1
$\begingroup$

I've been reading up on time series forecasting, but most of what I've come across assumes that we want to predict the next value in a series. As such, I have access to lagged values (maybe I look at t-4, t-3, ..., t when I predict the value at t+1).

Suppose I want to predict the next two weeks worth of values. Now I run into trouble, because I can make my first prediction using lagged values, but my second prediction (t+2) will be missing the true value for t+1. I'll have my prediction for t+1, but that seems far inferior to the using the ground truth.

Are there standard ways this is handled in?

Improve this question
$\endgroup$
2
  • 1
    $\begingroup$ software takes the prediction for the next period and uses it as an estimate of the actual to predict the subsequent period and so on ..... $\endgroup$
    – IrishStat
    Commented Mar 7, 2018 at 17:09
  • 1
    $\begingroup$ @IrishStat This is only correct when the model is linear. In general the conditional expectation must be computed. $\endgroup$
    – Chris Haug
    Commented Mar 7, 2018 at 18:16

1 Answer 1

Reset to default
1
$\begingroup$

'Recursive forecasting' uses your t+1 forecast value in its estimate of t+2. 'Direct forecasting' solves for a unique equation/set of coefficients for each future time period: t+1, t+2 etc. Rob Hyndman discusses this more in depth in the following link, and I personally have found direct forecasting to perform better long-term as it results in less propagation of errors.

Recursive and Direct Forecasting

Improve this answer
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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