--- title: ML & AI Methods url: https://www.ml-quant.com/topics/ml-ai-methods/ site: ML-Quant (https://www.ml-quant.com) updated: 2026-10-09 license: Summaries CC BY 4.0; links go to the original sources index: https://www.ml-quant.com/llms.txt --- # ML & AI Methods Machine-learning methods applied to finance: deep learning, boosting, RL and new architectures. 1135 papers featured; 63 in the last 12 months. Papers featured per quarter: 2023 Q2 27, 2023 Q3 55, 2023 Q4 88, 2024 Q1 79, 2024 Q2 154, 2024 Q3 202, 2024 Q4 168, 2025 Q1 119, 2025 Q2 146, 2025 Q3 29, 2025 Q4 27, 2026 Q1 2, 2026 Q2 1, 2026 Q3 11, 2026 Q4 27 ## Most cited - [Mamba: Linear-Time Sequence Modeling with Selective State Spaces](https://www.ml-quant.com/papers/arxiv/2312.00752/): 9474 citations. Sequence Modeling: Mamba, a neural network architecture that doesn't use attention or MLP blocks, provides faster inference and better performance in language, audio, and genomics than Transformers. - [Transformers are SSMs: Generalized Models and Efficient Algorithms Through Structured State Space Duality](https://www.ml-quant.com/papers/arxiv/2405.21060/): 2044 citations. The research identifies a link between state-space models and Transformers in deep learning, leading to the creation of a faster language modeling architecture, Mamba-2. - [Octo: An Open-Source Generalist Robot Policy](https://www.ml-quant.com/papers/arxiv/2405.12213/): 2026 citations. Octo is a large transformer-based policy for robotic manipulation, trained on a vast dataset, that can be instructed via language or images and adapted to new domains. - [Mastering Diverse Domains through World Models](https://www.ml-quant.com/papers/arxiv/2301.04104/): 1485 citations. Algorithm Mastery: DreamerV3, a universal algorithm, excels in over 150 varied tasks, including diamond collection in Minecraft without human input, expanding the scope of reinforcement learning. - [Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction](https://www.ml-quant.com/papers/arxiv/2404.02905/): 1209 citations. The article discusses Visual AutoRegressive modeling (VAR), a new image learning method that outperforms diffusion transformers in terms of speed, image quality, and scalability. - [SimPO: Simple Preference Optimization with a Reference-Free Reward](https://www.ml-quant.com/papers/arxiv/2405.14734/): 1206 citations. Simple Preference Optimization: SimPO improves reinforcement learning from human feedback by using the average log probability of a sequence as the implicit reward, enhancing training stability and computational efficiency. - [PixArt-α: Fast Training of Diffusion Transformer for Photorealistic Text-to-Image Synthesis](https://www.ml-quant.com/papers/arxiv/2310.00426/): 993 citations. PIXART-$\alpha$, a Transformer-based text-to-image model, generates high-quality images at a low cost, reducing CO2 emissions and offering a cost-effective solution for the AIGC community. - [Fourier Neural Operator with Learned Deformations for PDEs on General Geometries](https://www.ml-quant.com/papers/arxiv/2207.05209/): 765 citations. The study introduces geo-FNO, a new framework for solving partial differential equations on any geometry, proving to be faster and more accurate than standard and machine learning-based solvers. - [LightGaussian: Unbounded 3D Gaussian Compression with 15x Reduction and 200+ FPS](https://www.ml-quant.com/papers/arxiv/2311.17245/): 685 citations. LightGaussian is a new method that converts 3D Gaussians into a more compact format, enhancing efficiency in real-time neural rendering and reducing storage needs. - [Gated Linear Attention Transformers with Hardware-Efficient Training](https://www.ml-quant.com/papers/arxiv/2312.06635/): 578 citations. Efficient Training of Gated Linear Attention Transformers: The research introduces a more hardware-efficient version of gated linear attention Transformers that performs well against other models, especially in training on longer sequences. - [CAT3D: Create Anything in 3D with Multi-View Diffusion Models](https://www.ml-quant.com/papers/arxiv/2405.10314/): 490 citations. Multi-View Diffusion Models: CAT3D is a novel technique for generating 3D scenes from any number of images, surpassing existing methods in speed and efficiency. - [Capabilities of Gemini Models in Medicine](https://www.ml-quant.com/papers/arxiv/2404.18416/): 451 citations. Med-Gemini, an AI model for medical applications, outperforms previous models and human experts in medical benchmarks, indicating potential for real-world medical use. ## Latest - [Towards Financial World Modeling](https://www.ml-quant.com/papers/arxiv/2610.09048/) (2026-10-09): The study introduces a trillion-observation U.S. equity dataset and systematically evaluates 18 representation-learning methods across nearly two decades, finding that encoders with similar predictive performance organize market state very differently. - [STOCK-JEPA: Prior-Anchored Latent Revision Representation Learning in Equity Markets](https://www.ml-quant.com/papers/arxiv/2610.07006/) (2026-10-09): A joint-embedding framework learns predictable incremental revisions from market data relative to a financial prior, combining interpretability from classic models with the pattern-capture strength of deep learning. - [LiveMACE: Process-Aware Evaluation of LLM Agent Capabilities in Evolving Markets](https://www.ml-quant.com/papers/arxiv/2610.09872/) (2026-10-09): A benchmark evaluates frontier LLMs as live trading agents and finds that realized returns often diverge from capability-specific measurements, revealing outcome-capability gaps through decision traces. - [The Agentic ETF: How Agentic Trading Becomes an Asset Class](https://www.ml-quant.com/papers/arxiv/2610.06856/) (2026-10-09): Proposes that large-language-model-driven trading agents delegated to autonomous execution define a nascent asset class, mapping infrastructure layers and projecting potential trillions in assets by 2030. - [Trading Strategy Optimization via Textual Gradient](https://www.ml-quant.com/papers/arxiv/2610.03128/) (2026-10-09): Proposes an LLM-guided framework for trading strategy refinement using accumulated experience and cross-period robustness, achieving 27.99% annualized return on Chinese equities with 1.63 Sharpe ratio. - [The Price of Greenwashing: Algorithmic Verification and Market Discipline using Conformal Machine Learning](https://www.ml-quant.com/papers/arxiv/2610.02225/) (2026-10-09): Fuses SEC financials with EPA emissions data and conformal machine learning to quantify divergence between self-reported and physical emissions, finding algorithmic divergence predicts lower valuation and profitability. - [Does Autonomous Quant Research Improve? Leakage, Evidence Scarcity and Forward Tests in an LLM Factor-Discovery Loop](https://www.ml-quant.com/papers/ssrn/7555699/) (2026-10-09): An autonomous agent evolved 940 factors over 17 days and finds that reusing backtest data inflates edge by a quarter to a third and in-sample improvement predicts worse performance. - [Firm-level AI Exposure: Measurement, Determinants, and Pricing Implications](https://www.ml-quant.com/papers/ssrn/7575518/) (2026-10-09): A natural-language analysis of 13,757 firms finds AI exposure surged after ChatGPT and predicts 1.37 percent higher annual returns, though the effect is concentrated after the launch and driven by firm-specific factors. - [Agents Propose, Code Decides: An Autonomous Research Loop Framework to Build an Equity Alpha Model](https://www.ml-quant.com/papers/ssrn/7568638/) (2026-10-09): An autonomous research loop in which agents propose factor candidates and code verifies them builds multi-factor alpha models, screening 212 predictors and admitting 14 that raise baseline information coefficient out-of-sample. - [GFlowAlpha: A Quality-Control-Aware Neuro-Symbolic Framework for White-Box Alpha Factor Mining](https://www.ml-quant.com/papers/ssrn/7576913/) (2026-10-09): GFlowAlpha uses neuro-symbolic search and internalizes econometric quality control as online rewards to mine white-box factors, delivering mean out-of-sample daily ICIR of 0.485 with Fama-MacBeth t-statistics of 4.99 on Chinese stocks. - [Deflate by Bits, Not Trials: Why Query-Counted Deflated Sharpe Ratios Fail Under Adaptive Search, and a Sealed-Holdout Remedy](https://www.ml-quant.com/papers/ssrn/7557458/) (2026-10-09): The research proves that query-counted Deflated Sharpe Ratios certify noise in adaptive research and proposes sealed-holdout verification, where validity depends on bits revealed rather than strategy count. - [A hybrid graph neural network for realized volatility forecasting: Multi-relation attention and dual-path fusion](https://www.ml-quant.com/papers/ssrn/7576359/) (2026-10-09): A graph neural network combining temporal and cross-asset volatility patterns reduces mean squared error and quasi-likelihood loss by approximately 9.9% and 3.4% on Dow stocks. - [AI Intensity and Financial Market Responses to Economic Shocks: Evidence from US Industries](https://www.ml-quant.com/papers/repec/ces-ceswps-13000/) (2026-10-09): Industries with higher AI intensity show significantly higher returns and valuations when hit by supply and technology shocks but not demand shocks. - [Can AI Make Money in Crypto? Measuring the Gap from Backtests to Real Markets](https://www.ml-quant.com/papers/arxiv/2609.34510/) (2026-10-02): A benchmark compares machine learning, reinforcement learning, and large language model trading methods across historical backtests, paper trading, and live markets to measure the gap. - [AlphaPareto: Formulaic Alpha Discovery with LLM-Guided Multi-Objective Reinforcement Learning](https://www.ml-quant.com/papers/arxiv/2609.34188/) (2026-10-02): The research uses reinforcement learning with multi-objective rewards to discover formulaic alphas that work well together despite evolving reward functions and shifting environments. - [PALM: Point-in-Time Adaptation for Financial Language Models](https://www.ml-quant.com/papers/arxiv/2609.30316/) (2026-10-02): A new method adapts financial language models to avoid look-ahead bias using low-rank adapters instead of expensive annual retraining, matching performance of full pretraining. - [Same Text, Different Numbers: The Divergence of LLM-Based Measures](https://www.ml-quant.com/papers/arxiv/2609.31013/) (2026-10-02): Cross-model rank correlations for LLM-extracted sentiment, clarity and risk average only 0.52, and model choice significantly alters coefficient signs and significance in downstream analysis. - [KiT: A Foundation Model for Financial Time-Series Forecasting using DiffusionTransformers](https://www.ml-quant.com/papers/arxiv/2609.34507/) (2026-10-02): Diffusion transformer foundation model reformulates candlestick prediction as conditional trajectory generation, achieving 0.057 mean return RankIC across markets and timescales. - [The Price of Thought: Does Test-Time Reasoning Pay in LLM Trading?](https://www.ml-quant.com/papers/arxiv/2609.30705/) (2026-10-02): The study tests whether extra reasoning in large language models improves portfolio returns net of trading costs across multiple model families, finding no reliable gains. - [LiveOption: Evaluating LLM Agents in Structured Option Trading with Nonlinear Payoffs](https://www.ml-quant.com/papers/arxiv/2609.33470/) (2026-10-02): An evaluation framework tests language model agents on structured option trading tasks, finding current systems underperform in most real-world scenarios. - [EverMine: Dissecting the Self-Evolution of Research Capabilities in Long-Horizon Alpha Research](https://www.ml-quant.com/papers/arxiv/2609.33524/) (2026-10-02): A framework tests whether self-evolving research capabilities improve alpha discovery over time, finding no consistent gains from accumulated experience and tools. - [Deep kernel hedging](https://www.ml-quant.com/papers/arxiv/2609.34474/) (2026-10-02): A hybrid approach combining neural networks with kernel methods produces robust derivative hedges in low-data regimes and scales via random Fourier features with convergence guarantees. - [Can Language Models Learn to Forecast Stock Prices](https://www.ml-quant.com/papers/arxiv/2609.36914/) (2026-10-02): Post-training Qwen3-4B via supervised fine-tuning and policy optimization more than doubles direction-magnitude score on chronological stock-price predictions to 43.31. - [Residual Learning in Empirical Asset Pricing](https://www.ml-quant.com/papers/ssrn/7528558/) (2026-10-02): Deep residual networks outperform shallow models in asset pricing, achieving a long-short Sharpe ratio of 2.07 versus 1.92 for shallow versions by preserving and refining earlier layers. - [Dynamic Graph Neural Networks for Real Time Systemic Risk Surveillance: An Explainable AI Framework for Financial Stability](https://www.ml-quant.com/papers/ssrn/7522878/) (2026-10-02): The research proposes a temporal graph neural network with explainability tools for real-time systemic risk surveillance, achieving early warning signals 3-4 quarters ahead of financial distress on bank data. - [A SPOT in the dark: using AI to assess financial stability risks](https://www.ml-quant.com/papers/repec/ecb-ecbwps-20263262/) (2026-10-02): Large Language Models extract signals about potential trigger events from financial news, improving forward-looking estimates of downside risks and helping monitor financial stability threats ahead of major events. - [The Labor Market Effects of AI Automation and Augmentation](https://www.ml-quant.com/papers/repec/zbw-glodps-1821/) (2026-10-02): AI automation reduces occupational employment by 21 percent with little wage effect, while AI augmentation raises wages by 8 percent, revealing that AI's labor impact depends on the balance between these opposing channels. - [Propose, Don't Judge: An Anytime-Valid Referee for LLM Agents That Mine Investment Factors](https://www.ml-quant.com/papers/arxiv/2609.27051/) (2026-09-25): Proposes a statistical referee that judges investment factors proposed by language-model agents using out-of-sample market outcomes, ensuring false-discovery control at any stopping time. - [AlphaDiverse: Post-Training Local Quantitative Research Agents for Diverse Exploration in Alpha Factor Mining](https://www.ml-quant.com/papers/arxiv/2609.29014/) (2026-09-25): Proposes a multi-agent system with post-training that automates alpha factor mining locally, using diverse research paths and joint optimization to broaden exploration while maintaining prediction quality. - [Forecast-Dojo: Replayable Environments for Benchmarking and Training LLM Forecasting Agents](https://www.ml-quant.com/papers/arxiv/2609.28876/) (2026-09-25): Introduces a replayable environment combining 1,568 resolved prediction-market questions with 18.8M dated news articles to benchmark and train language-model forecasting agents on historical data.