AI × QuantPPR-01
Multi-Turn Tool-Using LLM Agents Enable Adaptive Alpha Mining
Existing formulaic alpha mining methods typically exploit only a subset of signals among market text, historical data, and quantitative evaluation. AlphaPilot integrates market descriptions, LLM financial knowledge, real-data evaluator feedback, and a retained factor pool within a multi-turn reinforcement learning framework, and performs day-by-day search at inference to adapt to shifting market regimes, thereby jointly leveraging multiple signal types for alpha mining.
#LLM Agents#Alpha Mining#Reinforcement Learning
金融研究PPR-02
Code-Interactive Candlestick Reasoning: Fin-Thinker Makes Evidence Verifiable
Existing financial candlestick QA often measures answer accuracy alone, without checking whether predictions are supported by chart evidence. This work introduces code-interactive financial candlestick reasoning: the model generates and executes code to obtain observations and perform computations, then uses returned results to guide subsequent reasoning. It also builds 32K replay-verified trajectories and a 1,200-question benchmark that separately assesses answer correctness
#financial candlestick understanding#code-interactive reasoning#multimodal large models
Quantitative ResearchPPR-03
Structured Continuous-Time SDEs Improve Realized GARCH Integrated-Variance Forecasting
The study proposes a structured continuous-time stochastic volatility model that embeds a realized-measure state directly into latent variance dynamics and captures return-linked dependence via correlated Brownian innovations. Across 17 equity indices, it reports stronger out-of-sample performance than several GARCH-family and HAR-type benchmarks under QLIKE, MSE and MAE; ablation indicates realized-measure feedback provides substantial incremental value while return-linked B
#continuous-time model#stochastic volatility#Realized GARCH
Quantitative ResearchPPR-04
Rough Martingale Optimal Transport: Finite Pricing Bounds for Non-Modelable Risk Factors under Sparse Data
Addressing the infinite audit bounds that classical Martingale Optimal Transport yields for non-modelable risk factors with sparse market data, this work proposes a Rough Martingale Optimal Transport framework that regularizes the transport plan with a rough volatility prior to obtain finite and asymptotically tight extrapolation bounds. Its novelty lies in identifiability theorems for rough volatility parameters and local identifiability of multi-asset correlation; its usefu
#rough volatility#martingale optimal transport#non-modelable risk factors
AI 研究PPR-05
TextualRL: Textual Reinforcement Learning for Context Management to Improve Agent Performance
TextualRL introduces a gradient-free reinforcement learning framework that treats multiple rollouts from the same task under the same context as the unit of policy critique, generates bounded textual patches through groupwise and cross-group review, and decides adoption via held-out validation rewards. It offers a new way to turn execution rewards into reusable context updates without fine-tuning model parameters, potentially improving LLM agents.
#LLM agents#context management#reinforcement learning
AI 研究PPR-06
Executable Environments Scale Agentic Tool-Use Supervision
The work introduces a dataset and model that scale environment coverage and outcome verification together, unifying restorable states, execution traces, outcomes, and verifiers as supervision substrate, and reusing states and verifiers to build preference data. It is directly relevant to researchers and engineering teams training reliable multi-step tool-use agents.
#agentic tool use#executable verification#supervised fine-tuning
Quantitative ResearchPPR-07
From Economic Explanation to Performance Improvement in Machine-Learning Portfolios
This study proposes a local first-order attribution-propagation framework that maps allocation-function feature exposures to local sensitivities of portfolio performance metrics, extending model-level feature importance to economically meaningful portfolio outcomes and decisions. Its novelty lies in deriving statistical inference for Sharpe-ratio attributions and using attribution-guided feature-exposure interventions for post-training adaptation without retraining the alloca
#machine-learning portfolios#attribution analysis#Sharpe ratio
Quantitative ResearchPPR-08
Neural Estimation of Irreversibility in Real and Simulated Limit Order Book Paths
This study extends time irreversibility measurement from scalar series to limit order book paths, defines an irreversibility rate for coarse-grained paths, and compares order-invariant and order-sensitive neural network estimates, finding a significant positive ordering increment on Nasdaq market-by-order data that a queue-reactive simulator fails to reproduce.
#limit order book#time irreversibility#neural estimation
AI 研究PPR-09
UDHRBench: A Benchmark for LLM Human Rights Alignment Reveals Efficiency and Neutrality Rationales as Main Violation Drivers
This study introduces UDHRBench, a benchmark covering all 30 articles of the Universal Declaration of Human Rights, with 1,124 human-rights-confronting questions and 4,496 responses from four large language models (ChatGPT, Grok, Gemini, Fanar). Expert annotations achieve high inter-annotator agreement (κ=0.801). Its novelty lies in operationalizing human rights compliance evaluation into a reproducible dataset and showing that violations typically arise from seemingly ration
#large language models#human rights alignment#benchmark evaluation
金融研究PPR-10
The Buffett Effect: Abnormal Premium from New-Position Disclosures by Recognizable Investors
This paper builds a recognition index from Wikipedia pageviews and ProQuest news mentions, compares 24 highly recognizable funds with 24 matched control funds, and finds a positive disclosure premium when recognizable investors reveal new positions, with the premium rising in recognition. A calendar-time portfolio that mechanically copies the most recognized investors' disclosures earns significant alpha that persists for months without reversal. Its value lies in adding inve
#investor recognition#disclosure premium#event study