Suppose we have sales of 3 products A, B and C which market share always sum up to 100%. How to model market share of product A using market shares of B and C? So that we will know that if share of A grows 1 percentage point then A has to steal that growth in some proportion form products B and C.

A means increase of share = A on Tuesday - A on Monday, and so on with B, C.

A = 0*A + b1*B + c1*C
B = a2*A + 0*B + c2*C
C = a3*A + b3*B + 0*C

How to estimate it? Is there any other method how to estimate structure which sums to 100% by the parts of the structure? Regression doesn't work here.

  • $\begingroup$ It seems you've made an edit without being logged in, thus it is showed as an anonymous one. Next time, please check it out before sending -- this way it will be applied right away. $\endgroup$
    – user88
    Jul 7, 2012 at 21:37
  • $\begingroup$ To alternatives I can think of, 1) estimate a model for the sales of A,B,C and calculate the shares afterward. 2) estimate a model for the shares of A and B and then calculate the shares of C by definition. $\endgroup$
    – John
    Jul 7, 2012 at 21:45

2 Answers 2


You did not really describe what you mean with model share, i.e. what the purpose of the modelling is, but the hts R package and hierarchical time series provide a way to create forecasts of time series which sum up to 100%, etc. Prof. Hyndman has at least two publications on using them: Hierarchical forecasts for Australian domestic tourism and Optimal combination forecasts for hierarchical time series. Besides documentation of the package I don't know other sources about hierarchical time series than the papers and the references therein.

  • $\begingroup$ The purporse of the model is to give answer how much growth of A is gained from B and how much from C. A(B,C) = ? If we estimate model A = aB + bC without constant then both parameters a and b equal -1. But empiricaly I know that product A grew from 0% share to 30% share. It "stole" 75% of this 30% share growth from product B and 25% from product C. But I have no idea how to estimate it. The purpose of the model is explanation, not prediction. $\endgroup$ Jul 9, 2012 at 21:25

Apart from the hierarchical approach already mentioned, an alternative is to model the three series as compositional time series. See the following papers for recent surveys of the literature.

  • $\begingroup$ I have access only to abstract. What is the concept of the compositional and hierarchical approach to time series? $\endgroup$ Jul 11, 2012 at 11:10
  • $\begingroup$ The additive log-ratio transformation! Good to finally put a name to the thing $\endgroup$
    – Taylor
    Jul 15, 2022 at 3:22

Your Answer

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

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