Model combination for capturing the inconsistency in the aggregate prediction

• What is the appropriate aggregation level for modeling when the purpose of modelingis aggregate prediction: is it to estimate a disaggregate model and aggregateindividual predictions or estimate an aggregate model for the aggregate prediction?There is no unique answer to this old question in the literature as well as no generalmethodology to address the problem. In this paper, we propose to tackle theaggregation problem by employing and developing model combination methods tocombine aggregate and disaggregate models. Dierent aspects of aggregation arecovered in this paper: aggregation over time, individuals and alternatives. We examinethe eect of aggregation on the prediction accuracy of a nested multinomiallogit (NMNL). The application of interest is to predict the monthly share of cleancars in the Swedish car eet. We investigate a situation wherein the large scalemodels are already estimated, and we are interested in improving their predictionperformance in a post-processing manner. We combine NMNL with a regressiontree to capture individual heterogeneity as well as a time-series model to capturedynamics of the market share of clean cars at the aggregate level. Models are combinedthrough a latent variable model and a Bayesian model averaging approach.We propose aggregate likelihood as the likelihood to be maximized for the modelselection and combination when the purpose of modeling is aggregate prediction.The results show the increase in the predictive power of combined models.

Nyckelord

model combination

aggregation

logit models

prediction

latent variable models

Bayesian model averaging

nite mixture model

Publikations- och innehållstyp

vet (ämneskategori)

ovr (ämneskategori)