Future of work

This study highlights an advance in the methodology of trade policy analysis by demonstrating how modern machine learning techniques can complement traditional economic modelling to understand complex policy impacts.  

Using Canadian trade data from 1997 to 2022, researchers from the University of Guelph applied three ensemble learning algorithms, which typically yield robust predictions, to examine how tariff changes affect the import values of Canadian merchandise (based on the monthly USD value). 

This new methodology addresses a key limitation of using traditional econometrics for analysis. Standard linear regression models assume a simple linear relationship between variables—but real-world trade patterns involve complex interactions between currency fluctuations, in global supply chain frictions, and policy changes. 

The researchers capture these complexities with three machine learning approaches: 

• gradient boosting trees to correct prediction errors 

• stochastic gradient boosting to randomly subsample data to prevent models from ignoring new data 

• random forests to average the predictions from multiple decision trees 

This research used a critical innovation called Shapley Additive Explanations (SHAP) to make the models’ insights more interpretable. Where traditional AI models are often criticized for being black boxes that provide predictions without explanations, SHAP solves this by explaining the contributions of individual factors involved in each prediction. This allows policy-makers to understand what a model predicts, why it makes those predictions, and the marginal effect of each explanatory variable, increasing overall transparency and accountability. 

The researchers’ models on tariff elasticity demonstrate the value of this transparency. Across all three models, the report shows that exports emerged as the strongest predictor of import volumes (accounting for more than 70% of explanatory power), followed by tariffs (approximately 22%), and exchange rates (less than 5%). The tariff elasticity estimates showed that a 1% increase in tariffs corresponds to a 0.18 to 0.23% decrease in imports—smaller than traditional estimates, but still economically meaningful, given Canada’s low average tariff rates. 

The decision tree analysis also revealed significant effects that linear models would miss. For instance, the AI models identified specific tariff levels (around 0.34% in log terms) where the impact on trade flows changes significantly. This suggests that even small tariff adjustments can have disproportionate effects when they cross certain thresholds. This sort of information is crucial for trade negotiators and policy-makers. 

This study demonstrates the value that certain AI models can offer to governments, consultants, and private sector entities seeking a greater understanding of the impacts that tariff adjustments will have on Canadian industries. The researchers also note that their approach provides a replicable framework for evidence-based decision-making that can be applied to other policy areas, including labour market interventions and environmental regulations.