Quantitative Finance Paper Digest: Key AI‑Driven Research Highlights (Feb 21‑27 2026)

Bayesian Parametric Portfolio Policies (Miguel C. Herculano) propose Bayesian Parametric Portfolio Policies (BPPP) that place priors on policy coefficients to correct the risk‑blindness of standard PPPs. The authors prove a strictly positive utility gap between PPP and BPPP proportional to posterior uncertainty and signal strength. In a mean‑variance approximation this adds an extra risk term to portfolio variance, causing PPP to over‑expose when signals are strong and risk aversion high. Empirically, using 242 signals and six factors over 1973‑2023, BPPP achieves higher Sharpe ratios, lower turnover, greater returns and reduced tail risk, with advantages increasing with risk aversion and peaking during crises.

Signed Network Models for Dimensionality Reduction of Portfolio Optimization (Bibhas Adhikari) introduces a time‑series‑based signed‑network model that constructs a full signed graph each trading day, where edge signs reflect relative log‑return behavior against average returns. By incorporating higher‑order moments, the authors show that maximizing skewness and minimizing kurtosis correspond to specific signed‑graph configurations, leading to an NP‑hard combinatorial problem for the former and a naturally satisfied property for the latter. They formulate a mean‑variance optimization with a hedging‑score metric for dimensionality reduction. Backtesting on 199 S&P 500 assets from 2006‑2021 validates the framework’s effectiveness for both Markowitz and equal‑weight strategies.

Toward Expert Investment Teams: A Multi‑Agent LLM System with Fine‑Grained Trading Tasks (Kunihiro Miyazaki et al.) address the limitation of coarse‑grained instructions in existing multi‑agent LLM trading systems. The proposed framework decomposes investment analysis into fine‑grained tasks, improving reasoning and decision transparency. In a controlled backtest on Japanese equities using price, fundamentals, news and macro data, the fine‑grained design yields significantly higher risk‑adjusted returns than coarse‑grained baselines. Further analysis shows that alignment between intermediate agent outputs and downstream decision preferences drives performance, and portfolio optimization exploiting low correlation with indices enhances results.

Overreaction as an Indicator for Momentum in Algorithmic Trading: A Case of AAPL Stocks (Szymon Lis et al.) investigates whether short‑term market overreaction can be systematically predicted using high‑frequency sentiment and machine‑learning models. The authors build an intraday dataset for AAPL, extracting transformer‑based sentiment from Twitter and volatility‑normalized returns. Overreaction is defined as extreme returns relative to contemporaneous volatility and transaction costs, modeled as a three‑class prediction problem. Experiments with XGBoost, Random Forest, deep neural networks and bidirectional LSTM across 1‑, 5‑, 10‑ and 15‑minute intervals show that ML models outperform a baseline overreaction rule at ultra‑short horizons, while classic behavioral momentum dominates at ~10‑minute frequencies. SHAP analysis highlights volatility and negative emotions (fear, sadness) as key predictors.

Janus‑Q: End‑to‑End Event‑Driven Trading via Hierarchical‑Gated Reward Modeling (Xiang Li et al.) tackles two challenges: lack of large event‑centric datasets and misalignment between language‑model reasoning and financially effective actions. Janus‑Q builds a two‑stage pipeline: (1) constructing a 62,400‑article financial news event dataset annotated with ten fine‑grained event types, associated stocks, sentiment labels and cumulative abnormal returns; (2) decision‑oriented fine‑tuning using supervised and reinforcement learning guided by a Hierarchical‑Gated Reward Model (HGRM) that balances multiple trading objectives. Experiments demonstrate that Janus‑Q surpasses market indices and strong LLM baselines, improving Sharpe ratio by 102 % and directional accuracy by over 17.5 %.

FinAnchor: Aligned Multi‑Model Representations for Financial Prediction (Zirui He et al.) proposes FinAnchor, a lightweight framework that aligns embeddings from multiple LLMs without fine‑tuning the base models. By selecting an anchor embedding space and learning linear mappings to align other models, the approach resolves feature‑space incompatibility. The aligned representations are aggregated into a unified vector for downstream prediction. Across several financial NLP tasks, FinAnchor consistently outperforms strong single‑model baselines and standard ensemble methods, demonstrating the effectiveness of anchored heterogeneous representations for robust financial forecasting.