Focused on AI applications in the fintech sector
This article reviews the PandaAI framework, a closed‑loop neural‑symbolic LLM agent that models market regimes, applies constrained Monte‑Carlo Tree Search for factor generation, and continuously adapts via back‑test feedback, achieving significantly higher Rank IC and lower drawdown on CSI‑300 data.
The article analyzes cross‑sectional stock prediction as a ranking task, identifies the prediction head as a bottleneck, introduces the RankGLU residual bottleneck GLU module that preserves a linear scoring path while adding bounded multiplicative interaction, and demonstrates an 11% IC improvement on CSI300 with extensive experiments and ablations.
The paper proposes a two‑stage Retrieve‑then‑Reason framework that first builds a sparse candidate graph from 10‑K text embeddings and then uses a large language model to filter edges for true economic relationships, resulting in a higher‑Sharpe, lower‑drawdown cross‑stock trading signal on S&P 500 constituents.
AlphaCrafter tackles the non‑stationary nature of financial markets by integrating LLM‑driven factor mining, market‑aware factor screening, and risk‑constrained execution into a closed‑loop multi‑agent system, and experiments on CSI 300 and S&P 500 demonstrate consistently higher risk‑adjusted returns, lower variance, and robust performance compared with five baseline methods.
The paper introduces MSH‑LLM, a multi‑scale hypergraph framework that aligns natural language and time‑series modalities via a cross‑modal alignment module and mixed prompts, achieving state‑of‑the‑art performance on 27 real‑world datasets across forecasting, classification, few‑shot and zero‑shot tasks.
The paper introduces FactorEngine, a program‑level factor mining system that injects domain knowledge from financial reports via LLM‑guided pipelines, separates logic correction, parameter optimization, and LLM usage, and demonstrates superior information‑ratio, Sharpe, and annualized returns over GP, neural, and other agent‑based baselines on CSI300 and CSI500 markets.
The article reviews the TimeART framework, which equips large language models with 21 ready‑to‑use time‑series analysis tools and a four‑stage training regime on the 100k‑trajectory TimeToolBench corpus, enabling an 8B Qwen‑3 model to act as a fully autonomous data scientist and achieve state‑of‑the‑art performance on multiple TSQA, prediction, and reasoning benchmarks.
This paper builds a multi‑scale candlestick‑chart dataset and a standardized evaluation framework to measure how well visual language models (VLMs) extract price information, using confusion‑matrix diagnostics and Information Coefficient (IC) metrics, and finds that VLMs excel only on monotonic trends and struggle with precise time‑based predictions.
TimeGMM introduces an adaptive Gaussian‑mixture‑model framework with reversible instance normalization, a dual‑branch time encoder, and a conditional decoder, achieving up to 22.48 % improvement in CRPS and 21.23 % in NMAE over state‑of‑the‑art probabilistic forecasting methods across multiple benchmark datasets.
ProbFM introduces a Transformer‑based framework that leverages deep evidential regression to separate epistemic and aleatoric uncertainty in time‑series forecasting, and demonstrates on cryptocurrency returns that this decomposition retains competitive prediction accuracy while enabling risk‑aware trading strategies with superior risk‑adjusted returns.
