AI × QuantPPR-01
Stochastic Pruning of the DPO Computation Graph: Faster Training and Lower Peak Memory
The study randomly samples and prunes response positions, response prefixes and preference branches in the DPO computation graph before the corresponding operations, skipping computation at each training stage. Its novelty lies in linking resource savings to changes in the training objective and using inverse-probability weighting for a conditionally unbiased single-branch gradient estimator, which is useful for practitioners concerned with DPO training efficiency.
#direct preference optimization#computation graph pruning#training efficiency
Quantitative ResearchPPR-02
Launch-Time Contract Information Predicts Trading Volume and Dispute Risk in Prediction Markets
Using a near-census of resolved Polymarket markets, this study predicts lifetime trading volume and settlement disputes solely from listing-time contract information (title, description, resolution source, metadata). It assesses interpretable features versus neural text models and documents performance degradation under platform change. The work offers a contract-text-based screening approach for market designers and researchers, but dispute prediction precision is too low fo
#prediction markets#trading volume prediction#dispute risk
AI × QuantPPR-03
JAX-Accelerated Multi-Agent Macro Model Trains in Seconds
New research rewrites the MARL-BC macroeconomic simulator in JAX, achieving 75–724x faster training and scaling to 200,000 independent learning households, cutting calibration experiments from over a month to two hours.
#multi-agent reinforcement learning#macroeconomics#JAX
金融研究PPR-04
GPU-Parallel Gabriel Graph Construction: Up to 8484x Speedup
Proposes CUDA parallelization strategies—implicit vertex-pair mapping, bitmask-based atomically compacted edge lists, and thread coarsening—reducing Gabriel graph construction time by two to nearly four orders of magnitude versus a sequential CPU baseline, without loss of classification accuracy on 14 public benchmarks, and enabling a clustering method that requires no explicit density radius.
#Gabriel graph#GPU parallelization#CUDA
金融研究PPR-05
PRISM: A Training-Free Financial Retrieval Framework Improving Recall and Cost Efficiency
PRISM proposes a training-free framework for financial document retrieval that combines prompt engineering, in-context learning, and a candidate refinement module. It improves Recall@100 on FiQA-2018 and achieves comparable end-to-end NDCG@5 at lower cost on FinAgentBench, while releasing an expert-annotated evaluation set and a 16-variant label-robustness sweep.
#financial retrieval#training-free#prompt engineering
AI × QuantPPR-06
Time-Series Foundation Models Enhance Financial Crime Detection
The study proposes TCEE, a framework that uses a frozen time-series foundation model to summarize account transaction histories and enrich edge representations with temporal context, improving PR-AUC and training speed on multiple anti-money laundering datasets, offering a new approach for financial crime detection.
#financial crime detection#time-series foundation models#graph neural networks
AI × QuantPPR-07
Lineage-Aware Search Control Improves Diversity and Predictiveness in LLM-Based Evolutionary Factor Mining
The study argues that conventional LLM-based evolutionary factor mining lacks lineage-level control, leaving search weakly directed and factors structurally similar. It proposes LineageEvo, which introduces lineage-conditioned LLM variation and lineage-stratified budget allocation to make semantic variation more direction-aware and prevent branch dominance, thereby improving factor predictiveness and lineage/expression diversity on CSI 500 and S&P 500.
#LLM-based evolutionary factor mining#lineage-aware search#symbolic factor mining
AI 研究PPR-08
Dynamic Scenario Synthesis Framework Improves LLM Agent Adaptability
DynaWeave automatically transforms static tasks into dynamic scenarios with exogenous events, generates verifiable interaction trajectories for supervised fine-tuning, and aims to improve LLM agents' adaptability in changing environments. Experiments show consistent performance gains across multiple benchmarks, offering a practical new approach for agent training.
#LLM agents#dynamic scenario synthesis#supervised fine-tuning
金融研究PPR-09
State-Space Foundation Model for Limit Order Books Encodes Decodable Order-Flow Imbalance
This study examines an 8.2M-parameter Mamba-3 state-space model trained autoregressively on limit order book message streams from eight US equities, finding that order-flow imbalance is linearly decodable from its internal state at R²≈0.59 and that this representation transfers to unseen stocks and extends to traded volume and volatility. It offers empirical clues for interpretability and controllability of market-simulation foundation models.
#limit order book#state-space model#order-flow imbalance
金融研究PPR-10
Quant-Arena: An End-to-End Benchmark for LLM Quantitative Research on Real A-Share Data
Existing benchmarks score formulas, code, or single trading decisions, leaving open whether LLMs can complete the full quantitative research loop from hypothesis to executable signal. Quant-Arena provides a benchmark and public leaderboard built on real A-share data: models construct features and stock-scoring code, receive feedback only on 2022–2023, freeze one strategy, and are assessed on the sealed year 2024 with trading constraints and costs, including market impact esti
#quantitative trading#LLM agents#benchmark evaluation
Quantitative ResearchPPR-11
Structural Features of Dynamic Portfolio Selection under RVaR with Logarithmic Returns
The study transforms a continuous-time mean-RVaR portfolio problem into a static quantile problem and characterizes optimal terminal wealth via a convex envelope, revealing state-dependent payoff structures induced by RVaR risk constraints. Its novelty lies in identifying two geometric configurations of the RVaR-specific convex envelope and limiting relationships to VaR and ES, offering a theoretical tool for portfolio design under interval-based quantile risk control.
#RVaR#dynamic portfolio#quantile optimization