Daily Summary for 2026-11-05

daily

Transcript

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

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

Warren: 31 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 Partially identified heteroskedastic SVARs.

The day's research: Ruth: Generative AI at Work studies how using generative AI assistants helps customer support agents increase their productivity and improve worker skills.

Warren: Access to this assistance generally boosts productivity by about fifteen percent on average, though it varies by experience level.

Iris: Evidence shows AI helps workers learn new skills and improves English fluency, particularly for international agents.

Ruth: The biggest gains happen when dealing with problems that are relatively rare, where human agents often lack prior training.

Warren: This assistance also makes customer interactions better; customers become more polite and less likely to demand a manager.

Iris: This effect builds upon findings from Generative AI at Work regarding support agents across different skill levels.

Ruth: Fixed-smoothing Uniform Inference for Quantile Regression provides a more robust framework for analyzing heterogeneity in predictive effects.

Warren: That paper handles heteroskedasticity well when performing quantile regression inference in time series data.

Iris: The behavioral modeling suggests new regulatory targets are needed beyond simple volatility limits in automated markets.

Ruth: Modeling rational arbitrageurs, manipulators, and retail herd-followers found that rational agents exhibit more stable risk-adjusted performance.

Warren: This is crucial because it supports targeting behaviors rather than applying uniform limits on volatility or turnover.

Iris: Agentic reinforcement learning systems use distinct reward functions for each archetype in this behavioral modeling.

Ruth: The rational arbitrageur maximizes risk-adjusted return with penalties on realized volatility and turnover.

Warren: Conversely, the manipulator receives a bonus proportional to turnover and operates under a short discount horizon favoring rapid cycling.

Iris: Retail herd-followers are rewarded for aligning exposure with recent momentum in this framework.

Ruth: Pilot results on hourly cryptocurrency data showed rational agents have the highest baseline volatility reflecting continual exposure adjustments.

Warren: Manipulators exhibit sharp spikes during stress periods, contrasting with higher volatility not automatically implying more stress.

Iris: This work connects to causal inference by proposing logit-based IV estimators as alternatives to two-stage least squares in binary treatment settings.

Ruth: These estimators offer a causal interpretation because their probability limits can be represented as convex combinations of treatment effects with non-negative weights.

Warren: This contrasts with the traditional 2SLS estimator which may assign negative weights.

Iris: GenAI-assisted writing suggests this technology lowers publication barriers for researchers, potentially increasing output by about seventy-two percent.

Ruth: This productivity gain is largest among early career researchers and those from linguistically distant countries.

Warren: Today's papers include Logit-based alternatives to two-stage least squares proposing new estimators for binary treatment variables.

Iris: The research on GenAI-assisted writing shows using generative AI to write papers can increase publication output among social science researchers.

Ruth: Agentic AI in Automated Markets models different types of trading agents, such as arbitrageurs and manipulators, to understand their impact on market stress and volatility.

Warren: This study models these agents to understand their impact on market stress and volatility.

Iris: What Drives Energy Use? Prices, Policy, and Inefficiency uses an economic model to explain high U.S. energy use despite consumption declines.

Ruth: That paper uses an economic model to explain why U.S. energy use remains high despite national declines in consumption.

Warren: The pursuit of happiness analyzes how individuals maximize happiness through social interactions and revise attitudes based on others' happiness.

Iris: This paper analyzes how individuals maximize happiness through social interactions and how they revise their attitudes based on others' happiness.

Ruth: We look at AI and the research assistant evidence from economics working papers comparing those that disclose AI use to those that do not.

Warren: This study compares economic papers disclosing AI use to those without, finding operational AI is more common in junior labor projects.

Iris: Information Intermediaries in Monopolistic Screening investigates how information intermediaries shape consumer choices and affect market efficiency.

Ruth: That paper investigates how information intermediaries shape consumer choices and ultimately affect market efficiency and welfare.

Lucky paper: 2403.06879: Ruth: Okay, we're diving into Partially Identified Heteroskedastic SVARs today. This paper tackles a real problem when we try to identify structural vector autoregressive models, which is when the variance of the structural shocks changes unevenly across different variables.

Warren: The core issue they address is that traditional point identification methods fail when those shifts in shock variances are statistically indistinguishable from one another, which makes it hard to isolate the true structure.

Iris: So, how does this new strategy propose to overcome that failure for point identification within the HSVAR framework?

Ruth: They propose a new methodology that combines heteroskedasticity assumptions with zero and or sign restrictions on model parameters, which they show allows for point identification even when those variance shifts are statistically indistinguishable.

Warren: That relies on rigorous analytical derivations, specifically Theorem one which establishes the necessary and sufficient conditions based on eigenvalue multiplicities and specific non-redundant zero restrictions like f ij equals m i minus j.

Iris: That sounds mathematically dense; could you explain what that combination of zero restrictions is actually doing for the identification process?

Ruth: It's setting up constraints that use those restrictions to force a unique solution for the model parameters, which then lets them isolate the structural shocks from the observed data, even under these tricky heteroskedastic conditions.

Warren: Beyond just point identification, they extend topological analysis to handle convex identified sets when dealing with these HSVAR models, leading to results in Theorem two.

Iris: So we're not just getting a single point estimate; this also helps characterize the entire set of possible structures that could explain the data.

Ruth: Exactly; it extends the topological analysis from standard SVAR literature to this more complex HSVAR context, giving analytical results for both point and set identification.

Iris: And then they adapt a robust Bayesian approach, which is interesting because it's originally used for set-identified models, to handle the violations caused by heterogeneous variance shifts.

Warren: The final major contribution is that concrete algorithm for estimation and inference using this Bayesian framework, presenting Generalized Least Squares, Maximum Likelihood, and a derived Bayesian estimator in Appendix E.

Iris: That gives researchers a concrete tool for estimating parameters when the standard assumptions break down due to those variance shifts.

Ruth: It’s about providing a practical way to estimate things even when the theoretical identification assumptions are violated in this complex setting.

Warren: This paper, Partially Identified Heteroskedastic SVARs, really pushes the boundaries on how we can apply structural models when data exhibits uneven variance across its components.

Lucky paper: 2504.06550: Ruth: Today we’re looking at Blameocracy, which is paper number two of five on the day's research.

Warren: It investigates how causal rhetoric, which is political language linking actions to outcomes through blame and credit, functions in politics.

Iris: So the study fine-tuned a large language model called RoBERTa-large on four point two million congressional tweets from two thousand twelve to two thousand twenty-three to classify them as blame, credit, or neither.

Ruth: And the finding there is that causal rhetoric is mostly the language used when politicians are in disagreement with each other.

Warren: They found that this style of language is primarily used during political disagreements.

Iris: But it's not just about disagreement; the paper also looks at how this rhetoric affects things like donation returns and voting intentions.

Ruth: Specifically, they used 2SLS estimates to see how blame and credit tweets impact donation revenue, donor counts, and average donation size.

Warren: The results show that blame is what actually yields significant political returns in terms of donations.

Iris: They found that exposure to blame generates gains in support—increasing the donor base by about one point six percent—while credit produces virtually no comparable aggregate response on donor counts.

Ruth: That’s a big difference there; blame gets the funding, but credit doesn't seem to move the needle on donations at all.

Warren: They also looked at voting intentions during rhetoric-only days, examining whether exposure to blame or credit rhetoric influences a respondent's intention to vote for the politician.

Iris: The analysis was robust across different specifications, including date fixed effects and tweet volume controls.

Ruth: It seems the exposure to blame has a measurable effect on voter behavior in this context.

Warren: On top of that, they quantified how rhetorical style relates to amplification by looking at retweets.

Iris: The study found that blame tweets receive significantly more retweets than tweets classified as either blame or credit, suggesting blame is a strong channel for political amplification and larger returns.

Ruth: So the mechanism for getting more attention seems tied directly to being blamed rather than just being credited.

Warren: They also explored how this rhetoric applies to economic discourse, noting that economic claims are often "underqualified yet plausible, politically convenient."

Iris: That context is important because it shows how this rhetorical style plays out in discussions about economics.

Ruth: The paper also points out a partisan alignment where blame attaches to opposing-party legislative action while credit aligns with one's own party's actions.

Warren: That alignment suggests the rhetoric is politically motivated rather than purely objective communication.

Iris: Finally, they tested if changes in Twitter's platform design or character limits have caused shifts in this causal rhetoric and compared the rhetorical styles of Donald Trump against members of Congress.

Ruth: They concluded that Trump relies more heavily on non-causal affective messaging when compared to the language used by members of Congress.

Lucky paper: 2105.10965: Ruth: So, we're diving into this paper now, Inference on Quantile Treatment Effects with Small Treatment Groups by Dias and Pouzo.

Warren: It’s interesting because it addresses the problem of estimating conditional quantile treatment effects when the treatment groups are actually very small.

Iris: The core idea is combining an estimate from a large reference group with an exact Bernoulli representation of those small treatment groups to keep things asymptotically valid as the sample grows.

Ruth: That sounds like a smart way to control the size of those treatment groups without losing accuracy.

Warren: They build this in a dynamic potential outcomes framework where the structural outcome function determines what you see based on past outcomes and observed characteristics.

Iris: The key assumption is a quantile analogue of selection on observables, meaning that conditional on past potential outcomes and observed characteristics, the treatment selection preserves the quantile of interest for an unobserved rank.

Ruth: So they are relying on this idea that the treatment assignment doesn't change the underlying distribution of what you are measuring at a specific point in time.

Warren: Exactly, and they have to satisfy certain restrictions on selection given past outcomes and observed characteristics for this framework to work properly across randomized experiments or observational models.

Iris: When testing, they use three steps: first estimating the structural conditional quantile function from that large reference group, second evaluating that estimate in the sample moments of those treatment-profile groups under a null hypothesis, and third comparing an aggregate of those moments against a critical value from their conditional Bernoulli distribution.

Ruth: That sounds like a very structured testing procedure to ensure they get a reliable result.

Warren: The paper proves asymptotic size control and derives an explicit lower bound on power, which means they can actually keep the treatment group sizes bounded while still controlling the error.

Iris: They show that this power bound depends on how much the moment displacement exceeds the critical threshold relative to the finite sample variability of those treated-profile moments.

Ruth: That's a concrete way to measure how detectable an effect is in practice, not just theoretically.

Warren: The Monte Carlo simulations confirm that holding the expected treatment group size fixed, this feasible test approaches its oracle behavior as the reference group grows larger.

Iris: But they also point out that power is constrained by the information and discreteness inherent when those treatment groups are extremely sparse.

Ruth: So even with this clever method, if you only have a tiny group, you still face limitations in how much power you can actually get.

Warren: The paper applies this framework to gauge the effect of large temperature changes on GDP growth, and it shows that conventional significance assessments can overestimate the strength of evidence when the number of observations informative about those effects is small.

Iris: That's a big warning because it means if you only have a few data points relevant to an economic effect, standard tests might make you think there's more evidence than really exists.

Ruth: The paper does not reject the absence of an effect in these GDP growth simulations, which is significant because conventional assessments can be misleading there.

Warren: The procedure itself becomes conservative when the reference group used for estimation is itself small, even though it can still have substantial power with a small treatment group.

Iris: It’s like you’re trading off precision in estimating the structural function against the ability to detect an effect in that sparse group.

Ruth: They also showed that the null rejection probability is largely insensitive to the shape and tails of the disturbance distribution, which is good because you don't have to worry too much about assuming a specific distribution there.

Warren: They also provide a method to construct confidence regions for finite-dimensional parameters indexing those structural quantile treatment effects.

Iris: The coverage result shows that as the sample size goes to infinity, the probability that the true parameter is within your confidence region approaches one minus alpha, which is what we want in practice.

Ruth: It’s helpful because it gives you a way to build confidence intervals for those specific quantile treatment effects when you can't run an infinite simulation.

Warren: The paper builds on literature about inference with few treated observations and many controls, showing how this new method fits into that existing research landscape.

Iris: It also relates to work on staggered adoption in difference-in-differences designs, which is relevant for many real-world policy questions involving gradual changes over time.

Lucky paper: 2609.05883: Ruth: Okay, we're looking at Fixed-smoothing Uniform Inference for Quantile Regression, Kaicheng Chen and the team developed this method to handle heteroskedasticity and autocorrelation in time series quantile regression.

Warren: So it's a new way to do inference across all quantile levels uniformly, accounting for dependence both over time and across those different quantiles.

Iris: I’m curious about how the 'Uniform-in-tau method' works specifically to address the non-pivotal limiting distributions under weak dependence.

Ruth: Well, they estimate the covariance structure and then simulate those distributions to get uniform confidence bands and Wald tests for joint hypotheses.

Warren: That simulation approach is key because it moves beyond just relying on standard Heteroskedasticity and Autocorrelation Consistent procedures.

Iris: And the 'Stack-Wald method' provides pivotal fixed-smoothing inference for tests involving a finite collection of quantile levels.

Ruth: The authors established the asymptotic validity for both the uniform-in-tau and stack-Wald approaches, showing they are robust even in complex time series settings.

Warren: Simulation results showed these fixed-smoothing methods improve size control compared to existing HAC procedures while keeping good statistical power.

Iris: Empirically, they applied this to predictive quantile regressions for stock returns and found significant heterogeneity across both quantiles and various forecast horizons.

Ruth: They noted that the stock variance has its strongest predictive effect at short horizons, especially within the tails of the return distribution, diminishing as the forecast horizon increases.

Warren: The net equity expansion showed a uniformly positive trend with stability across quantiles where confidence bands suggested statistical significance over a broad range of distributions.

Iris: When comparing Wald versus Sup-t tests, they found that while Wald tends to have higher null rejection rates as the number of restrictions goes up, the Sup-t test is more robust to an increasing number of restrictions.

Ruth: That means for many quantiles, the Sup-t test could be a better alternative when you have a large number of restrictions.

Warren: The choice of grid for quantile levels does matter; they found that a tail-based grid yields lower predictive power than a center-based grid.

Iris: This confirms that inference becomes more challenging when focusing on the tails, which is important context for applying this framework.

Ruth: The authors also showed a trade-off between size distortion and power across different smoothing parameters, which you have to navigate carefully when picking your bandwidth choices.

Warren: So for small numbers of restrictions, the Wald test has strong power, but that power drops substantially as you increase the number of restrictions.

Iris: We also saw that for a general trade-off between size distortion and power, the choice of grid for quantile levels matters significantly in terms of predictive performance.

Ruth: The main contribution here is providing a unified set of tools for studying heterogeneity by explicitly modeling dependence across quantiles in time series QR.

Warren: That unification is powerful because it gives researchers one consistent methodology to study how effects change across different parts of the distribution.

Iris: This framework enables simultaneous and uniform inference for quantile effects by modeling the dependence structure between quantiles directly.

Ruth: So, this paper really solidifies a unified way to conduct robust inference in time series QR while rigorously assessing heterogeneity across different distribution tails.

Lucky paper: 2505.22909: Ruth: So we're looking at Learning to Charge More, which is about how Q-learning agents learn to set prices above competition in repeated games.

Warren: It introduces a framework based on stochastic games with bounded memory, analyzing the learning process where firms update their pricing policies based only on observed profits.

Iris: The paper extends classical fixed-point theory to show the existence of one-memory subgame perfect equilibria in this setting, which allows for analysis of dynamic collusion.

Ruth: The authors study a variant of Q-learning with bounded experimentation to model how firms adapt their pricing policies based on observed outcomes.

Warren: They show that when the game admits both a one-stage Nash equilibrium price and a collusive-enabling price, and when the Q function satisfies certain inequalities at the end of experimentation, firms learn to consistently charge supracompetitive prices.

Iris: This learning process is formalized through Algorithm three which combines softmax based Q learning with bounded experimentation.

Ruth: The paper identifies three types of policy profiles that can support this learned supracompetitive behavior.

Warren: Those profiles are one naive collusion, where firms always choose the collusive enabling price, or grim trigger policies involving credible threats and dynamic escalation patterns.

Iris: Grim trigger policies involve a specific pattern of threats and escalation over time to sustain the higher prices.

Ruth: Let's talk about what that means for a firm operating in real-time markets. Does this mean they have to be explicitly programmed with these specific policy profiles?

Warren: Not necessarily, but the Q function satisfying those inequalities seems like a condition that emerges from the interaction of the learning process and the game structure.

Iris: From a representation learning standpoint, this suggests that simple memory structures combined with bounded experimentation are sufficient to capture complex dynamic collusion.

Ruth: It shifts our view on how firms adapt their pricing strategies based only on observed profits without needing complex equilibrium calculations upfront.

Warren: That’s a key point because it implies the adaptation is driven by the immediate profit signal, not a pre-calculated long-term plan.

Iris: The paper also focuses on how these policies sustain collusion in an infinite time horizon, which is important for understanding long-term market behavior.

Ruth: Does this learning mechanism hold up if we introduce more than one memory into the firms' pricing policies?

Warren: The current model uses a one-memory policy because it simplifies the analysis to establish the existence of subgame perfect equilibria in this specific setting.

Iris: If you increase the memory, you move away from this specific theoretical framework, but it might capture more realistic firm behavior.

Ruth: The authors specifically test how they learn to charge prices above competition when certain conditions are met by using bounded experimentation.

Warren: They use that bounded experimentation to model the adaptation process and show the learning mechanism itself is what drives the supracompetitive pricing.

Iris: So, even if a firm starts with naive collusion, it can learn to escalate its price based on observed outcomes within that learning loop.

Ruth: That’s interesting because it suggests that coordination, even simple coordination, can lead to an emergent strategy of overcharging.

Warren: It’s about how the interaction between the Q function and the bounded experimentation drives the system toward a higher price point than what a standard Nash equilibrium would suggest.

Iris: This has implications for how we model dynamic collusion in repeated games where firms are constantly interacting with each other.

Ruth: We need to consider that this is theoretical, so the real-world success depends on whether those specific inequalities hold up under messy, imperfect market conditions.

Warren: Right, the theory establishes the condition for learning this behavior, but empirical testing would be necessary to confirm if those exact conditions are met in practice.

Iris: We have to be careful not to assume that every firm will follow these specific policy profiles perfectly when they are actually running.

Ruth: The paper doesn't explicitly detail the cost of implementing such a learned supracompetitive price profile for the firm, which is a practical question.

Warren: That is something we would need to explore further; the theoretical focus here is on the existence and learning mechanism of the pricing behavior itself.

More episodes

← Home