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.
