Quantitative Finance Paper Digest (Sep 27 – Oct 10 2025)

STRAPSim: A Portfolio Similarity Metric for ETF Alignment and Portfolio Trades

Paper link: http://arxiv.org/pdf/2509.24151v1

STRAPSim (Semantic, Two‑level, Residual‑Aware Portfolio Similarity) matches portfolio components by semantic similarity, weights each match by its proportion in the portfolio, and aggregates the weighted matches using a greedy residual‑aware alignment. The method is benchmarked against Jaccard, weighted Jaccard, and BERTScore‑based variants on public classification, regression, and recommendation tasks as well as a corporate‑bond ETF dataset. Empirical results show STRAPSim achieves higher prediction accuracy, better ranking alignment, and the highest Spearman correlation for return‑based similarity, while remaining scalable and interpretable.

Bayesian Portfolio Optimization by Predictive Synthesis

Paper link: http://arxiv.org/pdf/2510.07180v1

The authors address the absence of known asset‑return distributions by applying Bayesian Predictive Synthesis (BPS), a Bayesian ensemble method for meta‑learning. Assuming investors have access to multiple return‑prediction models, BPS combines these forecasts through a dynamic linear model, producing a posterior distribution over mean returns that adapts to market uncertainty. Experiments construct mean‑variance and quantile‑based portfolios using the predictive distribution, demonstrating robustness to changing market conditions.

From Headlines to Holdings: Deep Learning for Smarter Portfolio Decisions

Paper link: http://arxiv.org/pdf/2509.24144v1

An end‑to‑end framework integrates three modules: (1) an LSTM to capture temporal price patterns, (2) a Graph Attention Network (GAT) to model evolving inter‑stock relationships, and (3) a sentiment‑analysis component that extracts market‑psychology signals from financial news. The combined network directly outputs daily portfolio weights, bypassing the traditional predict‑then‑optimize pipeline. Evaluation on nine U.S. stocks across six industries shows higher cumulative returns and Sharpe ratios than equal‑weight and CAPM‑based mean‑variance baselines.

The New Quant: A Survey of Large Language Models in Financial Prediction and Trading

Paper link: http://arxiv.org/pdf/2510.05533v1

The survey classifies over fifty studies of LLMs for stock‑return prediction and trading into a task‑centric taxonomy: sentiment/event extraction, numeric/economic reasoning, multimodal understanding, retrieval‑augmented generation, time‑series prompting, and tool‑based pipelines for backtesting and execution. It reviews empirical evidence of predictability, highlights design patterns such as retrieval‑first prompting and numeric‑tool verification, and discusses challenges including time leakage, hallucination, data coverage, deployment economics, interpretability, governance, and safety. Recommendations include standardized evaluation, auditable workflows, and multilingual, cross‑market research.

An Adaptive Multi‑Agent Bitcoin Trading System

Paper link: http://arxiv.org/pdf/2510.08068v1

The system builds specialized LLM agents for technical analysis, sentiment assessment, decision making, and performance reflection. A verbal feedback loop lets a reflective agent critique daily and weekly trading decisions in natural language; the critiques are injected into future prompts, allowing the system to adjust metric priorities, sentiment weights, and allocation logic without parameter updates or fine‑tuning. Backtesting on Bitcoin from July 2024 to April 2025 shows the quantitative agent delivers >30% excess return in bull markets (15% higher than buy‑and‑hold), the sentiment‑driven agent converts flat‑market losses into >100% gains, and weekly feedback improves total performance by 31% while reducing bear‑market losses by 10%.

Fuzzformer: A Neuro‑Fuzzy System for Interpretable Long‑Term Stock Market Forecasting

Paper link: http://arxiv.org/pdf/2510.00960v1

Fuzzformer combines multi‑head self‑attention with a fuzzy inference system. An LSTM and temporal attention compress multivariate inputs into fuzzy‑compatible features, which are then processed by the fuzzy module. Tested on the S&P 500, Fuzzformer matches ARIMA and LSTM prediction performance while providing interpretable internal information flow.

Hermes: Multi‑Scale Spatial‑Temporal Hypergraph Network with Lead‑Lag Structures

Paper link: http://arxiv.org/pdf/2509.23668v1

Hermes addresses two limitations of existing hypergraph‑based stock forecasting methods: (1) insufficient modeling of lead‑lag interactions between industries, and (2) lack of multi‑scale information. It introduces a moving‑aggregation module that uses a sliding window and dynamic time aggregation to capture lead‑lag dependencies, and a cross‑scale edge‑to‑edge message‑passing layer to fuse multi‑scale features while preserving consistency. Experiments on several real‑world stock datasets show Hermes outperforms state‑of‑the‑art baselines in both efficiency and accuracy.

Signature‑Informed Transformer (SIT) for Asset Allocation

Paper link: http://arxiv.org/pdf/2510.03129v1

Code link: https://github.com/Yoontae6719/Signature-Informed-Transformer-For-Asset-Allocation

SIT directly optimizes a risk‑aware financial objective by enriching asset dynamics with path signatures and embedding financial inductive biases (e.g., lead‑lag effects) via a signature‑enhanced attention mechanism. Evaluation on daily S&P 100 data shows SIT significantly outperforms traditional and deep‑learning baselines, especially compared with predict‑then‑optimize pipelines.

IKNet: Interpretable Stock Price Prediction via Keyword‑Guided Integration of News and Technical Indicators

Paper link: http://arxiv.org/pdf/2510.07661v1

IKNet extracts keywords using a FinBERT‑based contextual analyzer, projects each keyword embedding through a separate nonlinear layer, and fuses the resulting representations with technical‑indicator time series to predict next‑day closing prices. Shapley Additive Explanations provide keyword‑level attribution. On S&P 500 data (2015‑2024), IKNet reduces RMSE by up to 32.9% and improves cumulative returns by 18.5% relative to baseline RNN and transformer models.

Agent‑Based Genetic Algorithm for Crypto Trading Strategy Optimization (CGA‑Agent)

Paper link: http://arxiv.org/pdf/2510.07943v1

CGA‑Agent combines a genetic algorithm with a multi‑agent coordination mechanism. Real‑time market micro‑structure intelligence and adaptive performance feedback guide the evolutionary process, enabling dynamic optimization of trading‑strategy parameters in volatile, non‑stationary crypto markets. Backtests on three cryptocurrencies demonstrate statistically significant improvements in total return and risk‑adjusted metrics compared with static optimization methods.

Multi‑Agent Analysis of Off‑Exchange Public Information for Cryptocurrency Market Trend Prediction

Paper link: http://arxiv.org/pdf/2510.08268v1

The framework introduces (1) a theoretically guaranteed news‑analysis system that uses LLMs to quantify market impact, regulatory influence, volume dynamics, risk, technical relevance, and temporal effects; (2) an adaptive volatility‑conditioned fusion mechanism that dynamically combines news sentiment with technical indicators; and (3) a low‑communication‑complexity multi‑agent coordination architecture for real‑time heterogeneous data processing. Experiments on Bitcoin across three prediction horizons show statistically significant performance gains over recent NLP baselines.