Deploying data science for national economic statistics

Jeff Chen shares strategies for overcoming time series challenges at the intersection of macroeconomics and data science, drawing from machine learning research conducted at the Bureau of Economic Analysis aimed at improving its flagship product the gross domestic product.

Time series analysis sits at the heart of macroeconomics. It provides the trends and context of where the economy is moving, but the nature of the game is changing. Some data that feeds indicators may only be available after the preliminary release of macroeconomic numbers, necessitating the use of short-term prediction to plug gaps. In recent memory, modelers have turned to more timely alternative data and other sources that sit outside of macroeconomic theory to help boost the signal of the economy. This, in turn, raises both technical and theoretical challenges. By increasing dimensionality using modern data sources, the time series shrinks. Furthermore, input features may not be drawn directly to establish macroeconomic frameworks. Jeff Chen shares strategies for overcoming these time series challenges at the intersection of macroeconomics and data science, drawing from machine learning research conducted at the Bureau of Economic Analysis aimed at improving its flagship product the gross domestic product. Jeff presents this work in the context of a comprehensive horse race of algorithms, data sources, and feature selection methods that have yielded best practices for the field. He offers a number of practical tips for how social scientists can incorporate time series prediction into their work, covering core issues that underlie the success and failures in this space, namely how basic assumptions of algorithms affect their ability to detect regime change as well as the need to distinguish accuracy gains from each data and models.