AIQF Daily digest — 8 new papers

Quant Paper Radar · 4 of 8

1. 1

   From pixels to profits: trading arbitrage portfolios based on image representations

   Niklas Paluszkiewicz · OPENALEX · CNN-based Statistical Arbitrage

   we backtested it — Sharpe -0.62 · quant relevance 1.00

   The study uses CNNs to rank stocks by next-month factor-hedged residual returns, using RGB images of stock, hedge, and residual paths. Its strongest reported specification earns 0.82% gross monthly excess return with a 0.97 annualized Sharpe ratio; profitability depends on hedge estimation and trading costs.

   original paper

2. 2

   Certifying Reward Alignment in Portfolio Reinforcement Learning

   Matin Tavakoli, Seyedreza Tavakoli, Faezeh Salmani et al. · ARXIV · Portfolio Optimization / Reinforcement Learning

   quant relevance 0.96

   The paper develops a reward-alignment audit for mean–variance portfolio reinforcement learning, separating reward misspecification from actor and critic errors. Controlled experiments improve decision alignment, while historical US industry-portfolio tests preserve fitted allocation targets without improving realised investment performance.

   original paper

3. 3

   Optimal collective investment under regulatory uncertainty

   An Chen, Christian Laudagé, Shihao Zhu · ARXIV · Constrained Portfolio Optimization

   quant relevance 0.92

   The study derives optimal collective portfolio policies under a weighted probability constraint covering multiple terminal-wealth thresholds. It directly addresses investment allocation and downside protection, but its numerical welfare results are model-based rather than measured trading performance.

   original paper

The rest of the batch

• Tracking High-Frequency Trading Activity in China and Its Effects on Market Quality

  Li Guo, Frank Weikai Li, Ding Qi et al. · ARXIV · relevance 0.93

  The study identifies latency-sensitive trading through limit orders priced exactly at the prevailing opposite-side best quote and finds predictive information for mid-quote returns over 100 milliseconds to ten minutes.

and 4 more · see the full radar