Weekly Quantitative Finance Paper Digest (Aug 30 – Sep 5, 2025)

Robust MCVaR Portfolio Optimization with Ellipsoidal Support and RKHS‑Based Uncertainty

The authors propose a portfolio‑optimization framework that minimizes Mixed Conditional Value‑at‑Risk (MCVaR) while imposing an opportunity constraint on expected return and a cardinality constraint on the number of assets. The robust MCVaR model assumes that random returns lie within an ellipsoidal support set, avoiding any distributional assumption. To handle the opportunity constraint, the model incorporates an uncertainty set defined in a Reproducing Kernel Hilbert Space (RKHS), which reduces the problem to a second‑order cone program (SOCP). Experiments on six financial‑market datasets show that the robust model outperforms the nominal model, the market portfolio, and an equal‑weight portfolio in most cases, delivering higher expected returns, lower risk metrics, better risk‑return ratios, and higher Jensen’s α. The authors also evaluate the model across market regimes (bull, bear, neutral); it provides markedly better risk protection in bear markets, while its performance in bull and neutral periods is comparable to the nominal model.

Adaptive Alpha Weighting with PPO: Enhancing Prompt‑Based LLM‑Generated Alphas in Quant Trading

This paper introduces a reinforcement‑learning framework that uses Proximal Policy Optimization (PPO) to dynamically adjust the weights of multiple formulaic alphas generated by large language models (LLMs). Formulaic alphas are mathematically defined trading signals derived from price, volume, sentiment, and other data. Although recent work demonstrates that LLMs can produce diverse effective alphas, integrating them adaptively across varying market conditions remains challenging. The authors employ the deepseek‑r1‑distill‑llama‑70b model to generate 50 alphas for five major stocks (Apple, HSBC, Pepsi, Toyota, Tencent). PPO then optimizes the alpha weights in real time. Experimental results indicate that the PPO‑optimized strategy achieves strong returns and high Sharpe ratios on most stocks, surpassing an equal‑weight alpha portfolio and traditional benchmarks such as the Nikkei 225, S&P 500, and Hang Seng indices. The study underscores the value of reinforcement learning for alpha allocation and showcases the potential of combining LLM‑generated signals with adaptive optimization for robust financial prediction and trading.

Is All the Information in the Price? LLM Embeddings versus the EMH in Stock Clustering

The authors examine whether artificial intelligence can improve stock clustering beyond traditional price‑based methods. Operating under the semi‑strong Efficient Market Hypothesis (EMH), which asserts that prices fully reflect all public information, they benchmark three clustering approaches: (i) price‑based clustering using historical return correlations, (ii) human‑information clustering defined by the Global Industry Classification Standard (GICS), and (iii) AI‑driven clustering built from large‑language‑model embeddings of news‑headline text. Each method produces a daily classification that assigns every stock to a cluster. To evaluate the clusters, the authors translate them into a synthetic factor model within the Arbitrage Pricing Theory (APT) framework, enabling consistent out‑of‑sample, rolling‑window testing. Using S&P 500 constituents from 2022‑2024, they find that price‑based clustering consistently outperforms both the rule‑based GICS clustering (15.9 % lower RMSE) and the LLM‑embedding clustering (14.7 % lower RMSE). The paper contributes a generic conversion method for any stock grouping, the first direct head‑to‑head comparison of price, rule‑based, and AI clustering under identical conditions, and empirical evidence that short‑term return information is largely captured by price, supporting the EMH while providing a practical diagnostic tool for practitioners and a testing framework for scholars.

Exploring Diffusion Models for Generative Forecasting of Financial Charts

Recent progress in generative models has yielded impressive results in text‑to‑image, video generation, and related tasks, prompting exploration of their applicability to finance. While financial research typically focuses on time‑series transformation models, this work treats a time‑series as a single image pattern and leverages a text‑to‑image diffusion model to forecast stock‑price trends. Unlike prior studies that employ ResNet or Vision‑Transformer architectures to learn and classify chart patterns, the authors attempt to generate the next chart image directly from the current chart image and a textual instruction prompt using a diffusion model. They also propose a straightforward evaluation method that compares the generated chart image with the real future chart image. The approach opens a new direction for generative forecasting in quantitative finance.