Vera Chau

Technology, Investment & Growth

Working paper

Local Networks and Technology Adoption during Industry-Wide Technological Change*

Paper (PDF)Supplement (PDF)Web appendix (PDF)

AI tools were used to help edit the writing and to clean and implement code.

Four panels compare adoption and investment responses to peer performance and technique intensity across 2008–14, 2015–17, and 2018 onward.

Adoption and investment responses to local peer activity across three stages of the industry transition.

Figure notes

Rows distinguish peer performance from technique intensity. The left column measures adoption; the right measures horizontal well counts among firms already investing in horizontal drilling. Bars are coefficients on expiration-driven peer activity, lagged one quarter and interacted with the relevant local signal, estimated separately for 2008–2014, 2015–2017, and 2018 onward.

Whiskers show 95% confidence intervals using formation-block-clustered standard errors. Filled bars denote estimates significant at the 10% level; brackets report directly estimated differences between adjacent periods. Adoption becomes more responsive to both signals in 2015–2017, but subsequent responses differ by signal and by whether the outcome is starting or expanding horizontal activity. The panels use different vertical scales.

View full-size figure

I study how firms’ adoption responses to local network information change during the U.S. hydraulic-fracturing transition. I combine spatial panel models that characterize local network signals with a bunching estimator that isolates peer drilling as a plausibly exogenous source of new information. At the same local-development stage, the adoption response to peer activity is negative in markets that develop when the technology is young industry-wide and positive in markets that develop later. The cross-vintage difference is statistically significant. In later markets, the response weakens as local development accumulates. Early in the transition, the adoption response is negative but the investment response among investing firms is positive. A calibrated model uses this difference to distinguish returns to local evidence from starting costs as channels through which industry development affects firms’ use of local network information. The benchmark combines an early starting-cost response that declines with industry and local development with diminishing returns to local evidence.

Work in progress

What Does It Mean to Adopt a Technology?

Four panels compare firm types and horizontal output shares as industry learning and the cost of holding both technologies vary. Top-row shares are among firms surviving from 2007 to 2021.

Two forces in the model do different jobs.

Figure notes

Each column varies one model parameter. The top row splits the industry into three kinds of firm, as a share of every firm that survives from 2007 to 2021: non-adopters, switchers who let their vertical wells run down, and straddlers who keep drilling them. The bottom row shows how much of the industry's output comes from horizontal wells. Stronger industry learning, on the left, pulls firms into adoption, with non-adopters falling from 52 to 19 percent and switchers rising from 44 to 74, and carries production to the new technology, from 51 to 75 percent of output. A higher cost of running both technologies at once, on the right, mostly changes the form adoption takes, straddlers from 25 down to 4 percent and switchers from 26 up to 58, and moves production about half as far. Each point averages five simulations of the estimated model. These counterfactual results remain preliminary and have had no independent verification pass.

The figure is a counterfactual inside the estimated model, not observed data. Every point holds every other parameter at its estimated value and re-simulates the industry forward.

The base in the top row is firms that survive the whole 2007 to 2021 window, which is 46 to 55 percent of the simulated panel. The three groups are defined by comparing a firm's vertical capital at the end of the window with its own capital at the start, so a firm has to be there at both ends to be classified. A fourth group sits between the switcher and straddler cutoffs and is left out of the lines; it never exceeds 5 percent, and all four groups sum to 100.

Learning values below 0.06 are outside the box the solver was trained on, so they are not shown and no learning-off experiment appears here. The lowest learning value plotted, 0.06, is only about a third below the estimate, so the left column shows the direction and part of the magnitude, not an extreme.

The straddler line in the left column is not monotone. It rises with learning up to 0.12 and falls back after it. The figure plots that as it is.

The estimated model marked on the right column is the kappa of the current model, 0.47. An earlier estimation run of the same model gave 0.80; that run has been superseded and is not what any of these simulations use.

View full-size figure

New technology diffuses through the economy when firms incorporate it into production. How and why they do so are not well understood. Adoption can be a technological overhaul that replaces old technology or a diversification decision that expands the firm’s portfolio of productive capital without retiring older technology. I use the US shale boom as an empirical test case. Producers could adopt horizontal drilling and allow their existing vertical wells to run down or they could add horizontal wells while continuing to invest in vertical production.

I develop a heterogeneous firm model in which technology productivity, entry costs, and joint operating costs determine this choice. Switchers let vertical capital run down. Straddlers invest in both. A neural network conditioned on parameters solves firms’ dynamic problems. I estimate the model by simulated method of moments. The near zero estimated correlation means that conventional drilling productivity does not predict horizontal drilling productivity. Leaders under the old technology need not lead under the new. Horizontal drilling creates shared industry learning for all firms. Switchers generate two thirds of this learning. Lowering entry costs raises the share of adopters from 52 to 76 percent while horizontal drilling’s output share remains near 60 percent.

Work in progress

Technology Adoption Along a Moving Frontier

Three panels compare firm technology indexed to 100 in period zero under a full-path solution and a stationary AR(1) approximation, for slow, moderate, and rapid transitions.

Adoption paths from the operator and an estimated-AR(1) Bellman benchmark under three frontier transitions.

Figure notes

The three panels separate slow drift, a moderate S-shaped transition, and a rapid S-shaped transition. Solid blue lines show the full-path solution, which observes the entire transition; dashed orange lines show the stationary AR(1) approximation. Firm technology is indexed to 100 in period 0. The benchmark estimates an AR(1) process from observed frontier history, solves the stationary problem by value function iteration, and applies that policy to the realized path. The plotted paths are unchanged from the earlier exhibit; no model was rerun.

The Bellman benchmark produces 4% more lifetime firm value under slow drift. Under the two S-shaped transitions, it mistimes adoption and capital investment, and the operator produces about 9% more lifetime value. The full-path operator’s lower value under slow drift reflects approximation error in current training, not a structural advantage of recursion. These project results have not been independently verified.

View full-size figure

Many economic problems unfold along paths that are nonstationary, change regimes, or contain calendar breaks. Technological change, climate tipping risk, and dated policy changes are leading examples. Recursive methods can accommodate these environments only after the economist constructs a finite-dimensional sufficient state and law of motion tailored to each setting. I develop a transition operator that maps exogenous paths into optimal decision sequences. A history-conditioned version uses only the path observed to date and current endogenous states. A recurrent network learns the history summary relevant for decisions. I apply the method to firm investment and technology adoption along a potentially nonstationary frontier.

In the firm problem, optimal adoption effort reflects both the current technology gap and the shadow value of closing it. Adoption and capital investment are dynamically complementary because each raises the future return to the other. The shadow value depends on the expected frontier path. The same gap can support different decisions when observed histories imply acceleration, plateau, or reversal. As information accumulates, history-conditioned decisions become more specific to the realized transition. In current numerical exercises, the autoregressive Bellman approximation mistimes adoption in opposite directions across transition shapes, delaying it along a moderate S-curve and front-loading it along a rapid one. The full-path operator produces about 9 percent more lifetime value in both cases. Under slow drift, the Bellman benchmark produces 4 percent more lifetime value. The full-path operator observes the entire transition. Its lower value under slow drift reveals approximation error in current training, not a structural advantage of recursion.

Household Finance & Intermediation

Work in progress

Off-Balance-Sheet Costs and Credit Supply

Three estimates with 95 percent confidence intervals show pooled, purchase, and refinance origination responses to servicing-advance exposure after March 2020.

Origination responses to standardized Fannie servicing exposure after March 2020.

Figure notes

Bars report coefficients on standardized Fannie servicing exposure interacted with the post-March-2020 indicator: −0.145 for pooled originations, −0.173 for purchase lending, and −0.070 for refinancing.

Whiskers show 95% confidence intervals based on lender-clustered standard errors of 0.061, 0.048, and 0.051. The estimates are reported in the current manuscript and were not separately replicated for this website package.

View full-size figure

Operational costs increasingly shape mortgage lenders’ ability to extend credit. The rise of nonbank lenders contributes to this shift because these institutions rely on warehouse funding rather than deposits. The growing complexity of household finance also raises the resources required after a loan is made, including collecting payments and assisting borrowers who fall behind. Credit supply can therefore depend on operational capacity, such as servicing existing loans, rather than only on the capacity to fund new lending.

Operational costs for mortgage servicers take two forms. Advances require servicers to forward missed borrower payments to investors. They can be financed and are eventually reimbursed. Workouts, borrower contact and additional operating capacity are instead expensed as incurred.

I use plausibly exogenous differences in Fannie Mae and Freddie Mac advancing rules to study advance burdens during the COVID-19 forbearance wave. Lenders more exposed to Fannie Mae’s longer obligation reduced originations by about 14 percent per standard deviation of exposure. The decline was concentrated in purchase mortgages. Less-exposed lenders absorbed much of the displaced lending.

A structural model distinguishes financeable advances from expensed servicing costs. It will quantify how both crowd out lending when servicing and origination draw on the same funding sources.

Working paper

Integrated Intermediation & Fintech Market Power

with Greg Buchak and Adam Jorring

Paper

Two vertically stacked event-study panels show the relative probability of servicing integration and refinancing around the implementation of Basel III.

Servicing integration and refinancing responses around Basel III.

Figure context and notes

Panel A reports the relative probability that a bank-originated loan is serviced by its originator. Panel B reports the relative probability that the loan refinances in a given quarter. The shaded region marks the implementation of Basel III during the first two quarters of 2013.

The regressions include loan-level controls, lender-by-ZIP fixed effects, and ZIP-by-quarter fixed effects. Together, the panels show the paper’s central reduced-form evidence linking lower servicing integration to lower refinancing.

View full-size figure

Abstract

We study the financial disintermediation of loan origination and loan servicing, an increasingly salient feature of modern credit provision. In the context of the U.S. mortgage market, we show using plausibly exogenous variation from risk-based capital rule changes that borrowers with vertically integrated originator-servicers are significantly more likely to engage in beneficial refinance as compared to those with stand-alone servicers. Critically, borrowers with vertically integrated servicers are significantly more likely to refinance with the servicer’s associated lending business. Motivated by these facts, we build and estimate a dynamic structural model with adverse selection to disentangle a selection channel where originators retain those borrower types who are more likely to prepay (or less likely to default) from a treatment channel where vertical integration causally increases refinance probability through cost advantages and borrower convenience. Our estimated model finds that nearly the entire effect is causal rather than due to selection, and implies that disintermediated origination and monitoring can have significant downsides for borrowers.

Work in progress

Income Histories and the Marginal Propensity to Consume

On a ratio scale, model consumption responses rise from 0.14 to 1.43 while published comparison values rise from 0.05 to 0.69; model values are 2.1 to 3.1 times higher.

The model matches how steeply consumption responses rise, not how large they are.

Figure notes

The blue line reports model responses to one dollar of extra earnings of 0.14, 0.59, and 1.43; the orange line reports published comparison values of 0.05, 0.19, and 0.69. On the ratio scale, bottom-to-top responses rise 10.2 times in the model and 13.8 times in the comparison series. Model levels are 2.1 to 3.1 times higher. The model uses the one-asset benchmark solved by exact value function iteration.

The published comparison values are from Alan, Browning, and Ejrnæs (2018) and are percentiles of the fitted consumption-response parameter distribution, not group-level estimates for earnings-persistence deciles. The package flags an unresolved citation gap linking the three model income processes to a published table. The decile alignment should therefore be treated as provisional.

The package records independent verification of the model responses. No uncertainty bands are shown; the published table reports no standard errors and the model outputs are deterministic.

View full-size figure

I extend the permanent income hypothesis from a recursive policy function to an operator mapping income histories and dated environments into consumption plans. The operator handles heterogeneous ARMA dynamics and hidden regimes, both widely documented in empirical earnings research but cumbersome to accommodate in standard recursive formulations. The same framework accommodates long memory and irregular policy calendars. These features matter for income changes associated with business cycles, major technological transitions, and policies announced and implemented on different dates. A neural primal solver imposes feasibility by construction and maximizes expected lifetime utility. The neural solver makes simulated method of moments computationally feasible.

An MPC becomes an event-specific derivative that decomposes into the standard annuity response to cash and a news term measuring revisions to expected future income. Published household-panel estimates1 jointly recover an earnings process and the response of consumption to earnings innovations for each persistence decile. I feed the estimated earnings processes into the model. The model counterpart is the operator’s response to a one-dollar earnings innovation. At the current benchmark calibration, news accounts for 41, 85, and 94 percent of predicted responses at the bottom, median, and top deciles of estimated earnings persistence. The news share rises with persistence because an income change predicts more future earnings. The model predicts responses of 0.14, 0.59, and 1.44 across the three deciles, compared with published estimates of 0.05, 0.19, and 0.69. The exercise shows that estimated earnings-process heterogeneity can generate the steep cross-sectional gradient in responses to ordinary income changes without matching their levels.

Predicted windfall responses remain between 0.08 and 0.09 across income processes, well below empirical estimates of 0.2 to 0.5. In the current exercises, richer income dynamics account for variation in responses to earnings changes but do not explain large empirical windfall MPCs. Constraints, illiquid wealth, and present bias remain candidate mechanisms for that gap. Preliminary two-asset exercises preserve the cross-process gradient. A full assessment of the interactions among these mechanisms and their joint empirical fit remains open.

1 Şule Alan, Martin Browning, and Mette Ejrnæs, “Income and Consumption: A Micro Semistructural Analysis with Pervasive Heterogeneity,” Journal of Political Economy 126, no. 5 (2018): 1827–1864.