Daily Summary for 2026-11-10

daily

Transcript

Introduction to the show: ident: Economics. The day in economics: theory, mechanisms and the evidence beneath them.

Ruth: It's the tenth of November, twenty twenty-six, and this is the day's research.

Warren: 20 new papers came out today.

Ruth: I'm Ruth.

Warren: And I'm Warren, here with Iris.

Ruth: We'll take the day in one pass, then pull out the papers we're staying with.

Warren: Start us off with Does Frontier AI Strengthen Cyber Defense? Evidence from Project Glasswing.

The day's research: Ruth: We are looking at online Bayesian inference for Dirichlet process mixture models because traditional batch methods struggle with constant new observations.

Warren: That method uses a sequential Monte Carlo sampler with Pólya-Gamma-augmented partial-data posteriors and local particle-weight corrections for the data stream.

Iris: That sequential updating allows tracking posterior evolution with low latency, giving provisional summaries as trip records are processed one by one.

Ruth: That contrasts with batch Markov chain Monte Carlo, which simulation studies show yields similar final results but takes much longer.

Warren: The most pressing finding concerns how retraining programs affect unemployment outcomes following an initial dip in earnings.

Iris: That dip is mainly driven by industry switchers moving into healthcare, suggesting targeted postsecondary enrollment can yield long-term economic benefits for the unemployed.

Ruth: This is supported by work on handling complex estimation problems when many proxies are generated for a latent control variable.

Warren: The authors show that with enough rich controls, one can consistently estimate the parameter of interest without needing validation data.

Iris: Another important area involves understanding decision-making under uncertainty through correlation probabilistic sophistication.

Ruth: That framework shows how simultaneous event exchangeability leads to a unique solvable probability, simplifying binary comparisons into induced joint laws.

Warren: This builds on ideas where different models are tested, like the neural score-driven volatility model that uses a lightweight neural network.

Iris: Finally, there is the theoretical underpinning for how information value is calculated in optimization through regret in stochastic programs.

Ruth: That identity proves regret is determined by the covariance between the cost vector and the optimal decision, providing exact bounds on information value.

Warren: Flexible travel behaviour models can update posterior inferences as new data arrives by sequentially updating cluster structures and historical assignments.

Iris: Does Frontier AI strengthen cyber defense? Evidence from Project Glasswing access to frontier AI tools helps defenders find and fix more software flaws before attackers exploit them.

Ruth: Learning to design information designers can learn the subjective beliefs of their users through repeated interactions using robust learning algorithms that achieve no regret against strong benchmarks.

Warren: Optimality of separate selling with finite-support values characterizes when separate selling is revenue-optimal in multi-item auctions by showing it is equivalent to top selling under certain valuation distributions.

Iris: A general theory of dynamic mechanism design covers joint endogeneity, hidden actions, and information acquisition across various economic environments.

Ruth: Communication as voting shows that in informal elections, communication can only reveal information that induces the same preferred policy for both the receiver and the senders.

Warren: Full subgame-perfect implementation of optimal risk sharing on an infinite menu extends a mechanism to fully implement optimal risk sharing allocations in finance and insurance under complete information.

Iris: Priced in or bet on? Cram-School lessons, Jeonse Leases, and Expected Appreciation in Seoul shows access to private cram schools is capitalized into housing prices as a cost of the education race.

Ruth: Further education during unemployment evidence on the effectiveness of retraining unemployed workers primarily comes from evaluations showing that postsecondary enrollment leads to earnings gains for industry switchers.

Warren: Correlation probabilistic sophistication introduces a subjective reduction principle that extends probabilistic sophistication to binary choice under uncertainty based on joint outcome distributions.

Iris: Regression with many generated controls shows that practitioners can consistently estimate the parameter of interest in linear regression even when some covariates are endogenous by using many generated proxies as controls.

Ruth: Neural score-driven volatility and risk forecasting proposes a score-driven volatility model that uses a neural network to flexibly represent innovation density, performing competitively against parametric benchmarks.

Warren: Regret equals integrated covariance: exact identities for the value of information in stochastic optimization provides an exact identity showing that regret in stochastic optimization is determined by the covariance between cost and optimal decisions.

Iris: What is work for? Labor, politics, and meaning in the age of AI finds that non-monetary work motivations are evenly distributed across income levels and have implications for managing the transition to an AI economy.

Ruth: A sequential foundation for two-level nested logit shows that the two-level nested logit model arises exactly from a sequential choice process where shocks are revealed in stages.

Warren: Climate change, sea level rise, and the agricultural revolution tests hypotheses about agriculture's invention and spread, finding that technological progress was the main factor in extending its reach over time.

Iris: Information divergence and endogenous evolution of asymmetric information identifies that information asymmetry can arise from agents holding different predictive distributions due to differences in the age of their information.

Ruth: Online Bayesian Inference for a Dirichlet Process Mixture of Binary Logit Models flexible travel behaviour models can update posterior inferences as new data arrives by sequentially updating cluster structures and historical assignments.

Warren: Does Frontier AI strengthen cyber defense? Evidence from Project Glasswing access to frontier AI tools helps defenders find and fix more software flaws before attackers exploit them, but it creates security asymmetries between firms.

Iris: Learning to design information designers can learn the subjective beliefs of their users through repeated interactions using robust learning algorithms that achieve no regret against strong benchmarks.

Ruth: Optimality of separate selling with finite-support values characterizes when separate selling is revenue-optimal in multi-item auctions by showing it is equivalent to top selling under certain valuation distributions.

Warren: A general theory of dynamic mechanism design covers joint endogeneity, hidden actions, and information acquisition across various economic environments.

Iris: Communication as voting shows that in informal elections, communication can only reveal information that induces the same preferred policy for both the receiver and the senders.

Ruth: Full subgame-perfect implementation of optimal risk sharing on an infinite menu extends a mechanism to fully implement optimal risk sharing allocations in finance and insurance under conditions of complete information.

Warren: Priced in or bet on? Cram-School lessons, Jeonse Leases, and Expected Appreciation in Seoul shows access to private cram schools is capitalized into housing prices and deposits as a cost of the education race rather than expected appreciation.

Iris: Further education during unemployment evidence on the effectiveness of retraining unemployed workers primarily comes from evaluations showing that postsecondary enrollment leads to earnings gains for industry switchers.

Ruth: Correlation probabilistic sophistication introduces a subjective reduction principle that extends probabilistic sophistication to binary choice under uncertainty based on joint outcome distributions.

Warren: Regression with many generated controls shows that practitioners can consistently estimate the parameter of interest in linear regression even when some covariates are endogenous by using many generated proxies as controls.

Iris: Inference in clustered IV models with many and weak instruments develops identification-robust jackknife score statistics for instrumental variable models with many weak instruments and clustered data.

Ruth: Neural score-driven volatility and risk forecasting proposes a score-driven volatility model that uses a neural network to flexibly represent innovation density, performing competitively against parametric benchmarks.

Warren: Agreements have a clock: compute growth and the verification of AI restraint argues that agreements on pausing frontier AI are only supportable if the cost of monitoring does not exceed the gains from cooperation over time.

Iris: Regret equals integrated covariance: exact identities for the value of information in stochastic optimization provides an exact identity showing that regret in stochastic optimization is determined by the covariance between cost and optimal decisions.

Ruth: What is work for? Labor, politics, and meaning in the age of AI finds that non-monetary work motivations are evenly distributed across income levels and have implications for managing the transition to an AI economy.

Warren: A sequential foundation for two-level nested logit shows that the two-level nested logit model arises exactly from a sequential choice process where shocks are revealed in stages.

Iris: Climate change, sea level rise, and the agricultural revolution tests hypotheses about agriculture's invention and spread, finding that technological progress was the main factor in extending its reach over time.

Ruth: Information divergence and endogenous evolution of asymmetric information identifies that information asymmetry can arise from agents holding different predictive distributions due to differences in the age of their information.

Iris: Does Frontier AI strengthen cyber defense? Evidence from Project Glasswing access to frontier AI tools helps defenders find and fix more software flaws before attackers exploit them, but it creates security asymmetries between firms.

Lucky paper: 2611.06629: Ruth: So we're looking at the paper, Does Frontier AI Strengthen Cyber Defense? Evidence from Project Glasswing, and the core finding is that access to frontier AI tools helped defenders find and fix more flaws before attackers could exploit them.

Warren: That's a big claim because it suggests a direct benefit to security outcomes, but we need to look at the numbers on how much more disclosed information there was.

Iris: The study shows that firms with access more than doubled their vulnerability disclosures relative to Chinese and Russian tech firms without access, and they disclosed about one hundred fifty percent more high-impact vulnerabilities.

Ruth: That's a massive increase in reported flaws, but the paper also notes that treated firms fixed those reported flaws as quickly as before.

Warren: The study also looked at external researchers selling findings through bug bounties, and those researchers reported about half as many vulnerabilities in the code of treated firms compared to other firms' code.

Iris: And crucially, there was no evidence that attackers exploited these additional vulnerabilities; instead, the share exploited within a month of disclosure fell by about two-thirds because a smaller share was already exploited upon disclosure.

Ruth: That suggests frontier AI helped defenders find and fix more flaws before attackers could exploit them, but it creates a security asymmetry between firms in different countries.

Warren: That leads us to the second point of the paper, which examines how AI-enabled vulnerability discovery affects the allocation of independent research effort across software vendors.

Iris: The results indicate that when defenders and attackers search the same stock of flaws, a better search technology's value depends less on the sheer number of flaws found and more on finding them first.

Ruth: So access to the model allowed defenders to remove latent flaws before attackers used them, and researchers displaced by this access carried their effort to other code, spreading some benefit beyond the firms with direct access.

Warren: That brings up a geopolitical dimension because the exclusion criteria for Project Glasswing deliberately excluded firms based in or majority-owned by entities in China and Russia.

Iris: Relative to those excluded firms, the firms with access roughly doubled the latent flaws they removed from their software, and this restriction functions similarly to an export control allocating strategic technology along geopolitical lines.

Ruth: But timing measures regarding Chinese vulnerability disclosure around the launch were deemed weak, suggesting no discernible change in disclosure patterns immediately following that specific launch.

Warren: However, firm-level analysis showed high shares of vulnerabilities with CVE identifiers at zero point seven eight, zero point nine four, and zero point nine five for the treated groups, and significant activity from independent researchers crediting vulnerabilities for Chinese software across various channels.

Iris: Furthermore, the disclosure gap analysis indicated that Glasswing times Post had a high proportion of vulnerabilities with a gap of at most seven days at zero point five two four, and the analysis showed that three months exploit code was zero point five zero three with significance at the one percent level.

Lucky paper: 2611.05692: Ruth: We're moving into a paper called A General Theory of Dynamic Mechanism Design with Endogenous Information and Hidden Actions by Guoqiang Tian from Texas A andM University.

Warren: This paper develops a general theory of dynamic mechanism design across two dimensions: joint endogeneity and interaction between hidden actions, information production, and disclosure.

Iris: So it handles correlated private information, interdependent preferences, and potentially infinite-dimensional allocations within the framework.

Ruth: It introduces joint perfect correlated Bayesian equilibrium or joint PBE and establishes a dynamic revelation principle under that concept.

Warren: The incentive compatibility is recursively characterized using joint perfect correlated Bayesian incentive compatibility or joint PCBIC for direct mechanisms.

Iris: That requires four necessary and sufficient conditions for direct mechanisms: contingent choice optimality, the reporting-envelope first-order condition, global integral monotonicity, and residual history-report incentive compatibility.

Ruth: In private values environments, those same four conditions are used to characterize joint PCBIC.

Warren: The optimal contracting analysis yields an exact dynamic information-rent identity and conditional revenue equivalence.

Iris: It also yields a dynamic virtual surplus representation using controlled impulse responses to show how the dynamic virtual type weights the initial rent hazard by its controlled propagation.

Ruth: Comparing first best, second best with hidden information, second best with hidden action, and third best with both frictions allows us to characterize information rents and hidden-action implementation costs.

Warren: The paper shows that optimal incentive and disclosure design are generally nonseparable because informational costs and hidden-action implementation costs interact nonadditively.

Iris: The dynamic distortion hierarchy characterizes the optimal output rules among all admissible full-history jointly PCBIC contracts under fixed disclosure rules, providing specific allocation rules for different environments.

Ruth: It also shows that the transaction-cost perspective unifies informational and hidden-action implementation costs by decomposing losses into informational transaction cost and additional implementation cost.

Warren: Proposition fifteen establishes a marginal trade-off for disclosure intensity delta, showing the optimal rule depends on whether S of delta equals R IC of delta plus R zero of delta.

Iris: Proposition sixteen demonstrates that under a quadratic screening benchmark, full disclosure of available evidence is uniquely optimal in every contracting environment, and the optimal effort rule depends on the specific environment.

Lucky paper: 2505.14639: Ruth: Alright, we’re moving on to our third paper this day, and it’s titled Communication as Voting from Aalto University.

Warren: It looks like they're looking at information aggregation in informal elections using a cheap-talk model with multiple senders and one receiver.

Iris: The paper establishes an equilibrium with a no-conflict property where communication only reveals information that induces the same preferred policy for both parties.

Ruth: That’s interesting because it sets up two distinct types of conflict they are modeling: one from differences in payoff magnitudes, which appears under uncertainty, and another from opposing policy preferences in a disagreement state.

Warren: That disagreement state M is key because it prevents complete learning as the number of senders grows indefinitely.

Iris: The paper finds a discontinuity relative to binary benchmarks where for certain parameter ranges, assigning any positive probability to the disagreement state causes communication to unravel as population size increases sufficiently.

Ruth: So what’s driving that discontinuity? Is it uncertainty-based conflict demanding enough information to resolve uncertainty, while full-information conflict makes highly informative communication unsustainable?

Warren: That tension between resolving uncertainty and maintaining sustainable information is a central point.

Iris: They introduce a mediator who can enhance transmission and restore efficiency by committing ex ante to a mechanism that collects messages and recommends an outcome.

Ruth: And what does the paper ultimately conclude about mediated communication in this setting?

Warren: The central conclusion is that in cheap-talk communication, information fails to aggregate, and communication may unravel entirely, but mediated communication can asymptotically implement any Pareto-efficient outcome on the frontier between the receiver’s first best and the senders’ optimal Bayesian persuasion outcome.

Iris: The centralized benchmark suggests that under certain conditions when conflict persists under full information, transmission fails in a single fully informed sender setting but it persists in the decentralized setting with dispersed information.

Ruth: The final remark highlights that decentralizing information across a crowd can outperform centralization in specific scenarios, which seems like a practical implication for how we think about group decision-making.

Warren: Centralization corresponds to one sender observing multiple signals, which amplifies adverse selection and reduces receiver trust when the conflict of interest is large.

Iris: Conversely, in the multi-sender model with dispersed information, noise created by failed aggregation can act as a commitment device to mitigate conflict and improve communication, as shown by Blume et al. two thousand seven.

Ruth: So information transmission persistence depends on both sources of conflict captured by US(L)US(H)UR(H)UR(L) and q zero M being small.

Warren: The presence of the disagreement state M sharply alters these conditions, creating a discontinuity relative to the binary benchmark, specifically if q zero M is greater than zero.

Iris: If q zero M is greater than zero no informative equilibrium exists for sufficiently large N, and babbling becomes the unique outcome.

Ruth: They also establish limitations on sender welfare by showing a finite population size N* beyond which increasing the number of senders does not improve sender welfare.

Warren: Furthermore, informative equilibria generated under higher-conflict parameters imply existence under lower-conflict parameters with weakly larger x.

Iris: The model is extended to nonbinary signal and message spaces, showing that in monotone equilibria, the signal space is effectively binary.

Ruth: So even when we look at more complex signals, the structure simplifies back down under monotone conditions.

Warren: They conclude by confirming that mediated communication can asymptotically implement any allocation on the Pareto frontier using a binary voting mechanism and randomized qualified-majority rules.

Lucky paper: 2611.05767: Ruth: So we're looking at the paper titled Priced In or Bet On? Cram-School Lessons, Jeonse Leases, and Expected Appreciation in Seoul.

Warren: It looks like this study is looking at how families in South Korea spend money on private cram schools and whether that spending affects housing costs.

Iris: The core finding seems to be that proximity to high-school exam preparation hagwons increases both sale prices by eleven point eight percent and jeonse deposits by seven point three percent when you move within a five hundred meter radius of the ninety-seventh percentile of such facilities.

Ruth: That's a specific number, so we have to look at that detail next; it’s not just some general trend with housing prices.

Iris: They also found that holding the complex size fixed, this move is associated with sale price premiums of five point six percent and deposit premiums of five point nine percent.

Warren: That suggests the housing premium is capitalized in proportion into both prices and deposits, which they call a current-flow amenity rather than something based on expected appreciation.

Ruth: Wait, so if you fix the complex size, the change in sale price is five point six percent and deposit premium is five point nine percent for that proximity move?

Iris: That's right; specifically for the same flat type within the same complex and month, high-school hagwons produce no detectable change in the jeonse to price ratio, which is a coefficient of zero point zero zero two.

Warren: The paper also noted that convenience effects are distinguished because while hagwon access shows a larger premium in deposits than in prices across complexes, that gap narrows when commercial density is held fixed or when hagwons are measured by registry count.

Ruth: So the gap between the deposit premium and the sale price premium shrinks if we look at how many schools there are nearby, not just where one specific school is.

Iris: They also pointed out that new tenants pay no lasting entry fee because lease reforms capped deposit increases for renewing tenants.

Warren: This implies that what you're seeing is a temporary markup from the lease renewal, not something that sticks with the new resident.

Ruth: So, to summarize, the overall conclusion is that families pay a housing cost associated with competition for university places which isn't accounted for in measures of tutoring spending.

Iris: That’s right; it shows that access to these services warrants a premium in housing costs and expected appreciation.

Lucky paper: 2312.17123: Ruth: Okay, we're looking at the paper on Further Education During Unemployment from Cornell University. It shows earnings gains of six percent three to four years after enrolling, following an initial dip in earnings during the first two years.

Warren: That initial dip is mainly driven by industry switchers moving into healthcare, and that’s a key detail to keep in mind when we talk about the paper on Further Education During Unemployment.

Iris: It also contributes methodologically to the dynamic treatment effect literature by connecting studies that estimate the treatment-now versus later and treatment versus no-treatment effects of training.

Ruth: The study estimates that for industry switchers, the earnings difference between enrollees and non-enrollees in quarter sixteen is one hundred three percent of the overall earnings gain.

Warren: That’s a pretty big jump, so we have to be careful how we frame that when we discuss the paper on Further Education During Unemployment.

Iris: The cost-benefit analysis indicates that the net present value of average post-tax earnings impacts is thirty-seven thousand fifty-six against an out-of-pocket cost of six thousand one hundred twenty-one resulting in a private net benefit of thirty thousand nine hundred thirty-six.

Ruth: And the social net benefit is estimated at twenty-nine thousand one hundred ninety-two which suggests the worker receives about a dollar for every six dollars invested.

Warren: That's an eight percent social internal rate of return, so we have to weigh that against how much training actually affects wage rates four years after enrolling.

Iris: The study concludes that training mainly affects employment and likely has minimal impact on wage rates four years after enrolling, but the long-run effect is estimated at a thirteen percent gain in the tenth year after enrollment.

Ruth: That’s a specific number to hold onto when we discuss the paper on Further Education During Unemployment because it shows that long-run potential.

Warren: The data also examines subgroup effects, and gains to enrollment are larger for men in the four years post-enrollment at nine percent compared to women at four percent.

Iris: There’s a difference in school type and quality too; technical center enrollees showed a more substantial lock-in effect than community college enrollees.

Ruth: The paper also finds that workers laid off when unemployment rates peaked during the Great Recession have earnings gains of eight to twelve percent in the third and fourth years post-enrollment.

Warren: That comparison against cohorts laid off prior to the recession in two thousand six to two thousand seven shows gains indistinguishable from zero, which is a very important contrast for understanding the timing aspect of this paper on Further Education During Unemployment.

Iris: The effect is larger among manufacturing workers than non-manufacturing workers in the third and fourth years post-enrollment, but for young unemployed workers, additional training is less beneficial if they already possessed some college experience.

Ruth: And there’s also a detail about enrollment timing: later enrollees have a smaller lock-in effect in the first two years but larger gains from enrollment up to twelve percent in the third and fourth years.

Warren: That timing variance is something we should keep focused on when we talk about the paper on Further Education During Unemployment because it shows it isn't just about showing up.

Iris: The overall treatment effect is estimated to be approximately six percent in the third and fourth years after enrolling, with all estimates being highly statistically significant across tested specifications.

More episodes

← Home