Regensburg 2016 – scientific programme
Parts | Days | Selection | Search | Updates | Downloads | Help
SOE: Fachverband Physik sozio-ökonomischer Systeme
SOE 15: Economic models II
SOE 15.3: Talk
Wednesday, March 9, 2016, 12:45–13:00, H36
Evaluating Multilevel Predictions from Trading Data — •Sven Banisch1, Robin Lamarche-Perrin2, and Eckehard Olbrich1 — 1Max Planck Institute for Mathematics in the Sciences, Inselstrasse 22, D-04103 Leipzig, Germany. — 2Laboratoire d'Informatique de Paris 6, UPMC - Sorbonne Universités, Paris, France.
If one wants to predict the behaviour of a complex, multi-level system such as the economy one can use observables on different levels of aggregation. Naively, one might think that going to a higher aggregation level always deteriorates performance because it means losing information and therefore using a microscopic description is the best. However, this is usually not the case since no complete microscopic model is available in most applications. Instead, the predictor has to be inferred from data which may become practically infeasible due to high-dimensional microscopic state spaces and exponentially increasing data requirements. We study the trade-off between the higher information content of less aggregated descriptions and the better inferrability of higher-level aggregates on real world data on international trade. We compare different predictors for GDP growth considering aggregations over meaningful groups of products representing mesoscopic levels of the export structure and highly aggregated measures of economic complexity (Hidalgo/Hausmann 2009) and fitness (Tacchella et al. 2012) previously shown to have predictive power regarding the growth potential of countries. We present evidence that mesoscopic observables may outperform these highly-aggregated measures while still allowing proper inference of the predictor from the limited amount of data.