Distributions or Quantiles? A Convex Theory of Worst-Case Risk
WhyUnifies worst-case spectral risk analysis under a convex variational theory, deriving closed-form solutions for several ambiguity models with strong analytical support and moderate field impact.
Measuring High-Frequency Trading in China and Its Impact on Market Quality
WhyProposes an order-level HFT proxy and uses cross-channel latency variation in a difference-in-differences design, yielding strong novelty and credible causal evidence.
Tokenised Sukuk: a verifiable system for asset evidence and Shariah trading rules
WhyEncodes Shariah tradability rules as on-chain policies and compares them with identity-only baselines, revealing trade-offs, though evaluated only on a synthetic instrument.
Stars are a rough AI assessment based on the title and abstract: each paper is scored for novelty, strength of evidence and breadth of impact, then mapped to 1–5 stars (about 2% get five stars and 10% get four). The site only marks four- and five-star papers as Notable, with a one-line reason; no mark does not mean a paper is weak. The full text is not read, so treat the stars as a reading aid, not a verdict on quality. "Code" means the abstract states that code, data or tools are released.
In briefOptimal quote placement for market makers under a state-dependent Hawkes order flow model
SummaryThe paper studies how a market maker should place quotes when order flow follows a state-dependent Hawkes process. It links market-order counts to execution volumes through a block-based linear demand model and formulates a continuous-time, discrete-state dynamic program. Under a quadratic terminal inventory penalty, it derives inventory-dependent quotes and backward recursions for the value function coefficients. Numerical tests on LOBSTER data show higher aggregate terminal wealth than fixed-spread benchmarks at one and three ticks, but lower than the two-tick benchmark.
In briefPricing Catastrophe Bonds in Incomplete Markets Using Riemannian Geometry and Financial Entropy
SummaryThis paper proposes a framework for pricing catastrophe bonds in incomplete markets by optimizing financial entropy with Riemannian geometry to select the pricing measure. It uses a compound Poisson and generalized Pareto loss model with Wang distortion and derives closed-form expressions for entropy and the Riemannian metric. Applied to a hypothetical earthquake bond calibrated to Golden State Re II (Series 2014-1) spreads, the study compares risk-neutral and physical trigger intensities and splits the risk premium between frequency and severity components.
In briefDeposits Should Not Reduce Returns: Supplier Income Monotonicity and Rate Curve Design in Pooled Lending
SummaryThe paper studies utilization-based interest rate models in which adding liquidity can lower a supplier's total interest income even as its balance rises. It gives the condition under which a supplier's income does not decrease with a marginal deposit and derives the resulting constraint on admissible rate curves. The authors also check how often the problematic case occurred on two public Drift markets.
In briefSimple, low-complexity ETF portfolios for private investors
SummaryThis empirical study tests portfolio construction methods for private investors using liquid US ETFs, with monthly trading costs and no short selling or leverage. The portfolios are built as convex optimization problems using interpretable inputs estimated from public data. The authors report that simple portfolios of as few as three ETFs deliver consistent returns and outperform a standard 60/40 allocation over the past two decades.
In briefRisk measures for infinite-server queues driven by a Cox-BESQ process
SummaryThis paper proposes an infinite-server queue model whose arrival intensity follows a Cox process driven by a squared Bessel (BESQ) process. Using the affine structure of the dynamics, it derives closed-form joint moment-generating functions for the arrival intensity, the number of customers, and the time integral of the customer count, extendable to two or more dates. These explicit functions are used to construct risk measures, such as joint moments conditional on a variable exceeding a threshold, computed via Fourier representations. Numerical experiments illustrate the practical implementation of the method.
In briefEvent definition and branching ratio: Hawkes estimates of price-change clustering in consolidated quotes
SummaryThis paper fits Hawkes models to price changes in the consolidated US quote feed, where raw estimates show a sub-millisecond kernel and a near-critical branching ratio. The authors find that in large-tick stocks about half of NBBO changes are locked or depleted quotes resolved within a millisecond, which the model misreads as excitation. Defining events as a mid-price level persisting for one millisecond removes the sub-millisecond component.
In briefFinLLM-KG: Knowledge-Graph-Enhanced LLMs for Financial Sentiment in Algorithmic Trading
SummaryThis paper proposes FinLLM-KG, a framework that combines causal large language models with a financial knowledge graph built from SEC 10-K filings. It supplies company-level context at inference time without updating model weights, either as a profile or as relevance-ranked factual triples. Sentiment scores are attributed to individual sentences and graph facts to make predictions more explainable. Simulations reportedly show consistent portfolio performance gains over a state-of-the-art baseline.
Giorgos Iacovides, Oliver Hannibal, Wuyang Zhou et al.
In briefScenario matching for credit dependence stress tests with correlation constraints and empirical risk bounds
SummaryThe paper pairs two sets of simulated market and credit paths by minimizing the absolute error in temporal correlation relative to a target, with each market path matched to exactly one credit path so both full-scenario distributions are preserved. A perfect-matching test checks whether a uniform correlation-error tolerance is feasible and finds its smallest value, and further assignments yield exact empirical bounds on mean loss and exceedance probability. A synthetic bond-forward credit stress test illustrates matching precision and tail results. The method does not depend on the scenario generator and can be used to examine dependence assumptions in existing risk engines.
In briefDet-MASO: Diversity-Regularized Multi-Agent Sequential Optimization for Factor Generation
SummaryThis paper proposes Det-MASO, a framework for generating alpha factors in quantitative investing. It adds a differentiable log-determinant penalty on the inter-agent correlation matrix to the joint objective, discouraging agents from producing similar outputs while keeping computational cost linear in the number of assets. The authors report that the method outperforms baselines and improves the Sharpe ratio by 0.33.
In briefDo round-number prices act as barriers? Order book study of liquidity and price departures
SummaryPrices often cluster at round numbers, which is sometimes read as evidence those levels act as barriers. This paper uses limit order book data and a tick-grid framework to separate barrier effects from a mechanical explanation, where deeper resting liquidity makes prices slower to leave. Analyzing five Nasdaq stocks throughout 2025, it finds round levels attract more liquidity but do not show longer depth-adjusted durations. Intel and SoFi show directional effects, with departures biased away from round levels.
In briefChecking Whether Portfolio RL Agents Follow Their Intended Financial Goal
SummaryThis paper proposes a decision audit for portfolio reinforcement learning that separates reward misspecification from learning error, checking whether a learned policy implements its stated mean-variance objective. The method is tested in controlled markets and on historical US industry portfolios. The authors report that it quantifies whether a learned decision implements its financial objective, separately from backtest performance.
Matin Tavakoli, Seyedreza Tavakoli, Faezeh Salmani et al.
In briefWhat FLARE's Published Scores Actually Measure
SummaryThis paper audits the scoring pipeline of FLARE, the financial benchmark behind PIXIU and FinBen, and finds that in many cells the score depends on the answer-extraction rule rather than the model. On identical generations, two extraction rules differ in accuracy by up to 0.4216, and no leaderboard states which rule it used. The authors argue that reading a FLARE score requires ambiguity rate, coverage, and a majority-class floor, and they release floorbench to compute these from existing generations.
In briefOptimal investing when a green transition's timing is uncertain
SummaryThis paper studies an investor's optimal portfolio choice when a possible future green transition, whose timing is uncertain, affects stock prices. The transition is modeled as a jump in which green stocks rise in value and brown stocks fall. The investor is ambiguous about timing and designs a strategy robust to this risk, with preferences that penalize ambiguity. A case study with one green and one brown stock shows that a transition-aware investor shifts holdings toward the green stock and away from the brown stock.
In briefTokenised Sukuk: a verifiable system for asset evidence and Shariah trading rules
SummaryThe paper argues that security-token standards check only who holds a token, not what backs a sukuk, so an identity-only token can keep trading after non-debt assets are sold. The authors encode the 2023 Securities Commission Malaysia Shariah Advisory Council ruling on Sukuk Wakalah bi al-Istithmar as issuance and transfer policies that read signed asset evidence. In 26 scenarios on a synthetic instrument, the identity-only baseline wrongly allowed 13 operations that the full policy denies, while the full policy matched every expected outcome.
WhyEncodes Shariah tradability rules as on-chain policies and compares them with identity-only baselines, revealing trade-offs, though evaluated only on a synthetic instrument.
In briefTesting whether deep learning across many convertible bonds can beat simple price persistence for next-day forecasts
SummaryThe paper uses a daily panel of 73,806 observations from 100 Chinese A-share convertible bonds (2022–2025) to test whether cross-bond neural networks beat a random walk baseline. Under strict temporal splitting and dependence-aware inference, the price-change MLPs show slightly higher out-of-sample RMSE than the random walk, negative out-of-sample R², and non-significant test results. The authors conclude that without contemporaneous equity market information or explicit contractual event conditioning, deep panel models do not deliver next-day predictability beyond persistence.
In briefBuilding implied volatility surfaces and forecasting with survey sampling, tempered inverse-probability learning, and activity clocks
SummaryThis paper uses split-adjusted NVIDIA equity and options data from 2018 to 2025 to study how to compile implied volatility surfaces and forecast derivative risk. It combines finite-population survey sampling, tempered inverse-probability sequence learning, and causal activity clocks. The findings show that the sampling design should match the target measure, short-horizon prices are not forecastable beyond a random walk with drift among the tested models, and forward implied volatility shows descriptive mean-reversion that is not statistically significant under session-clustered tests.
In briefMeasuring High-Frequency Trading in China and Its Impact on Market Quality
SummaryThe paper proposes an order-level proxy for high-frequency trading based on limit orders priced exactly at the best opposite-side quote. Using Shenzhen Stock Exchange data, the authors validate the proxy through its links to quote responsiveness, cancellation speed, and mid-quote returns. A difference-in-differences design using order dissemination latency differences at market open finds that lower latency raises the share of such orders, which increases short-horizon price variation, wide-spread incidence, and adverse-selection exposure. Realized spreads decline and pricing error falls modestly.
WhyProposes an order-level HFT proxy and uses cross-channel latency variation in a difference-in-differences design, yielding strong novelty and credible causal evidence.
In briefContinuity of conditional distributions and stratified sampling for derivative pricing
SummaryStratified sampling reduces Monte Carlo estimator variance by partitioning the sample space using a low-dimensional factor and sampling within each stratum, which suits path-dependent claims in high-dimensional models. The paper studies constructing neural-network stratification factors for derivative pricing under proportional allocation in continuous-time models. It establishes continuity results for conditional expectations and distributions, and gives counterexamples showing the limits of this theory. Combined with neural network density results, it yields a universal approximation theorem for admissible factors, and numerical experiments show the learned factors outperform all benchmarks considered.
In briefOptimal collective investment when future regulations are uncertain
SummaryThis paper studies optimal collective investment under a Value-at-Risk-type constraint that reflects uncertainty about future regulatory requirements. The constraint includes several wealth thresholds and requires a minimum weighted average probability of meeting them. The authors characterize the optimal terminal wealth and show that welfare losses can be unevenly distributed among investors with different capital contributions.
In briefBalancing Risk Pricing, Coverage and Anti-Discrimination in Insurance Tariffs
SummaryThe paper proposes a framework for insurance tariff design built on risk classification, aggregation of benchmark risk prices within classes, and balanced transfers. It evaluates tariffs on risk adequacy, portfolio coverage, and anti-discrimination, and frames the trade-offs as a multi-objective problem with a Pareto frontier. The authors prove several mathematical characterizations of these trade-offs and illustrate the results with a numerical motor-insurance study.
In briefWhether digital investment and energy use are reshaping the Kitchin business cycle, tested with U.S. and Irish data
SummaryThis paper uses quarterly U.S. and Irish data from 1995 to 2025 to explore whether IT investment and data-centre electricity use are reshaping the Kitchin cycle around energy and compute assets. It reports a structural break in the U.S. IT-GDP correlation in 2014Q3, and finds that the strong Irish data-centre electricity and GDP correlation largely disappears when GNP replaces GDP. The authors state that all coefficients are exploratory given small samples.
Bartosz Ziollko, Adam Niedzialkowski, Bartosz Szyma
In briefContext-rich trading strategy using actor-critic deep reinforcement learning with sentiment for stock trading
SummaryThis paper proposes a deep reinforcement learning approach to stock trading that partially and deeply incorporates sentiment analysis to capture external factors such as social sentiment and political instability. It compares actor-critic agents using only policy optimization (A2C, PPO) with those combining policy optimization and Q-learning (DDPG, TD3, SAC). The authors report that agents combining both principles consistently improve in profit maximization and risk-averse trading when sentiment is deeply infused.
In briefSurvey on quantifying uncertainty in financial foundation models
SummaryThis survey examines uncertainty in financial foundation models used for question answering, disclosure analysis, forecasting, and decision-making. It proposes a source-aware framework that separates six uncertainty sources across evidence and context, model and inference, and outcome and action, and a mechanism-based taxonomy for assessing methods across financial tasks. The authors map finance-specific empirical studies onto this taxonomy. They find that most studies address single parts of a workflow, with little work linking evidence and reasoning reliability to downstream prediction and constrained action.
In briefOptimal order size for market makers under a state-dependent Hawkes process model
SummaryThe paper models a market maker's executed volume as the product of average order size, an exponential demand function, and a Prony-series multiplier that depends on quote depth, with market-order counts built from a state-dependent Hawkes model. An empirical study on limit-order-book data examines calibrated models and sensitivity to inventory restrictions. In retrospective evaluations conditioned on realised order-flow paths, the selected calibration achieves higher terminal utility than a baseline while respecting the inventory bound, and the authors separate the effects of trading revenue from inventory control.
In briefProspect Theory in LLMs: Meta-Analysis of Loss Aversion and Related Biases
SummaryThis paper meta-analyzes 102 fitted estimates from 11 studies covering 32 language models to test whether LLMs show the same prospect-theory properties as people. Loss aversion is modest, with a pooled λ of 1.26, far below the strongest human estimate. Probability weighting shows a human-like inverse S shape, but diminishing sensitivity depends on the measurement method used.
In briefIdentifying level-k reasoning in repeated games: strategies, beliefs, and cognitive ability
SummaryThis paper identifies level-k reasoning operating at the level of supergame strategies in repeated games. The authors build a model covering both strategy choices and beliefs about others' strategies. Using elicited strategies and strategy beliefs from the indefinitely repeated Prisoner's Dilemma and a new balance-optimal error tolerance measure, they classify a substantial share of subjects as level-1 or level-2. They also find that cognitive ability and experience predict higher-level reasoning.
In briefInstitutional Traders' Intraday Forecasting Skill and Liquidity Provision
SummaryThis study uses high-frequency exchange data to examine how institutional traders' forecasting ability and liquidity provision relate to their proprietary intraday profits. It finds persistent forecasting skill, with flow-following strategies based on these traders' executions yielding annualized Sharpe ratios of 16 to 63 across trading frequencies. Higher forecasting ability and more passive executions within a position are both associated with higher profit and loss per matched position.
In briefDirectionally Correct Expert Views in Portfolio Optimization
SummaryThis paper, inspired by the Black-Litterman framework, asks whether incorporating expert views improves portfolio allocation. In a simple Markowitz setting, the authors show that even a small amount of directionally correct information can materially improve investment decisions. They also discuss using entropy-based ideas to add expert views to end-to-end portfolio models. Empirically, simpler Markowitz-type models perform as well as an end-to-end model once directionally correct experts are included.
Ojas Kumar, Sreejit Dey, Sarvesh Ravichandran Iyer et al.
In briefMultichoice Modeling: Exploring Decision Configuration Space Under Uncertainty
SummaryThis paper proposes multichoice modeling, a framework that makes large spaces of interdependent decision configurations explicit and explorable under uncertainty. The authors treat such problems as contingent performance landscapes, drawing on rugged landscape logic, and combine variance-based sensitivity analysis with conditional distribution analysis in a single simulation. Three case studies (a manufacturing process upgrade, M&A strategy, and mortgage design) show how the approach reveals unexpected complementarities and near-best configurations with different risk profiles.
Mariia Kozlova, Roman Stepanov, Alexander Walzer et al.
In briefEquilibria and Learning in Competitive Market Making: A game-theoretic study of market makers and a non-strategic trader, proving pure-strategy Nash equilibria exist and examining learning under discrete prices.
SummaryThe paper proposes a market-making game among N ≥ 2 strategic market makers and one non-strategic representative trader. Because quotes are continuous and payoffs are non-convex and discontinuous, standard finite-game and classical continuous existence results do not apply. Under minimal assumptions on the asset value distribution, the authors show at least one pure-strategy Nash equilibrium always exists, and that the equilibrium bid-ask spread widens strictly with asset value volatility. With discretized prices and an added regularity condition, they show time-averaged play of no-internal-regret learners approaches the continuous Nash price envelope as tick size vanishes.
In briefJoules-to-GDP: AI Economic Value per Unit of Energy
SummaryThis paper measures how much economic value an AI system creates per joule of energy consumed, using two metrics: net dollar value and time saving. The authors study 220 public GDPval tasks across 44 occupations, running five open-weight models locally and evaluating four accelerator types. They find that occupations and models that rank highest on success, dollar value, or time saving do not necessarily deliver the greatest economic return per joule.
Amirreza Zeinali, Jon Saad-Falcon, Avanika Narayan et al.
In briefThe Delegation Divide: An Open Index of AI-Agent Delegability by Occupation and Country
SummaryThe paper introduces the Agent Delegability Index (ADI), in which two language models rate all 18,796 O*NET 30.2 task statements against a frozen, preregistered rubric of delegation levels and barriers. Tasks are mapped to ISCO-08 occupation groups and weighted by ILOSTAT employment across 146 countries. The authors find that delegable employment rises with income, from 22.5% in low-income to 48.7% in high-income countries, and that ADI correlates strongly with existing generative AI exposure measures at the occupation level.
WhyIndexes task delegability to AI agents across 146 countries, complementing exposure metrics, with human blind ratings and model agreement checks.
Cristian Espinal Maya (Universidad Nacional de Colombia), Santiago Jim\'enez Londo\~no (Universidad EAFIT)
In briefWhen News Gets Old: Portfolio Allocation with Reinforcement Learning and LLMs
SummaryThis paper studies how aging financial news affects portfolio decisions, comparing keeping old news with timing information against filtering it out before selection. A fixed large language model extracts signals from public SEC disclosures, and these treatments are tested in reinforcement learning policies and LLMs that select portfolios directly. Over historical tests from 2023 to 2025, measured after trading costs, filtering old news gave the LLM the highest average net return of 18.43% per year when news delivery was delayed by four weeks, exceeding all RL strategies that used news. RL choices responded to timing information in a controlled test, but adding it showed no consistent benefit in historical returns.
Bolin Song (Cornell University), Surui Xiang (United Nations)
In briefGoverned self-evolution for autonomous investing: a multi-agent AI system with a protected core
SummaryThe paper presents Steward, a multi-agent LLM system for autonomous investing with nine specialised agents, a seven-layer shared memory built by code, and nightly and quarterly self-review. Self-edits are blocked from an invariant core covering order arithmetic, risk rules and discipline-agent prompts. Reported results include a five-month live deployment over 102 trading days and a 2015–2026 backtest.
In briefModeling how liquidity is distributed in concentrated-liquidity trading pools over time
SummaryThe paper models how liquidity is spread across price ranges in Uniswap V3 and similar pools, using a stochastic partial integro-differential equation (SPIDE). The equation combines an operator that keeps liquidity concentrated near the market price with an integral operator capturing redistribution from providers repositioning. The authors prove a mild solution exists and is unique, and reduce the model to linear stochastic equations for level, slope, and curvature. Calibration on three Uniswap pools shows lower active-band forecast error than a martingale benchmark on two pools consistent with the stationary specification, while the third pool's out-of-sample results suggest nonstationarity.
In briefLineageEvo: a method that uses factor evolution lineages to guide LLM-based mining of market prediction factors.
SummaryThis paper addresses LLM-based evolutionary mining of formulaic alpha factors, where search tends to concentrate on a few structurally similar factors. The authors propose LineageEvo, which controls search at the lineage level by conditioning LLM variation on lineage-derived operator priors and by allocating search budget across lineages to limit concentration. Experiments on CSI 500 and S&P 500 report that LineageEvo achieves the strongest predictive performance on both markets, while LSPS yields a more balanced lineage distribution and broader factor pool.
In briefDistributions or Quantiles? A Convex Theory of Worst-Case Risk
SummaryThis paper develops a convex variational theory for worst-case risk evaluation using spectral risk measures, unifying and extending existing analyses. It uses a representation principle that picks the quantile or distribution formulation preserving convexity of the ambiguity set, and optimality conditions identify the worst-case law and its risk value. The theory yields closed-form or semi-closed-form solutions for several models, including Wasserstein and Kullback–Leibler ambiguity sets, and supplies projection-based algorithms with fixed-grid convergence. In a high-dimensional portfolio application, an exact projection identity under a Gaussian reference reduces robust portfolio selection to a second-order cone program.
WhyUnifies worst-case spectral risk analysis under a convex variational theory, deriving closed-form solutions for several ambiguity models with strong analytical support and moderate field impact.
Jianjun Gao, Jonathan Yu-Meng Li, Zhen Wang et al.
In briefWhat launch-time contract details reveal about prediction-market trading volume and disputes
SummaryThe paper analyzes 802,863 resolved Polymarket markets listed from January 2023 to May 2026, using only information available at listing, such as titles, descriptions, resolution sources, and metadata. Contract information predicts trading volume well, with market topic contributing the most. Performance drops sharply when platform changes and new listing formats appear in later periods. Dispute risk is ranked better than chance but is too imprecise for predicting individual disputes.
In briefJaxMARL-BC: Fast Multi-Agent Reinforcement Learning for Macroeconomic Models
SummaryThis paper presents JaxMARL-BC, a JAX-accelerated simulator that trains MARL-BC economies 75 to 724 times faster than the reference implementation and scales to 200,000 independent learning households. The authors verify its correctness against known DSGE limiting cases and demonstrate it through a proof-of-concept calibration of an extended model with heterogeneous returns on capital, completed in two hours rather than over a month. The tool is released as open source.
Daniele Montagnani, Andrea Coletta, Antonietta Mira et al.
In briefDSINC method extended beyond Heston to 3/2 and other non-affine stochastic-volatility models
SummaryThis paper extends the damped-sinc (DSINC) method for pricing American options from the affine Heston class to general one-factor stochastic-volatility models. It shows that the collapse identity, which reduces early-exercise integration to a single spectral integral, still holds for these models. Tests on 3/2, SABR, GARCH and lognormal-volatility models agree with fine finite-difference benchmarks to within one basis point of the strike.
In briefHow a state-space model for limit order books internally represents market dynamics
SummaryThe study examines what an 8.2M-parameter Mamba-3 state-space model learns after being trained autoregressively on tokenized limit order book message streams from eight large-cap US stocks. The authors report that order-flow imbalance is decodable from the model's internal state with R² ≈ 0.59, and that this representation extends to unseen stocks and to traded volume and volatility. They also report that steering the internal state shifts the model's next-message predictions in line with the implied feature.
Valentin Mohl, Alexandre Bismuth, Aramis Marti-Shahandeh et al.
SummaryThe paper proposes a framework that separates changes in the number of tasks an occupation performs from changes in occupational uniqueness, meaning how distinct its task mix is relative to other occupations. Using 196,563 task statements from the Dictionary of Occupational Titles and O*NET covering 1977 to 2018, the authors find that occupations becoming less unique are associated with lower wage growth. The wage associations of task additions and removals depend on how widely those tasks are performed elsewhere.
In briefSpeculative Imbalance Positions by Non-Physical Traders and Their Effect on Great Britain's Electricity Balancing
SummaryThis paper provides the first empirical analysis of speculative imbalance positions taken by non-physical traders in Great Britain's electricity system from 2021 to 2024. Using counterfactual simulations that remove the aggregate non-physical trader imbalance, the authors measure effects on balancing volumes, balancing costs and supplier imbalance costs. They find these positions reduced balancing costs by 5.9 to 6.7 percent and supplier imbalance costs by 65 to 75 percent, suggesting consumer bills would have been higher without them.
In briefArithmetic Models for Pricing Commodity Futures, Vanilla, and Calendar Spread Options
SummaryThe paper proposes an arithmetic multi-factor stochastic volatility model for commodity futures and options, giving a low-dimensional Markovian representation of the futures curve and semi-closed-form prices for vanilla, basket, and calendar spread options. It is estimated with an unscented Kalman filter on WTI crude oil futures, vanilla options, and calendar spread options from 2013 to 2024. The authors find three volatility factors are needed, and that including calendar spread options improves the model's treatment of dependence between adjacent futures contracts. The arithmetic model is also reported to fit better than its geometric counterpart.
In briefStronger AI breakthroughs accelerate the exhaustion of their own ideas
SummaryThe paper analyzes 36 years of computer science, including 7.2 million papers, 73,089 peer reviews and 3.7 million researchers. It finds that attention narrowed sharply after the Transformer, and that each major breakthrough was followed by steeper exhaustion of ideas. The authors call for institutions that protect diverse attention in AI.
WhyAnalyzing millions of papers, reviews and researchers, it shows breakthroughs deplete diverse attention, corroborated by structural biology and simulations.
In briefDeep catastrophic-risk-aware reinforcement learning: training neural policies with extreme value theory to control rare-event losses
SummaryThis paper extends an Extreme Value Theory approach for CVaR optimization to high-dimensional neural-network policies. It uses a peaks-over-threshold representation and method-of-moments estimation to obtain a differentiable tail-risk objective whose finite-sample gradient can be computed via automatic differentiation. In a deep-hedging study with rare extreme market moves, EVT-trained policies behaved similarly to empirical-CVaR-trained ones under normal conditions but had lower hedging losses in extreme-scenario evaluation.
Carlos Octavio P\'erez Mendoza, Fr\'ed\'eric Godin
In briefLOB-Impact Benchmark: Market Impact Emerging in Generative Order Book Models
SummaryThis paper introduces the LOB-Impact benchmark for testing whether generative limit order book models reproduce realistic market impact. The authors inject execution-style metaorders into order flow generated by five autoregressive models across five NASDAQ stocks and measure price responses using empirical impact estimators. Models trained only for next-token prediction show emergent impact, with exponents between 0.63 and 1.21, slightly above the square-root law's 0.5. No impact decay during relaxation was observed, which the authors tentatively attribute to short context windows; the benchmark is released as open source.
George Nigmatulin, Valentin Mohl, Sascha Frey et al.
In briefExact risk analysis of teacher training, fixed features, and calibration for candidate-scoring models
SummaryThe paper studies three training choices that determine what a candidate-scoring decision model's probabilities mean: soft targets from a teacher model, a frozen pretrained backbone, and temperature calibration. It analyzes these in Gaussian surrogate models where finite-sample risk can be computed exactly, and tests them on 231,521 CFPB credit card complaints about whether a complaint ends in monetary relief. The authors find that a student inherits a teacher's bias, and that calibration corrects only the probability scale, so probabilities must be checked against realized outcomes.
In briefMulti-agent reinforcement learning for market making on Polymarket order books
SummaryThis paper studies market makers trained with multi-agent reinforcement learning on Polymarket, a prediction market with thin order books. The authors reconstruct order-level LOBSTER-format data from Polymarket and train agents on 417 real markets, splitting the maker into a quoter and a risk manager with reward shaping. Their redesigned single-agent maker quotes more often and has smaller held-out losses than the retrained baseline. However, none of the configurations achieves positive average portfolio value in periodic held-out evaluation, and each was trained with one seed.
In briefProbability-weighting functions from return distribution changes: three orthogonal channels and information-theoretic indices
SummaryThe paper builds probability-weighting functions from changes between prior and posterior normal tempered-stable return distributions rather than a parametric family. Scale, skewness and tail thickness are isolated one at a time with variance held fixed, each producing a distinct sign pattern in the deviation from identity. Three information-theoretic indices summarize the distortions, and the framework is applied to S&P 500 ETF daily returns from 2015 to 2025.
Akash Deep, Bhathiya Divelgama, A. Alexandre Trindade et al.
