In briefCombinatorial Arbitrage in Prediction Markets: a study of arbitrage across logically related prediction-market contracts
SummaryThe paper develops a framework to identify and quantify arbitrage among logically related prediction-market contracts, proving that prices outside the convex hull of admissible payoff vectors admit arbitrage and using integer programming to find optimal arbitrage portfolios. Using venue metadata and semantic analysis of resolution criteria, the authors recover 1.9 million logical relations covering more than 255,000 contracts. Replaying full-depth Polymarket order books, they find that an ideal arbitrageur earns a risk-free return of 3.5% a week on settled portfolios after spreads and fees. Opportunities persist longer and attract less observed participation where relation discovery and execution are more costly.
In briefWhere Data Value Comes From: Consumer Disclosure, Access Design, and Platform Monetization
SummaryThe paper develops a game-theoretic model of a commissioned platform, two differentiated sellers, and privacy-sensitive consumers to study how consumer disclosure rights shape sellers' data purchases and the platform's incentive to provide data. Under mandatory collection, the platform can monetize defensive adoption even when bilateral personalization reduces seller revenue. Under voluntary disclosure with anonymous purchasing and positive privacy costs, unilateral disclosure is unattractive in the baseline. The authors also discuss welfare effects and suggest capping anonymous offers at their no-data benchmark.
In briefBasis-Only Funding Is Not Dividend Fair: Settlement Design for Crypto Equity Perpetuals
SummaryThe paper examines perpetual futures on dividend-paying equities listed on crypto venues. It argues that the common premium-index funding mechanism, which sets funding from the basis, cannot replicate shareholder cash flows for traders entering and exiting at arbitrary times, because funding serves both basis control and settlement. A predictable dividend behaves as a measure-theoretic atom, so fairness requires a discrete event-based settlement. Exchange-native data are cited to show one venue posts such a discrete funding payment at ex-dividend dates while a standard venue does not.
In briefReading News Without Hindsight: NoLBERT Topics and Stock Returns
SummaryThe paper builds news topics with NoLBERT, an encoder pretrained only on 1976–1995 text and adapted by contrastive fine-tuning, and compares it with LDA, MiniLM and FinBERT readers. The topics are used to forecast returns for the 1,000 largest US stocks over 2022–2024, a held-out period after training on 2013–2019 and validation on 2020–2021. The best validation-selected experiment loses money at every rebalancing frequency. An equal-weight combination of all readers and models earns a monthly net Sharpe ratio of 1.12, significant against zero but not against a price-only benchmark.
WhyUses an encoder pretrained only on 1976-1995 text to avoid hindsight, with strict out-of-sample tests across 1.26 million articles.
Kevin Foley, Jonathan Hartadi, Shivesh Prakash et al.
In briefOrderFlowDiT: A diffusion model that forecasts limit-order supply, with better price and volume accuracy
SummaryThe paper studies how much limit-order volume will be submitted at each price over the next five minutes, given the current order book and recent order flow. It introduces OrderFlowDiT, a conditional Diffusion Transformer trained with an unbalanced optimal-transport auxiliary loss, and a full-field evaluation framework that scores price placement, total volume, and uncertainty together. On 1,950 validation conditions across 30 U.S. equities, adding the transport loss cut full-field transport cost by 45.8% and the transport energy score by 26.1% versus mean squared error alone, with both improving for every stock.
In briefHow momentum, commentary tone, and earnings surprise shape LLM financial forecasts
SummaryThe paper builds synthetic earnings announcements that independently vary 20-day momentum, uninformative commentary tone, and whether actual EPS beats a manipulated consensus. It tests six instruction-tuned language model checkpoints from four families on 2,016 prompts each. The authors measure matched changes in conditional answer-prefix scores rather than calibrated market probabilities. Tone and momentum often exceed the EPS contrast, though answer-encoding and other controls qualify these results. The study measures observable input reliance and numerical recognition, not internal reasoning or real return predictability.
