In briefRetrospective Analysis of Team-Following Trading in Pumpfun Meme Coins via Funding Networks
SummaryThis paper examines how much following token-creation teams on Pumpfun could have returned across ten separately recorded windows from April 2024 to September 2026. Each window is rebuilt from real SOL reserve states and finite accounts starting with one SOL, with teams defined as directed funding groups identified in hindsight. Results are strongest when exit thresholds are chosen with hindsight, while using only information available online yields weaker outcomes. The authors state these are conditional retrospective model accounts, not executable profits, and not evidence of a causal effect of the SOL price.
In briefOptimal trade execution with reinforcement learning for fractional price dynamics
SummaryThis paper studies optimal liquidation with linear market impact, following the Almgren and Chriss framework, but with asset prices driven by fractional Brownian motion. For Hurst parameter H not equal to 1/2, the deterministic Almgren and Chriss solution is shown to no longer be optimal. The authors propose a strategy that exploits correlations between price increments, derived from a neural-network-parameterized policy gradient algorithm trained to maximise CARA utility, which yields significant improvements.
In briefJUBILEE: A Debt-Collections Benchmark for Off-Policy Evaluation
SummaryThis paper presents JUBILEE, a benchmark environment built on a debt-collections simulator that generates million-loan panels with competing risks, rare recovery events, and hidden factors affecting both outcomes and assignment. The authors compare tabular and gradient-boosted estimators against reference values from independent policy simulations. When hidden factors influence assignment, several tabular methods underestimate the field-visit policy's score, while randomized decisions show no comparable large error.
Diego Silva Prado, Tiago Amaral Penteado de Jesus, Bruno Caldas de Souza et al.
In briefRisk-sensitive reinforcement learning for options market makers: an empirical study with real S&P 500 options
SummaryThis paper studies how market makers hedge option inventory risk, comparing a reinforcement learning hedging agent trained on historical market data against delta hedging, simple hedge-ratio rules, and agents trained in simulated markets. The authors build an empirical replay environment using S&P 500 option data from 1996 to 2017 and evaluate out of sample over 2018 to 2023. The agent trained on real data reduces Conditional Value-at-Risk tail loss by 3.1% relative to delta hedging, reaching 8.9% in volatile falling markets but producing 7-10% higher costs in rising markets.
In briefWinning reviews, losing discovery? A co-evolutionary game of AI research and peer review
SummaryThis paper uses a two-population evolutionary game to study the feedback between autonomous research systems and automated reviewers. Authors choose between acceptance-oriented and exploratory research, while reviewers choose between proxy evaluation and careful assessment. The authors derive conditions under which a stable regime rewards publication success while lowering the scientific value of accepted work. Numerical illustrations show how publication counts can hide scientific loss.
