Research

I work on monetary policy and firms, mostly on what a rate shock does to investment and R&D. The job market paper below is the core of it. A longer account of the agenda is in my research statement.

Job market paper

2026

Monetary Policy and Firm Innovation

Job market paper · JEL E52, O31, O32, O34

This paper studies how monetary policy shapes firm-level innovation, using linked Compustat, USPTO and CRSP data with local projection methods. The R&D response to monetary policy operates almost entirely through the intensive margin: expansionary shocks raise R&D and knowledge capital among firms already engaged in innovation, while other firms remain largely unresponsive. On the output side, expansionary shocks lower the aggregate number of patents but raise average citations and, over the medium run, patent importance and the share of patents in the upper tail of the importance distribution, consistent with firms reallocating R&D from incremental patenting toward higher-impact projects as financing conditions ease. A two-stage analysis indicates that this reallocation operates through the R&D response itself rather than through a direct effect of the shock on patenting. The effects are also asymmetric in sign: contractionary shocks reduce both the quantity and quality of innovation, and this reallocation-toward-quality channel is cleanest among economically large shocks. Monetary policy therefore changes the direction of firms’ innovation as well as its scale.

Impulse response of average citations per patent to a monetary policy loosening over twenty quarters, rising from zero to about five, with confidence bands
Average citations per patent after a 100 basis point loosening, by quarter since the shock. Citations barely move in the first year and rise after that.

Data

The panel joins five sources. CRSP-Compustat gives quarterly financials from 1990 to 2020; USPTO filings give patents granted from 1990 to 2017, with citations, the Kogan, Papanikolaou, Seru and Stoffman market-value measure and text-based importance; and the monetary policy shocks are the high-frequency series used by Ottonello and Winberry, 164 shocks from January 1990 to December 2007. Patent assignees and firm identifiers do not line up cleanly across three decades of mergers and delistings, so a large part of the work was the matching and the change log that lets any estimate be rebuilt from raw files. Knowledge capital is built from R&D by the perpetual inventory method, following Peters and Taylor. The Great Recession is left out of the estimation sample because high-frequency identification stops working at the lower bound.

Method

Local projections trace the response of R&D, knowledge capital and patent outcomes to a shock over twenty quarters. Probit and Tobit models separate the extensive margin, whether a firm starts doing R&D, from the intensive margin, how much an active firm spends. A two-stage design then asks whether the patent response runs through the change in R&D: the first stage fits R&D on the shock, the second regresses patent outcomes on the fitted R&D. Interacting shock sign with shock size separates asymmetry from magnitude.

What the paper finds

  • Firms that already do R&D spend more after a loosening. For firms that do none, the probability of starting is indistinguishable from zero at every horizon.
  • Patent counts fall in the short run. Average citations and importance rise over the medium run, and so does the share of patents in the upper tail of importance.
  • The patent response runs through R&D. Once fitted R&D is in the regression, the shock itself adds little.
  • A loosening lowers patent counts and raises quality, but a tightening lowers both. The move toward quality is clearest for economically large shocks.
  • Citations and text-based importance rise after a loosening. Per-patent market value does not move reliably.

Why it matters

Work on monetary policy and heterogeneous firms sorts the investment response by leverage, liquidity, age and size. For innovation the split that matters most is simpler: whether a firm does R&D at all. The paper also gives work on monetary policy and growth a firm-level channel, from the rate shock through R&D spending to the kind of patents firms produce. For innovation research, the disagreement between quality measures means a study’s answer can depend on which proxy it uses.

Line chart of average R&D spending and the physical investment rate for US firms from 1990 to 2017, both rising, with R&D rising more steadily
Average R&D spending (solid, left axis) and the physical investment rate (dashed, right axis) for Compustat firms, 1990 to 2017. R&D grows through every downturn, which is part of why its response to policy looks different from physical investment.
Compustat USPTO CRSP Local projections Probit and Tobit Stata

Work in progress

Two papers with coauthors look at other kinds of firm investment, in AI and in physical capital.

Draft

Artificial Intelligence and Inventory Management: Evidence from U.S. Retail Industry

With Shuang Chen, Min Fang, Boyuan Li and Baolian Wang · draft available

Firms hold inventory, and working capital more generally, because they cannot forecast demand and supply exactly. AI lowers the cost of forecasting (Agrawal, Gans and Goldfarb, 2019), so we ask two things about retailers. Does AI change how they manage working capital, and does it help them when a supply chain breaks?

Firms do not report where they use AI, so we measure it from the near-universe of U.S. online job postings: Lightcast data from more than 40,000 sources, 2010 to 2023. We flag AI vacancies by their machine-learning skills and sort them into marketing and supply-chain jobs with a tiered keyword taxonomy that favors precision over recall. The measures are linked to Compustat for more than 1,300 public firms.

In long differences over 2010 to 2023, rising AI adoption goes with a shorter cash conversion cycle for retailers, by about a week of receivables per standard deviation of supply-chain AI, and with essentially no change for other firms. The shortening comes from receivables. Days receivables outstanding falls for all seven AI measures we build, days inventory outstanding does not move in any specification, and days payables falls for two of the seven. Listed retailers had already squeezed most of the slack out of inventory before 2010, so the customer side is where AI still had room to work.

The second test treats the 2018 U.S.-China tariff announcements as a supply-chain shock, following the event-study design of Huang et al. (2023). We interact pre-shock China-import exposure, retail status and 2017 AI intensity, with buy-and-hold abnormal returns over 20 to 80 trading days as the outcome. China-exposed retailers with more supply-chain AI lost significantly less market value. The effect is zero on impact, shows up by about 40 trading days and keeps growing through 80, and it does not appear for marketing AI. Without the import-exposure control the sign flips, because AI-intensive retailers are also the larger and more import-reliant firms that the tariffs exposed.

AI adoption Working capital Job postings Event study
Ongoing

Solving the Macro and Micro Investment Elasticity Inconsistency with Incomplete Information

With Jonathan Adams, Min Fang and Eugenio Rojas · draft in progress

Heterogeneous-firm investment models cannot match two elasticities at once. One is the macro elasticity, how much aggregate investment responds to the real interest rate. The other is the micro elasticity, how firms move between the extensive and intensive margins of investment. Fixed costs small enough to fit the micro evidence (Cooper and Haltiwanger, 2006; Khan and Thomas, 2008) leave aggregate investment far too sensitive to interest rates, and fixed costs large enough to fit the macro elasticity (Winberry, 2021; Koby and Wolf, 2020) distort the micro margins.

We think fixed costs alone cannot fit both, so we add incomplete information. Firms in our model pay fixed, partially irreversible and quadratic adjustment costs, and each sees only a noisy signal of its own productivity. I solve the model by value function iteration in Julia and compute the stationary distribution with the non-stochastic method of Young (2010), so the moments carry no simulation error. Starting from the Cooper-Haltiwanger and Khan-Thomas calibrations, we change only the volatility of the information noise. More noise dampens both elasticities and narrows the gap between them, and in some calibrations it closes the gap.

Firm dynamics Structural model Julia

Next projects

Next

A model of the quality margin

The main empirical result of the job market paper is that easier financing moves research from incremental toward high-impact projects. I plan a heterogeneous-firm model in which firms choose between project types under financial frictions, and easier financing lowers the cost of long-horizon, higher-risk projects. With it I can ask how much of monetary policy’s long-run effect on productivity comes from the direction of research and how much from its volume.

Next

Why patent quality measures disagree

After a monetary shock, citation-based and market-value measures of patent quality move differently. I want to work out when they diverge and what each one measures, since the answer matters for any study that uses patents to measure the returns to innovation policy.

Next

Technology adoption and financial conditions

With the job-posting data linked to firm financial records, I can test whether credit conditions change how fast firms adopt technologies such as AI. That would tie the AI paper to the investment results in the job market paper.