Research
Job Market Paper
When the News Matters: Newspaper Text and Real-Time Forecasts of U.S. Growth
Latest draft (PDF) · Slides (PDF)
Does newspaper text help to forecast the U.S. economy in real time, and when? I build monthly text series, topic shares, policy-uncertainty counts and sentiment, from 11.76 million articles of seven U.S. newspapers over 1990–2020, and race them against 101 FRED-MD series in unrestricted MIDAS regressions estimated with four machine-learning methods on expanding real-time windows, scored as RMSE relative to an autoregression over 2003Q1–2020Q1. First, on all articles text alone is behind the hard data for GDP at the nowcast and for investment, and adding it to the hard data with the method held fixed gives 14 significant improvements and 7 deteriorations in 144 comparisons, about 7 of each being expected by chance; the improvements are for consumption. Second, which articles are used matters: an automatic filter for the economic articles, checked against hand labels, improves the text forecasts in 27 of 144 comparisons and worsens them in 4, in normal times as well as in the recession, under every selection rule tried. Third, the gain over the benchmark is a matter of timing: four quarters of 2008–2009 hold between a third and a half of the benchmark’s squared error; in those quarters text beats the benchmark by 10–13% for GDP and 20–35% for consumption, and outside them neither text nor hard data beat it. The same-method gain from adding text survives scoring against first releases; the gain over the benchmark at the nowcast does not.
Working Papers
Pricing a Good That Costs Nothing to Copy: Discount Dynamics and Demand on Steam
with Arseniy Stolyarov
Slides (PDF)
A released video game costs nothing to copy, is bought only once and cannot be resold, so its price serves only to sort buyers over time. Using 5.6 million posted-price changes on the US Steam store since 2020 and more than 150 million dated user reviews, we document how sellers set that price and how demand responds. The average game is discounted about three times a year but changes its list price only about once a decade. Discount timing follows the platform’s sale calendar and cooldown rules, and discounts deepen as games age. We measure demand with reviews and player counts, and it rises by similar proportions at the start of shallow and deep discounts. This suggests that a discount works partly as a notification to consumers waiting for a sale. We are developing a dynamic model of a seller facing a stock of waiting consumers to evaluate the platform’s sales calendar and discount rules.
Instruments of Change: The Rise of Identification Designs in Economics
Slides (PDF)
I document how the identification designs of the credibility revolution spread across the economics profession. Using full texts of roughly 114,000 articles from about 200 economics journals published between 1990 and 2024, I build a phrase-based classifier for seven designs (difference-in-differences, regression discontinuity, instrumental variables, randomized experiments, event studies, synthetic control, and structural estimation), validate it on 589 hand-labelled papers, and correct the resulting series for measurement error and for the uneven availability of full text across years and publishers. The share of articles using at least one of these designs rises from 17.5% in the 1990s to 42.6% in the 2020s, a 2.4-fold increase that is only partly driven by longer papers. A pre-specified comparison across the journal hierarchy shows that observational designs—difference-in-differences, regression discontinuity, and instrumental variables—converged on the rest of the profession about a third faster per decade than resource-intensive designs such as randomized experiments and structural estimation, a pattern that survives dropping any single design and reweighting on research topic. Exploratory evidence from co-author and colleague networks shows that design choices cluster within these ties, but the clustering is not distinguishable from shared taste.