Beyond the Streetlight

This REMARK contains Christopher Carroll’s discussion, at the Federal Reserve Board’s R&S Centennial Conference on November 6-8, 2023, of the paper by Carol Corrado and Arthur Kennickell, together with the code and data behind the discussion’s figures. The paper’s final version, with Tomaz Cajner as a third author, is Corrado, Kennickell and Cajner (2025), “Beyond the Streetlight: Economic Measurement in the Division of Research and Statistics at the Federal Reserve”, Finance and Economics Discussion Series 2025-019, https://doi.org/10.17016/FEDS.2025.019. The conference draft was titled “100 years of Economic Measurement in the Division of Research & Statistics: Beyond the Streetlight”.

The argument

The discussion treats measurement as information production. Beyond primary data (such as the Survey of Consumer Finances) and secondary data (such as industrial production), the Fed produces “tertiary” information: the staff forecasts in the Greenbook and Tealbook, staff memos, and similar work. The Greenspan Fed’s call of a “new economy” in the mid-1990s is the leading example of that information being used well. The discussion asks whether the Fed was just lucky then, or whether forecasting has improved over time, at the Fed and elsewhere, as better measurement would imply.

What the code does

The analysis compares two sets of forecasts with what actually happened:

• Greenbook/Tealbook (GB): the Federal Reserve Board staff forecasts, from the Federal Reserve Bank of Philadelphia’s Greenbook data set;
• Survey of Professional Forecasters (SPF): the mean private-sector forecast, also from the Philadelphia Fed;
• Realized values: the unemployment rate (FRED series UNRATE) and real personal consumption expenditures (DPCERA3Q086SBEA).

For each quarter it builds a year-ahead forecast of the change in unemployment and of real consumption growth, computes the absolute forecast error, and regresses the error on a time trend with heteroskedasticity-robust (HC3) standard errors, over the full sample (144 quarters, labeled 1983Q1-2018Q4) and from 1995 on.

Findings

Errors in forecasts of real consumption growth have fallen significantly, for both forecasters. Over the full sample the trend is -0.039 percentage points a year for the Greenbook and -0.031 for the SPF (both p < 0.001), and it remains negative after 1995. Errors in forecasts of the change in unemployment show no significant trend. The two figures in the slides show the consumption results.

Reproducing the results

./reproduce.sh regenerates every table and figure from the raw data, checks them against the committed versions, and renders the slides; it takes about 30 seconds once the environment (pinned in uv.lock) is installed. It needs no network access or API key: the realized values come from a committed snapshot of the FRED data used for the published results. See README.md for Docker, conda and Binder instructions and for a description of every output.