MDCEV Forecasting alpha-profile

[Anna_R]

Hi,

I am trying to implement MDCEV forecasting into my activity model and I've been using the forecasting algorithm proposed by Pinjari and Bhat (2011). I don't know if I'm missing something, but the formulation in that paper does not allow for forecasting using an alpha-profile, i.e. where alpha values are different for each alternative.

Now Apollo lets us define an alpha-profile and forecast based on the estimated parameters. How did you circumvent the issue of the forecasting issue?

I did take a look at what's implemented in apollo_mdcev (I am looking at the outside good case):

Code: Select all

while (!stopping) {
                use = orderofV < M
                lambda_1 = b + sum(p * g * use)
                lambda_21 = ph0^(1/(1 - a0))
                lambda_22 = sum(p * g * use * phk^(1/(1 - a0)))
                lambda_2 = lambda_21 + lambda_22
                lambda = (lambda_1/lambda_2)^(a0 - 1)
                if (M > sum(phk > lambda) || M > sum(phk > 0)) {
                  x0_1 = lambda_21 * lambda_1
                  Xintra[i, 1] = Xintra[i, 1] + x0_1/lambda_2
                  xk_1 = phk^(1/(1 - ak)) * lambda_1
                  Xintra[i, 2:s$nAlt] = Xintra[i, 2:s$nAlt] +  use * (xk_1/lambda_2 - 1) * g
                  stopping = TRUE
                }

It looks like alpha_0 is used when calcluating lambda, and alpha_k are then only used once the consumption is calculated. Do you have a paper that provides more background on the reasoning behind this?

Thank you!
Anna

[stephanehess]

Anna

apologies for the slow reply, your post got lost somehow.

This is very odd as Apollo actually checks for this and does not make predictions when alpha is not generic. What version of Apollo are you using?

Best wishes

Stephane