Payments-Data Nowcasting — the Bank of Canada's Machine-Learning System for Real-Time Macroeconomic Policy Analysis

Official Canadian GDP data lag roughly eight weeks and CPI data roughly two weeks — a serious constraint when COVID-19 hit and policymakers needed a near-real-time read on the economy.

Bank of Canada researchers James Chapman and Ajit Desai aimed to nowcast GDP, retail trade and wholesale trade using payments data and machine learning to close this data-lag gap.

They combined 18 series of values and volumes from Canada's high-value (Lynx) and retail (ACSS) payment settlement systems, spanning March 2004-December 2020, with five ML methods (elastic net, support vector machines, random forest, gradient boosting regression and neural networks), training on March 2005-December 2018 data and testing on January 2019-December 2020, a window spanning the pandemic's onset.

A factor model using payments data cut nowcasting RMSE by 20-25% relative to a traditional-indicator benchmark at a one-period-ahead horizon; adding gradient boosting regression increased the reduction to 35-40%. Gains were larger during the COVID-19 crisis window (up to 35%) than in the normal period (15-25%). The original 2021 staff working paper summarised the headline finding as a 15-45% reduction in forecast error versus a benchmark linear model.

A May 2026 Bank of Canada staff analytical paper describes this line of work evolving into an ongoing institutional practice using ML, NLP and generative AI on non-traditional data for policy analysis, alongside a governance framework for trust, transparency, reproducibility and model risk. No source confirms the nowcasts have directly changed a specific interest-rate or programme decision, as distinct from serving as one monitoring input among many.