BOTTOM-UP MACHINE LEARNING FOR ECONOMIC ESTIMATION

This paper concerns the tradeoff between accuracy and precision in economic models when deploying teachable algorithms – i.e., machine learning (ML). It advocates lowest-level supervised learning first, to create fully informed equations as baseline estimators, whether standalone or as component models of a hierarchy. Bottom-up modeling focuses on accuracy from real data, statistics that are usable, and proper aggregation. Simple linear equations support bottom-up; they are often constrained by near-additive coefficients, so that between-model agreement, although never assured, is more likely; small models respond better to efforts to treat aggregation bias, to produce high quality coefficients or likelihoods. The lowest level is characterized by overall lowest estimation error. It is easier to identify, compare and correct error within, rather than between, bottom-up learning models. Because local samples represent current real-world conditions of an area, they suit economic problems best. Bottom-up estimation is the antidote to hasty summation and improper post-weighting in complex models.1 Random variates from small and local domains are priors, potential building blocks or components of a hierarchy; alone, they should not be treated as mere ancillary placeholders for the top-down allocation of biased aggregates