In briefFactorBench: A Portfolio-Aware Benchmark for Automated Factor Mining
SummaryThis paper introduces FactorBench, a benchmark that compares about 5,000 factors produced by nine automated mining methods across five equity markets. It evaluates factor validity, temporal generalization, predictiveness beyond risk and style exposures, pool distinctness including similarity to Alpha101, and the quality of composite signals and after-cost portfolio performance. The authors find that no single mining paradigm consistently outperforms the others.
In briefRobust enhanced index tracking portfolio selection under uncertain return distributions
SummaryThis paper studies building portfolios that aim to beat a benchmark index when the true joint distribution of asset and index losses is unknown. The authors propose robust models, guided by Pareto-optimality, that minimize a convex combination of worst-case mean loss and downside risk under a worst-case mean return constraint. Empirical tests on real market data report better out-of-sample cumulative wealth and Sharpe, Sortino, and Omega ratios than the tracked index, an equal-weighted portfolio, and other robust models.
In briefCities of Signals: Compression and Novelty Geometry Explained
SummaryThe paper treats each daily cross-sectional forecast as a point on a sphere and defines its "city" as the direction of its time-averaged vector, a compressed summary of signal history. It proves probability bounds on when the angle between two cities conservatively estimates the angle between their histories. The authors test these results on positive-IC signal pairs from AlphaNova's May 2026 competition, finding agreement that is mostly positive but not fully explained by the uniform null model.
In briefAgent Policy-Value Audit: Splitting Trade Mix from Event Picking
SummaryThe paper argues that financial LLM agents are often judged by return differences against a baseline, which cannot tell event-selection skill apart from passive market exposure. The authors propose an audit that keeps each type of action change at its observed count and randomly reassigns them across eligible events, using the average payoff as a composition benchmark. Applied to 723 earnings events at 44 U.S. consumer-facing firms, the agent's gross value of +15.2 bps per event splits into a +25.8 composition benchmark and a -10.6 selection value, showing no detectable outperformance over matched random assignment.
In briefWhen a Correct Reward Is Not Enough: Diagnosing and Guiding PPO in a Solved Broker-Trader Game
SummaryThis paper places a proximal policy optimisation (PPO) agent in an analytically solved continuous-time broker-trader game, where it chooses the broker's trading speed. With stochastic uninformed order flow, the tested PPO networks remain inaccurate, and their critics do not reliably rank nearby actions. Under partial information, a causal certainty-equivalent controller based on observable history stays closer to the reference than PPO. Freezing the analytical policy and training PPO to adjust it after a cost change closes 2.22% of the gap to the changed-cost reference.
Siu Tung Wong (1), Carlo Campajola (1, 2) ((1) Institute of Finance et al.
