0
$\begingroup$

I'm using Holt Winters to predict sales revenue from past performance. Seasonality and changing trends exist in the data.

One of the reasons chosen for Holt Winters is that it is fairly simple (implementable in excel) and explainable to non-statisticians. If other methodologies are more appropriate I’m happy to hear about them.

A key implementation detail is to choose coefficients $\alpha, \beta$, and $\gamma$. Are there “standard” values for these?

The other reason for this methodology is that it auto-updates. This will be useful as the series will change with market shifts. I was planning to adjust the “standard” value by eye, to try to mitigate likely changes. Is this a reasonable approach?

$\endgroup$
3
$\begingroup$

You should not "choose" coefficients, you should estimate them from the data. The values of $\alpha$, $\beta$ and $\gamma$ control how quickly the model components change over time. What values are appropriate depend on the data. You can estimate them by minimizing the MSE, for example. Even in Excel you can do that using Solver.

$\endgroup$
  • $\begingroup$ Thanks excellent help, would you minimise the MSE on out of time data or over the entire series? $\endgroup$ – pipie314 Nov 3 '11 at 10:00
  • $\begingroup$ On the entire series $\endgroup$ – Rob Hyndman Nov 3 '11 at 22:57
  • $\begingroup$ @RobHyndman Assume we are given a time series for training, where the ts spans 2 time periods. We usually use this training set to bootstrap the estimation of level, trend, and seasonality. How can we use this training set for estimation of α, β and γ instead? Because to calculate MSE, we still need to be able to estimate the L,T, and S first. I hope I am making sense. $\endgroup$ – sandyp Aug 18 '16 at 23:22
  • $\begingroup$ Use a nonlinear optimizer. $\endgroup$ – Rob Hyndman Aug 18 '16 at 23:35
0
$\begingroup$

You asked "If other methodologies are more appropriate I’m happy to here about them." A generalized ARIMA can be easily expressed as a lagged auto-regression ADL or PDL . THis model easily adpats to changes in levels, trends , parameters , seasonal pulses , variability. You are assuming a structure rather than allowing the data to suggest the structure.

$\endgroup$

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.