AI for Finance: Quantum Asset Clustering, Causal Market Troughs, Multimodal Forecasting, Diffusion SDEs

Toward Quantum Utility in Finance: A Robust Data‑Driven Algorithm for Asset Clustering

The paper addresses asset clustering based on return correlation, which is essential for portfolio optimization and statistical arbitrage. Traditional clustering methods require lossy transformations of signed correlation matrices and often assume a fixed number of clusters. The authors propose the Graph‑Based Community‑Structure Generation algorithm (GCS‑Q), which directly clusters signed, weighted graphs without preprocessing. Each partition step is formulated as a Quadratic Unconstrained Binary Optimization (QUBO) problem, enabling quantum annealing to explore an exponential solution space. Experiments on synthetic data and real‑world financial datasets compare GCS‑Q with SPONGE and k‑Medoids. Results show that GCS‑Q dynamically determines the number of clusters and consistently achieves higher clustering quality, measured by the Adjusted Rand Index and a structural‑balance penalty metric.

Predicting Market Troughs: A Machine‑Learning Approach with Causal Interpretation

The study applies a flexible Double‑Machine‑Learning (DML) average partial‑effect framework to identify causal drivers of market bottoms. Robust DML estimates reveal that option‑implied risk aversion and market‑liquidity volatility are key causal factors, which simpler linear models either misinterpret or fail to detect. The causal analysis is implemented in a high‑performance instant‑prediction model that leverages the identified triggers, providing high‑frequency empirical support for asset‑pricing theory.

FinZero: Launching Multi‑modal Financial Time‑Series Forecast with Large Reasoning Model

The authors construct a multimodal financial image‑text dataset (FVLDB) and introduce Uncertainty‑Adjusted Group Relative Policy Optimization (UARPO), a reinforcement‑learning fine‑tuning method that enables a model to output both predictions and associated uncertainty estimates. FinZero, a multimodal pretrained model, is fine‑tuned with UARPO on FVLDB for inference, prediction, and analysis. Extensive experiments demonstrate strong adaptability and scalability. After UARPO fine‑tuning, FinZero improves high‑confidence prediction accuracy by approximately 13.48 % relative to GPT‑4o, illustrating the effectiveness of RL‑based fine‑tuning for financial time‑series tasks.

Data‑driven Generative Simulation of SDEs Using Diffusion Models

The authors present a model‑free, data‑driven approach that employs conditional diffusion models to generate sample paths of unknown stochastic differential equations (SDEs). Given a finite set of observed SDE trajectories, the diffusion model learns to synthesize new paths of the same SDE without requiring explicit drift or diffusion coefficients. Simulation experiments compare the method with neural‑SDE and other baselines, confirming superior path generation quality. In an empirical study, the synthetic paths are used to augment a continuous‑time mean‑variance portfolio‑selection reinforcement‑learning algorithm, resulting in improved performance, thereby highlighting the potential of diffusion‑based generative simulation for financial analysis and decision‑making.