In briefCollusion Without Deterrence: Auditing Punishment in Q-Learning Pricing
SummaryThe paper replicates a simulated Bertrand duopoly in which independent Q-learning pricing agents converge to above-competitive prices. Auditing frozen policies, the authors find the reward-punishment explanation does not hold: a single attractor covers most of the state space, and re-pairing agents collapses coordination by 80%. They also report that the one-period undercut is unprofitable in 92% of pairs, but optimal deviations earn 17.5% on average, so collusion persists largely because deviation is never attempted.
WhyFreezing policies and computing optimal deviations by dynamic programming challenges the standard punishment-mechanism account of algorithmic collusion, with broad evidence across many converged pairs
In briefSix Gaps Between Backtests and Live Trading in an Intraday System
SummaryThe paper reports six defects found by comparing a US equity intraday trading system's live days with its backtest replay. Three stem from data issues, two from code, and one from configuration. For each defect, the authors describe the symptom, error size, cause, and the test or control now used to prevent it. The paper makes no claim about whether the system is profitable.
In briefHow language model agents change price competition
SummaryThis paper studies LLM pricing agents in repeated Bertrand competition, a standard benchmark for algorithmic pricing. LLMs adapt faster than standard online learners, but this advantage largely disappears against full-information learners. Depending on deployment choices, some agents can sustain supracompetitive prices while others move toward the competitive region, producing dynamic pricing regimes.
WhyShows LLM pricing agents escape Bertrand Nash outcomes depending on deployment, a novel finding, though evidence is mainly simulation-based and limited in scale.
Martin Bichler, Philip Hierhager, Kassian K\"ock et al.
In briefWhen Does “Post-Call” Begin? An Availability Contract for Earnings-Call NLP Benchmarks
SummaryThe paper argues that financial NLP benchmarks can suffer a timing mismatch within events, where the prediction target starts before the text or audio input is complete. It proposes an Availability Contract requiring benchmarks to state the input boundary, availability evidence, outcome start, and supported claims. Applied to 552 earnings calls from 24 firms, the authors show that changing the timing proxy flips return signs for 9.4% of calls and changes quintiles for 28.3%.
In briefHow much screening can offset slower quote updates for liquidity providers
SummaryThe paper studies the spread a liquidity provider needs to break even when counterparties can exploit stale quotes, comparing faster repricing with screening of informed flow on one scale. Empirically, across four USDC-WETH pools benchmarked against Binance ETH-USDT, one-second price innovations are sharply non-Gaussian, with 73% of seconds showing no price change. Moving from twelve-second to one-second cadence sheds 84.5% of the expected shock, so a twelve-second maker must screen at least that share of informed flow to match.
WhyPlaces repricing speed and informed-flow screening on one break-even scale, validated on four real pools with non-Gaussian data.
Aleksei Adadurov, Eduard Airapetian, Anton Chtepine et al.
In briefValuing American commodity options with a two-factor price model and an efficient numerical scheme
SummaryThis paper studies how to value American commodity options under the Schwartz two-factor model. It uses a double-sweep alternating direction explicit (ADE) finite difference method and compares it with Saul'yev's single-sweep variant. The authors derive the governing PDE, map the semi-infinite domain to a rectangle, and set boundary conditions using Fichera theory. The solver is checked against a closed-form European solution and Monte Carlo, then benchmarked against Longstaff-Schwartz Monte Carlo for accuracy and convergence. Following feedback from a commodities trader, the framework was extended to exercise on futures prices and to use seasonal parameters.
In briefStochLOB: Robust limit order book trend prediction using heavy-tailed noise in training
SummaryThis paper proposes StochLOB, a training-only score-matching framework that replaces the usual Gaussian noise in diffusion models with jump-Lévy and clipped α-stable noise to better reflect heavy-tailed, bursty limit order book order flow. The deployed predictor remains deterministic and needs no reverse-time diffusion sampling at inference. On chronologically held-out Coinbase LOB data, the variants stay competitive on uncorrupted BTC and altcoin horizons and improve several tail-window stress diagnostics. An exploratory A-share transfer study on Feishu data is also included.
Arshia Abolghasemi, Seyed Reza Tavakoli, Parsa Naderi et al.
In briefLoss-versus-rebalancing with external transaction costs
SummaryThis paper studies liquidity provision in automated market makers when trading on external markets incurs proportional transaction costs. It compares the pool with a self-financing benchmark that holds the same risky inventory and trades at external bid and ask prices. Under continuous prices and pure arbitrage, the two have identical liquidation wealth whenever the pool fee or external spread is positive. The paper also constructs pool designs that implement decreasing inventory schedules and constant no-trade boundaries.
In briefRegime Detection in Latent Factor Models: A Bayesian Decision-Theoretic Approach
SummaryThis paper studies when an agent should act on inferred regime changes in models with latent factors and hidden regimes. It frames online regime detection as a Bayesian decision problem that balances false alarms, detection delay, and misclassification, solved approximately with particle filtering and Monte Carlo dynamic programming. In numerical experiments on Heston and regime-switching Vasicek models, the latent-information procedure outperforms feasible plug-in rules.
