ThGaskin/OT_Trade

Modelling Trade with Optimal Transport

This repository contains all the datasets and trained neural networks used in our paper Modelling Global Trade with Optimal Transport. Evaluation and training code can be found in the Github repository. The data is sourced from FAO trade matrix dataset for the years 2000-2022 and pooled such that the countries constituting 99% of imports and exports are listed, and all others subsumed in an 'Other category'. The pooled FAO datasets are contained the data/FAO_data folder, alongside yield and total production datasets.

Our own estimates are collected in seperate folders for each commodity. These contain:

• The torch.Tensor objects used to train the neural networks: training_mu.pt and training_nu.pt are the two marginals; training_T.pt is the FAO transport plan
• FAO_data_mask.nc: the masked FAO values (NaN values);
• C_mask.pt: the mask applied to the cost matrix estimates: values where this is true are set to 1 to allow for unique inferrability
• samples_stats.nc: the estimates of the transport plans and cost matrices, obtained by sampling matrices from the FAO dataset (randomly selecting importers and exporters) and pushing these through the trained neural networks to obtain a cost matrix sample
• trained_models: an ensemble of 10 trained neural networks. Each contains the network parameters and configuration file needed to load the trained network as a PyTorch object
• gravity_estimates: gravity coefficients for each dataset; the gravity models are described in the paper

In addition, the data folder contains a number of additional datasets used in the Analysis.