Weekly AI Finance Paper Digest (Nov 1‑7 2025)

DeltaLag: Learning Dynamic Lead‑Lag Patterns in Financial Markets

DeltaLag is the first end‑to‑end deep‑learning model that discovers dynamic lead‑lag structures with specific lag values and uses them for portfolio construction. It employs a sparse cross‑attention mechanism to identify relevant lead‑lag pairs, then extracts raw features from the leading stocks aligned with the lagged ones to predict future returns of the lagged stocks. Empirical evaluation on real‑world market data shows DeltaLag substantially outperforms fixed‑lag and self‑lead‑lag baselines, as well as statistically pre‑computed lead‑lag graphs, and beats a range of time‑series and spatio‑temporal deep models, delivering better trading performance and enhanced interpretability (Zhou et al., 2025).

MS‑HGFN: Gated Fusion Enhanced Multi‑Scale Hierarchical Graph Convolutional Network for Stock Movement Prediction

MS‑HGFN addresses two overlooked aspects of multi‑scale graph neural networks for stock prediction: (1) subtle intra‑stock attribute patterns that affect inter‑stock correlations, and (2) bias toward coarse‑ or fine‑grained features during multi‑scale sampling. The model introduces a hierarchical GNN module that builds dynamic graphs from internal attribute and feature patterns across time scales, capturing spatio‑temporal dependencies. A top‑down gating mechanism fuses multi‑scale spatio‑temporal features while preserving essential coarse and fine information. Experiments on U.S. and Chinese stock datasets show MS‑HGFN improves prediction accuracy by up to 1.4 % over traditional and state‑of‑the‑art baselines and yields more stable return simulations (Xue et al., 2025).

LiveTradeBench: Seeking Real‑World Alpha with Large Language Models

LiveTradeBench is a real‑time trading environment designed to evaluate LLM agents under continuously changing market conditions. It follows three principles: (1) streaming live market prices and news to avoid offline back‑testing and information leakage; (2) abstracting portfolio management to multi‑asset allocation, integrating risk management and cross‑asset reasoning; (3) assessing agents across heterogeneous markets, namely U.S. equities and the Polymarket prediction market, which differ in volatility, liquidity, and information flow. At each step the agent observes price, news, and its portfolio, then outputs a risk‑adjusted allocation. A 50‑day live evaluation of 21 LLM series reveals that high LMArena scores do not guarantee superior trading results, models exhibit distinct portfolio styles reflecting risk preferences, and some LLMs can effectively adapt decisions using real‑time signals (Yu et al., 2025).