagation of errors. Rob Hyndman discusses combining recursive and direct forecasting in the following link:

Combining Recursive and Direct Forecasting

| cite | improve this answer | | | | |

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.