I need some hint over what is the general prediction solution to modelling products prices in such a case:

  • I have several models (types) of the product
  • I want to predict prices for each of these models separatedly, for a few weeks long time window
  • I have great historical data for each of the model from last 1-2y (which makes time series analysis my first obvious choice, doesn't it?)
  • I know there is a number of external factors which may have a VERY important and immediate effect on the product price (possibly more or less different for each model);

My questions are:

  1. Is time series prediction a good way to follow? Maybe some kind of regression would be more approproate in this case?
  2. Which approach to choose to include these external factors effects in the best way?

The preferred approach is to construct a Transfer Function where Price is the dependent variable and your suggested variables are the predictors. This is also called a dynamic regression. Care should be taken to identify and deal with unusual values that are either one-time events or level/trend shifts. Oftentimes there can be a lead effect on the dependent variable such as holidays. If you have daily data then day-of-the=week patterns and/or weekly patterns my be important. Perhaps you could post your data and I can be of more specific help.

  • $\begingroup$ IrishStat, thank you for you attention! I have daily product price data. The other important variables I mentioned above are of both types: 'singular values' (e.g. feature of the product of) and 'multiple values' by which I mean numeric values of one measure, measured let's say quarterly. Main goal of this analysis is to predict trend of the price (will it go up or down?) and the next step is to evaluate the level of this up/down movement. $\endgroup$ – Marta Karas Sep 9 '14 at 8:34
  • $\begingroup$ Thus I am not sure whether regression models are appropriate to 1) simulate several future values (second step mentioned above) 2. include efficiently all the information without serious overfitting problems (how should I incorporate all past daily information in the regression variables?). I would also be grateful for some further references for the model you mentioned, of which you think that are worth reading. [I wish I could grant it with more bounty as acknowledge of your attention, unfortunalety I cannot.] $\endgroup$ – Marta Karas Sep 9 '14 at 8:34
  • $\begingroup$ product.half.ebay.com/… isa text that introduces concepts of multivariate Box-Jenkins. Care should be taken to validate the Gaussian Assumptions using the ideas of unc.edu/~jbhill/tsay.pdf among others.As I suggested please consider posting your data using dropbox.com/… or send it to me via my contact info $\endgroup$ – IrishStat Sep 9 '14 at 11:35
  • $\begingroup$ stats.stackexchange.com/questions/66825/… discusses daily data. Your problem is easily resolved by incorporating user-known predictors. Good software will automatically detect what lags are appropriate for each of your predictors. $\endgroup$ – IrishStat Sep 9 '14 at 18:38

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.