AI traders lose 22% on Kalshi but near break-even on Polymarket
57-day test shows platform design decides which frontier models can profit from real trades
· “Prediction Arena: Benchmarking AI Models on Real-World Prediction Markets”
econ
Economics
sort pith recommended most recent
DiD estimator picks pre-trend length to minimize MSE
Balances bias from longer windows against variance from shorter ones for lower overall error in small samples.
No design can approach the efficiency bound unless d << log n
Even with thousands of units, only a few covariates can be balanced; new designs target smaller function classes instead.
· “The Limits of Experimental Design: Covariate Balance Beyond Low Dimension”
One budget inequality decides what exploration can achieve
Dynamic learning and stopping collapse to a static convex program, priced by a shadow cost of information over time.
EADA's expected rank is O(log log n)
In i.i.d. random markets, every Pareto-efficient improvement of Deferred Acceptance provably breaks the logarithmic average-rank barrier.
· “Efficiency Adjustments Break the Logarithmic Rank Barrier”
Optimal no-transfer mechanism never requires the agent's report
Disclosure plus delegation replaces transfers and keeps the agent's benefit private.
One tree organizes every risky multidimensional preference
Seven axioms pin down a recursive utility tree that unifies joint, separate, and conditional risk evaluation.
Mid-size model tops energy-agent benchmark at a quarter of the cost
Why care: per-family scores show routing by task type beats picking one big model.
· “EnergyAgentBench: Benchmarking LLM Agents on Live Energy Infrastructure Data”
AI markets will pay premiums for verified human presence
As synthetic substitutes erode middle-tier knowledge work, governance must treat provenance verification as labor infrastructure to support
· “Human-Provenance Verification should be Treated as Labor Infrastructure in AI-Saturated Markets”
AI buyer agents leak willingness to pay from dialogue
Seller inference recovers budgets nearly one-for-one from natural-language profiles, and confidentiality instructions do not stop it.
· “When Agents Shop for You: Role Coherence in AI-Mediated Markets”
Senior staff write clearer AI prompts
Field data from 713,564 prompts shows sophistication rises with rank and function, with no lasting training gains.
· “Sophistication in GenAI Use: Field Evidence from a Large Firm”
First explicit orthogonal basis for consistent comparison matrices
Closed-form projection replaces iterative fitting and unifies logarithmic, Saaty, and SVD windowing.
Fanning-out makes time-averaged regressions super-consistent
When the cross-section variance of averaged data grows with time, ordinary regressions beat root-N; mean-reverting data do not.
· “Cross-Section Estimation of Long-Run Relations Using Time-Compressed Data”
Posterior converges to the true identified set
Consistent Bayesian inference for set-identified models, with no covariate binning or ad hoc moment selection.
· “Nonparametric Bayesian Inference for Partially Identified Discrete Response Models”
Groundwater markets are stochastic games—and math can manage them
A tractable model of pumping, trading, and banking gives regulators a simulation lab before rules are locked in.
Ending a solar rebate beats extending it
A pre-announced deadline triggers an adoption surge, so limited-duration rebates deliver more rooftop solar per dollar.
Deliberation revises choices only toward the emphasized dimension
A revealed-preference theorem says full-horizon rankings must extend a protected partial order, giving sharp tests.
· “Directional Revision under Two-Horizon Deliberation: A Revealed-Preference Analysis”
Forecast-error PCA finds business cycles without detrending
One OLS regression per series isolates stationary cyclical components, removing the need to pre-test for unit roots or choose…
· “Principal Component Analysis for a Mix of Stationary and Nonstationary Variables”
Random validation turns biased AI into credible measurement
Label a random sample, learn the model's errors, and black-box AI can serve as credible measurement for economics.
· “The Measurement Revolution? Credible Measurement and Inference in the Age of AI”
New test detects nonignorable panel attrition
Compares reweighted distributions to test selection on observables in panels with refreshment samples.
· “Testing selection on observables in parametric models with refreshment samples”
One in four U.S
Y02 errors are systematic: digital tech is overcounted while heavy-industry decarbonization is undercounted.
· “Systematic Bias in Green Patent Classification: Silent Green and False Green”
This paper finds the best way to assign a fixed set of prizes in a competition when the…
In contests with fixed prizes, the effort-maximizing grading policy irons out locally misordered incentive returns, pooling adjacent ranks…
Individualistic societies see more constitutional compliance
Using 115 countries, the study links a nation's culture to whether its government keeps constitutional promises.
One test can declare an effect positive
The procedure returns an error-controlled class or "inconclusive", instead of only rejecting or failing to reject a null hypothesis.
A p-value of 0.05 survives replication only 10-25% of the time
Same-size replications of p=0.05 findings usually fail: original studies were underpowered, not selectively published.
E-value tests show inflation forecast miscalibration comes in bursts
Sequential monitoring spots overconfident Bayesian forecasts during crises that static tests miss.
· “Sequentially valid inference for probabilistic inflation forecasts”
Female GOP nominees for governor lose 6.5 points since 2016
Stacked county comparison finds no comparable penalty for Democratic women; pooled nulls hid the asymmetry.
· “Female Nomination and Party Vote Share in US Gubernatorial Elections”
Causal direction in extremes emerges when tails differ
Tail-heaviness gaps give a clean asymmetry, and proxy-adjustment recovers it when heavy-tailed confounders interfere.
· “Identification and Inference for Causal Effects in Extremes under General Conditions”
Randomization test flags misspecified network spillover models
It uses graph-weighted correlation between treatment and residual to test exposure mappings in experiments.
· “Randomization tests for model specification in causal inference under network interference”
Papers in spatial economics often assume a fixed rule for how neighbors influence each…
An operator-orthogonal double machine learning estimator delivers root-n normal inference for spatial autoregressive parameters when the…
· “Double/Debiased Machine Learning for Functional-Form-Robust Spatial Autoregression”
This paper studies team contests made of many small battles
In majority-rule team contests, finer information (more disclosure or a finer schedule) spreads each battle's incentives without changing…
· “Outcome Disclosure and Temporal Refinement in Multi-Battle Team Contests”
Pathway contributions to survival are now identifiable
Decomposes drug effects into direct, transition, and path-specific pieces without sequential ignorability.
· “Estimating Pathway Treatment Effects in the Presence of Intermediate Events with Multi-State Data”
Closed form recovers treatment curve with nonseparable first stage
Plug-in kernel estimator skips regularization, matches the optimal nonparametric rate, and gets a uniform bootstrap band.
Fixed-T spatial panels: common shocks without N loading parameters
Projecting factor loadings onto covariates and spatial lags removes the fixed-T bias; county data highlight education.
High minimum values force simple deterministic two-item auctions
With nondecreasing valuation densities, the best mechanism is either two posted prices or one bundle price.
Uniform bands track quantile effects in network experiments
A Gaussian bootstrap approximates the whole curve, no limiting degree distribution required.
· “Uniform Inference on Quantile Effects under Network Interference”
One two-way link can suffice for consensus in any population
Consensus needs reciprocity between merged groups, not between pairs, so one seed dyad can absorb the whole society.
One Newton step gives uniform size control at the boundary
A one-step corrected estimator keeps Wald and LR tests at nominal size even on the parameter-space boundary.
· “Uniformly Valid Inference Under Interactive and High-Dimensional Constraints”
18% debt relief or 10% remittances could offset half of aid losses
But 18 countries are locked out of debt relief and 16 out of remittance mobilisation—and they are different countries.
Click goes to the product with no close visual twin
It rises with query fit, doubles when text is uninformative, and recruits clicks that convert less often.
· “Contextual Visual Distinctiveness in Online Product Search”
Exact rejection sampling now practical for non-Gaussian state paths
Tangent-line twisted proposals give a computable sharp bound, so exact smoothing draws stay practical at long horizons.
· “Exact Rejection Sampling for Non-Gaussian State Space Models”
Better Netflix algorithms shift viewing toward the middle tail
A 60-day test on 8.5 million subscribers shows algorithm upgrades lift engagement and cut superstar concentration.
· “Recommendation Quality and the Concentration of Consumption: Experimental Evidence from Netflix”
Synthetic benchmark embeds real ad-spend endogeneity for MMM validation
Spend reacts to budgets, promotions, TV bursts, and a bidding rule; true sales effects are known weekly.
· “A Synthetic Benchmark Dataset with Endogenous Marketing Spend for Validating Marketing Mix Models”
Differencing geo-experiments recovers true MMM parameters
Instead of one ROAS number, the full time series identifies adstock, saturation, and effectiveness.
· “Structural Estimation of Marketing Mix Model Parameters from Geo-Experiments”
Only one-coordinate 'dictators' survive unit changes
Standard multivariate objective functions would fail a natural unit-change invariance, so their solutions may depend on arbitrary scales.
· “Characterizations of continuous adequate objective functions for ordinal or interval scaled data”
A formula turns ad randomization into a privacy budget
Firms can set one number that caps how far a tracker's beliefs about a visitor can move.
· “A Privacy Budgeting Framework for Online Experimentation”
Contracts jump, must jump, or cannot jump: one curve decides
The optimum is continuous, forced to jump, or chooses to jump — a single geometric meeting decides which.
· “Monotone Allocations without Single-Crossing: When to Bunch and When to Jump”
Plurality is the unique rule passing a one-rank clone test
Anonymity, neutrality, consistency, and 1-clone invariance hold together only for Plurality.
Two-voter coalitions already decide which set-valued rules survive
Pessimistic voters force single-valued medians; optimistic and best-worst voters permit exactly target-set rules.
Joint family trips get their own model in transport simulations
A pipeline classifies joint tours, picks their mode, and assigns drivers, so HOV and fare policies can finally be tested.
· “Accounting for intra-household joint travel in agent-based transport simulations”
Best AI player is not best AI coach in five of seven work tasks
Automation rankings diverge from assistant rankings; leaderboards miss a separate capability.
· “CentaurBench: Benchmarking LLM Capabilities on Augmenting vs. Automating Real-World Work Tasks”
Stable matchings exist under peer-dependent preferences
A cutoff supply-and-demand picture yields existence in large markets and approximate stability in finite samples.
· “Stable Matching with Peer-Dependent Preferences: Existence and Cutoff Characterization”
Path-conditional DiD identifies direct effects without parallel trends
Compare treated and untreated units with the same mediator path to separate direct from indirect policy effects.
· “Difference-in-Differences Models in the Presence of Time-Varying Mediators”
Slight noise uncertainty makes long-chain social learning impossible
With infinitely many sources, a Bayesian receiver still learns nothing from distant chains if mutation rates are only slightly uncertain.
Level-0 beliefs and actions diverge in the Extreme 11-20 game
Both depart from the standard 20-point anchor, recasting the 19/20 cluster as low-level reasoning rather than level-9.
Grand coalition, once reached, is never left
Every absorbing state is grand at every discount factor; failure means cycling forever between non-grand states.
Work-related AI use shows where humans still steer
From personal to work use, AI prompts carry more direction; chat modes show iterative control more than APIs.
One demand process carries signaling
Equilibrium spending rules are computable by path-integral control, and simulated revenues track two midstream firms.
· “Bayesian Signaling and Entry Decisions under Uncertain Market Conditions”
Messy democracy is rewriting the mathematics of collective choice
Voting, participatory budgeting, and citizens' assemblies now demand data-grounded theory and practical algorithms.
Correcting job-search optimism can backfire
A single reemployment-expectation question can tell pessimists who need encouragement from optimists whom true information may demotivate.
· “Biases-Informed Job Search Guidance: Characterization, Implications, and Targeting Support”
A modest researcher shift toward big metros can raise output
Production losses are small; the break-even access elasticity is far below existing estimates.
· “The Geography of Research: The Trade-Off Between Knowledge Production and Access”
Crop choices expose U.S
Portfolio-style analysis of planting decisions yields the first large-scale historical risk-aversion map and links it to tractor adoption.
· “The Yeoman's Portfolio: Measuring Historical Risk Preferences Using Crop Choice”
Housing rules passing three axioms share TTC's outcome envelope
Even mechanisms that are not strategy-proof give every agent the same best and worst houses as top trading cycles.
· “Non-obvious Manipulability with Groups in Shapley-Scarf Housing Markets”
Three route-choice models share one soft Bellman equation
Recursive logit, dynamic discrete choice, and maximum-entropy IRL are formally connected — yet estimate different objects.
Variance-reducing survey weights are eigenvectors of one matrix
Hajek, AIPW, and TMLE all sit on the same inverse-probability basis; the optimal basis beats them.
· “Optimal Control Variates for Survey Sampling and Causal Inference”
Same financial path, different felt time
Why care: financial observables alone would not reveal how time feels.
· “A Neurofinance Framework for Subjective Temporal Perception, Risk, and Investment Behavior”
Handovers need only preserve the task's predictive state
A theory bounds the bits a session handover needs and the regression error a memory cap costs.
· “Handover of In-Context Learning State Across Session Boundaries”
Lottery complexity is distance from the nearest sure outcome
A new axiom system recovers the exact metric behind perceived complexity, so 'twice as complex' becomes meaningful.
Censoring failed searches restores the classic cutoff rule
Bayesian searchers who only see their wins keep a myopic stopping rule; full information usually destroys it.
Capital release in crises lifts GDP growth one standard deviation
Regime-dependent estimates for Hungary find the buffer works: release helps, build-up costs little.
· “Macroprudential Policy and Downside Risk: Regime-Dependent Effects of Capital Regulation”
Acquiring irrelevant information can be the rational voter's defense
When a future self may be swayed by scandal, worst-case optimal disclosure separates irrelevant states.
Word blacklist cut flagged topics in federally funded economics papers
Economists at universities most reliant on US federal money cut politically flagged wording after 2025.
Same owner means same news: +0.33 coverage similarity
Co-owned outlet pairs also report tone 0.342 more alike, even after fixed effects.
Submission fees only cut referee workload if authors quit or move down
A 500-journal census shows fees are a residual price on top of waiting time, not a cure for congestion.
Six axioms characterize persuasion under worst-case leakage
A sender who fears extra information is modeled by a leakage set; new axioms make that model testable.
Dilutable debt turns interbank networks into automatic bail-ins
Long-term, dilutable interbank debt implements the optimal triage of shocked banks with no regulator and no bailout.
· “Systemic Risk in Financial Networks Revisited: Debt Dilution as a Backdoor Bail-in”
One household block can be estimated in isolation
Identified shocks serve as instruments, so misspecification elsewhere cannot contaminate block-level estimates
· “Limited-Information Estimation of Heterogeneous Agent Models”
Brazil's instant payments lifted small-firm wages and cut inequality
Gains went to the bottom half of the wage distribution, opposite the usual skill-biased pattern.
· “Financial Technologies, Labor Markets, and Wage Inequality: Evidence from Instant Payment Systems”
One latent draw per step is enough to match maximum likelihood
The paper proves the averaged stochastic-score estimator inherits the MLE's limiting distribution and sandwich inference.
· “Scalable likelihood-based inference for limited dependent variable models”
Minimax trial error is 1/n minus 1.617/n^{4/3}
A Bernoulli design with nonlinear shrinkage is the optimal procedure, beating complete randomization for small trials.
