Key Quantitative Finance Papers from WWW2025 – Summaries & Insights

CSPO: Cross‑Market Synergistic Stock Price Movement Forecasting

Paper: https://arxiv.org/pdf/2503.22740 Code: https://github.com/Lamsoda1123/CSPO Problem: Stock price prediction is complicated by (1) stock exogeneity – each stock reacts to market factors differently, and (2) volatility heterogeneity – volatility varies across stocks and time.

Method: A deep neural architecture that (a) incorporates external futures market knowledge to enrich stock embeddings via cross‑market information, and (b) introduces a pseudo‑volatility term that models per‑stock prediction confidence, allowing the loss to be dynamically weighted during optimization.

Results: Extensive experiments on industry benchmark datasets show CSPO achieves higher prediction accuracy and robustness than prior methods; ablation studies confirm the contribution of both the cross‑market embedding and the pseudo‑volatility module.

MERA: Mixture of Experts with Retrieval‑Augmented Representation for Modeling Diversified Stock Patterns

Paper: https://dl.acm.org/doi/pdf/10.1145/3701716.3715513 Code: https://github.com/chenchen1104/MERA Problem: Existing deep‑learning stock‑prediction models degrade because they treat all market patterns uniformly, ignoring the diversity of patterns.

Method: MERA builds a Mixture‑of‑Experts (MoE) architecture. A backbone learns coarse‑grained representations of all stocks. Independent expert networks specialize on distinct patterns. A GateNet dynamically routes each input to the most suitable expert. Since explicit pattern labels are unavailable, MERA retrieves semantically similar samples using self‑supervised pretrained embeddings; neighbor labels serve as weak signals for pattern discrimination.

Results: Large‑scale experiments on real‑world stock markets demonstrate significant performance gains over baselines. MERA is designed as a plug‑and‑play module that can be attached to any deep‑learning‑based stock predictor.

Ploutos: Towards Explainable Stock Movement Prediction with a Financial Large Language Model

Paper: https://dl.acm.org/doi/pdf/10.1145/3701716.3715254 Code: https://github.com/hanstong/Ploutos Problem: Traditional deep‑learning approaches struggle to fuse textual news and numerical market data and lack interpretability.

Method: The framework consists of two components. PloutosGen is a multi‑expert module that processes multimodal inputs (e.g., news text, price series) and produces trading strategies. PloutosGPT consumes the strategies and generates natural‑language rationales that explain the predictions.

Results: Empirical evaluation shows the combined system outperforms existing baselines in both prediction accuracy and explainability metrics.

DINS: Domain‑Informed Negative Sampling Strategies for Dynamic Graph Embedding in Meme‑Related Stock Social Networks

Paper: https://dl.acm.org/doi/pdf/10.1145/3696410.3714650 Code: https://github.com/YunmingHui/DINS Problem: Dynamic graph embedding (DGE) models rely on negative sampling strategies originally designed for static link prediction, which fail to capture the distinctive interaction patterns of meme‑stock communities on platforms such as Reddit.

Method: An analysis of real meme‑stock social graphs and financial domain knowledge motivates three novel negative‑sampling heuristics (e.g., sampling temporally distant but topically similar non‑edges, avoiding sampling within identified meme clusters). These strategies are integrated into standard DGE training pipelines.

Results: Experiments demonstrate that the domain‑informed samplers improve downstream tasks such as link prediction and stock return forecasting on meme‑stock datasets compared with conventional uniform negative sampling.

HedgeAgents: A Balanced‑Aware Multi‑Agent Financial Trading System

Paper: https://arxiv.org/pdf/2502.13165 Problem: LLM‑driven and agent‑based trading systems can incur large losses during rapid market declines or high‑volatility periods.

Method: HedgeAgents introduces a hierarchical multi‑agent architecture comprising a fund‑manager agent and multiple hedging‑expert agents. The agents exchange information through three coordination meetings that leverage LLM reasoning to decide when to activate hedges.

Results: Over a three‑year back‑test, the system achieves a 70 % annualized return and a 400 % cumulative return, comparable to top human fund managers.

HRFT: Mining High‑Frequency Risk Factor Collections End‑to‑End via Transformer

Paper: https://dl.acm.org/doi/pdf/10.1145/3701716.3715235 Code: https://github.com/wencyxu/IRF-LLM-accepted-at-WWW25- Problem: Existing neural approaches extract latent risk factors but do not produce explicit, formulaic expressions that are interpretable for traders.

Method: The Intraday Risk Factor Transformer (IRFT) treats symbolic mathematics as a language modeling task. It uses a hybrid symbolic‑numeric vocabulary where symbols encode operators and stock features, and numeric tokens represent constants. The Transformer generates complete risk‑factor formulas (including constants). Predicted constants are refined with the Broyden‑Fletcher‑Goldfarb‑Shanno (BFGS) optimizer to mitigate non‑linearity.

Results: Trained on high‑frequency trading (HFT) datasets, IRFT outperforms ten SRBench baselines on the HS300 and S&P 500 datasets, delivering roughly 30 % higher investment returns and inference speeds that are orders of magnitude faster than prior methods.

Pre‑Finetuning with Impact Duration Awareness for Stock Movement Prediction

Paper: https://arxiv.org/pdf/2409.17419 Problem: Most text‑based stock‑movement models ignore the temporal duration of a news event’s impact on markets.

Method: The authors release the Impact Duration Estimation Dataset (IDED), which annotates news articles with investor‑perceived impact intervals. Language models are pre‑finetuned on IDED to learn impact‑duration representations.

Results: Compared with sentiment‑based pre‑finetuning, IDED‑pre‑finetuned models achieve statistically significant improvements in stock‑movement prediction accuracy, demonstrating the value of modeling impact duration.