2
$\begingroup$

ALL of the examples go through a lot of explanation then only forecast one data point. I need to forecast more than on data point, but I can't find an example anywhere to do that.

If you use the example as in the books you end up reverting to the last value and getting the same number over and over again.

Can someone point me in the right direction of what I need to learn to do what I want ?

Just made up a series:

1.00000 2.00000 1.00000 3.00000 4.00000 1.00000 1.00000 2.00000 6.00000 4.00000 4.00000 5.00000 5.00000 4.50000 4.50000 4.75000 4.75000 4.62500 4.62500 4.68750 4.68750 4.65625 4.65625 4.67188 4.67188

Using alpha = .5 you see how the extrapolation always becomes 4.6 ? Surely that is avoidable ?

Improve this question
$\endgroup$
3
  • 1
    $\begingroup$ What book are you using? Are you doing this on paper? Are you using a statistical suite or programming language? Can you show us how you generate your prediction and the data used? $\endgroup$
    – Gullydwarf
    Commented Nov 14, 2014 at 12:24
  • $\begingroup$ Just made up a series :1.00000 2.00000 1.00000 3.00000 4.00000 1.00000 1.00000 2.00000 6.00000 4.00000 4.00000 5.00000 5.00000 4.50000 4.50000 4.75000 4.75000 4.62500 4.62500 4.68750 4.68750 4.65625 4.65625 4.67188 4.67188 Using alpha = .5 you see how the extrapolation always becomes 4.6 ? Surely that is avoidable ? (sorry man I'm having trouble using this site !) $\endgroup$
    – user3528592
    Commented Nov 14, 2014 at 12:33
  • 1
    $\begingroup$ I edited your question to add the information you provided. One thing that is not clear yet is where the data ends and the prediction starts. Also, how do you compute this? $\endgroup$
    – Gullydwarf
    Commented Nov 14, 2014 at 12:39

1 Answer 1

Reset to default
4
$\begingroup$

If you're using simple exponential smoothing then you should expect forecasts to stick at a constant level, though as you go further ahead they become less accurate. (The one-step-ahead forecast at $t$ is the level component at $t+1$ calculated from the observation & level component at $t$; why would the level component at $t+2$ be different without another observation at $t+1$ to take into account? It's not: with SES the forecast for subsequent periods is the same as the one-step-ahead forecast.) If you want them to trend then use an exponential smoothing method with a trend as well as a level component; if you want to take seasonality into account add seasonal components. Note that a forecast isn't a simulation of what the series might look like in the future.

Improve this answer
$\endgroup$
6
  • $\begingroup$ What do you mean by your last sentence, exactly? $\endgroup$
    – mugen
    Commented Nov 14, 2014 at 13:54
  • $\begingroup$ @mugen: Sometimes people new to forecasting tack a forecast for several periods on to the end of a time series plot of past observations and are then surprised that it looks so much smoother. $\endgroup$
    – Scortchi ♦
    Commented Nov 14, 2014 at 14:27
  • 1
    $\begingroup$ you'll always forecast a mean (smoothed) value while in simulations you'd get any randomly drawn value for the process that'd account for volatility... makes perfect sense now, thanks! $\endgroup$
    – mugen
    Commented Nov 14, 2014 at 14:38
  • $\begingroup$ Thanks so much for these answers ! Really helped clear things up. $\endgroup$
    – user3528592
    Commented Nov 14, 2014 at 21:46
  • $\begingroup$ Do you know of a good book to learn about forecasting ? $\endgroup$
    – user3528592
    Commented Nov 15, 2014 at 1:06

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.