当前位置：首页 > news >正文

如何快速定位做市商经典论文

news 2026/3/27 4:11:30

你提到在 arXiv 上搜索 “market making” 有 3400+ 篇论文，想快速：

定义研究范畴（什么是“做市”？）
识别经典/高质量论文
过滤掉水文（低质量、重复、无实质贡献）
用 AI 工具高效完成上述任务

下面我为你提供一套系统化 + 可落地的 AI 辅助筛选策略，结合最新工具（截至 2025 年），让你在几小时内完成从“3400 篇”到“30 篇核心文献”的提炼。

🔍 第一步：明确定义“Market Making”研究范围

避免把“订单流预测”“价差预测”“高频交易”等混进来。建议聚焦以下子方向：

子类关键词典型问题

理论做市模型 Avellaneda-Stoikov, inventory risk, optimal quoting 如何在价差收益与库存风险间权衡？
强化学习做市 RL market making, DQN, PPO, multi-agent 如何用 RL 学习最优报价策略？
实证做市分析 order book dynamics, liquidity provision 做市商行为对市场微观结构的影响？
加密货币做市 crypto market making, CEX/DEX 高波动、高滑点下的做市挑战？
多资产/跨市场做市 portfolio market making, cross-listing 多品种协同报价？
✅ 建议搜索式（arXiv 高级语法）：
text
(“market making” OR “market-maker”) AND (“inventory” OR “liquidity” OR “quoting” OR “Avellaneda”)

排除明显无关项：
text
NOT (“market prediction” OR “price forecasting” OR “sentiment”)

🧠 第二步：用 AI 工具快速筛选高质量论文（推荐组合）
✅ 推荐工具链（全部免费或试用）：

工具功能链接

AlphaXiv 输入关键词 → 自动生成综述 + 排名靠前论文 + 摘要总结 https://www.alphaxiv.org
Consensus 用自然语言提问，返回经评分的论文（含“是否被引用”“是否开源”） https://consensus.app
Scite.ai 查看论文是否被“支持性引用”（vs 仅提及） https://scite.ai
Connected Papers 输入一篇经典论文 → 生成文献图谱 https://www.connectedpapers.com
Elicit 自动提取论文“研究问题、方法、结论” https://elicit.org

🚀 实操流程：4 步快速淘金
🔹 Step 1：用 AlphaXiv 找“奠基性论文”和近期热点

访问：https://www.alphaxiv.org/assistant
输入：
“What are the most influential and foundational papers on market making in quantitative finance? Focus on theoretical models like Avellaneda-Stoikov and recent deep reinforcement learning approaches.”

✅ 它会返回：
经典论文（如 Avellaneda & Stoikov 2008）
被引 >500 的高影响力工作
近 2 年 SOTA 方法（如使用 Transformer 的做市）
💡 技巧：点击任意论文 → 自动跳转到 AlphaXiv 解读页，含 “核心思想”、“公式”、“实验设置” 摘要。

🔹 Step 2：用 Connected Papers 构建“核心文献网络”

找到 Avellaneda-Stoikov (2008) 的 arXiv 或 SSRN 链接（实际发表于 Quantitative Finance，但可找后续扩展版）
粘贴到 Connected Papers
生成文献图谱 → 中心节点是经典，外围是改进/应用

✅ 你会看到：
哪些论文是“真正推进了理论”（靠近中心）
哪些是“换个数据集跑 RL”（边缘、孤立节点 → 水文）

🔹 Step 3：用 Consensus / Elicit 批量评估质量

在 Consensus 输入：
“Does this paper propose a novel market making strategy with empirical validation?”

它会返回：
论文是否包含真实回测（vs 仅仿真）
是否开源代码（GitHub 链接）
是否被顶会接收（NeurIPS, ICML, AISTATS, WSDM 等）
📌 水文特征（可自动过滤）：
无实验对比（only “our method is better”）
未与 AS 模型 baseline 对比
使用合成数据且未验证鲁棒性
标题含 “A Novel Deep Learning Approach…” 但方法 trivial

🔹 Step 4：人工快速验证（10 分钟/篇）

对 AI 筛选出的 30~50 篇候选，快速检查：

检查项合格标准

是否引用 AS (2008)? 是 → 有理论根基
是否有明确的 reward function? 是 → 可复现
是否报告 inventory risk / PnL volatility? 是 → 关注做市核心指标
作者单位来自 Citadel/Jump/Two Sigma/Top CS Lab？→ 更可信

📚 附：做市领域必读经典清单（帮你锚定）

论文贡献年份

Avellaneda & Stoikov 奠基性存货控制模型 2008
Gueant, Lehalle & Fernandez-Tapia 扩展 AS 到多期、有限库存 2013
Spooner et al., “Market Making via Reinforcement Learning” 首个 RL 做市实证 2018 (AAMAS)
Belletti et al., “Deep Reinforcement Learning for Market Making” 使用 Dueling DQN 2017
Kanagalakshmi et al., “Multi-Agent Market Making” 多智能体博弈 2021
Zhou et al., “Dealing with Inventory Risk in Crypto MM” 加密货币做市挑战 2022
这些论文在 AlphaXiv 或 Connected Papers 中都会高频出现。

✅ 总结：你的高效筛选 pipeline

mermaid
graph LR
A[arXiv 3400+ 篇] --> B(AlphaXiv: 生成综述 + Top 50)
B --> C(Connected Papers: 构建文献图谱)
C --> D(Consensus/Elicit: 过滤无实验/无代码论文)
D --> E[人工快筛: 30 篇核心文献]
E --> F[精读 + 复现]

💡 最后建议
不要试图读完所有论文，聚焦“被广泛引用 + 开源 + 有清晰数学框架”的工作。
用 Zotero + AI 插件管理文献，自动打标签（如 #RL、#theory、#crypto）。
如果你做量化实盘，优先选择有实盘回测细节的论文（哪怕不是顶会）。

如果你愿意提供具体研究方向（如“我想做加密货币做市”或“想用 LLM 做动态价差”），我可以进一步缩小范围，给出定制化文献列表！

Market Making: From Avellaneda-Stoikov to Deep Reinforcement Learning

Research on market making has evolved from classical stochastic control frameworks to sophisticated deep reinforcement learning models that can handle the high-dimensional and non-stationary nature of modern limit order books. The foundational theory remains rooted in the work of Avellaneda and Stoikov, which introduced the concept of balancing bid-ask spreads against inventory risk. Recent literature has extended this into the realm of neural networks, focusing on high-frequency dynamics and multi-agent interactions.

Foundational and Theoretical Models

The following papers represent the bedrock of market making theory, focusing on stochastic control, inventory risk, and the evolution of the Avellaneda-Stoikov framework.

Optimal market making (9 years ago): This is a comprehensive review and extension of the field, building directly on the foundations of Avellaneda-Stoikov to solve the complex optimization problem of setting bid and ask prices while managing inventory risk under various utility functions.
Optimal market making under partial information with general intensities (7 years ago): This work extends the Avellaneda-Stoikov framework by relaxing the assumption of perfect information, considering a market maker who must estimate order arrival intensities to maximize expected utility over a finite horizon.
Macroscopic Market Making (2 years ago): This paper proposes a transition from discrete point processes to continuous processes, aiming to bridge the gap between microscopic order-by-order modeling and macroscopic market dynamics.
Optimal High Frequency Trading with limit and market orders (15 years ago): A key early paper that models the bid-ask spread as a Markov chain and derives optimal policies for a market maker using both limit and market orders in a high-frequency setting.
High-frequency market-making with inventory constraints and directional bets (14 years ago): This paper extends the classical inventory-based models by incorporating the market maker’s view on the drift of the mid-price, allowing for directional positioning alongside passive liquidity provision.

Deep Reinforcement Learning Approaches

Contemporary research focuses on using deep reinforcement learning (DRL) to overcome the “curse of dimensionality” and adapt to the non-linear dynamics of real-world markets.

Deep reinforcement learning for market making in corporate bonds: beating the curse of dimensionality (6 years ago): This is a significant application of DRL to Over-The-Counter (OTC) markets, demonstrating how neural networks can manage thousands of different assets simultaneously where classical closed-form solutions fail.
Market Making with Deep Reinforcement Learning from Limit Order Books (3 years ago): This paper focuses on training DRL agents directly on raw limit order book (LOB) data, allowing the agent to discover complex quoting strategies that react to the current state of liquidity and order flow.
Reinforcement Learning-Based Market Making as a Stochastic Control on Non-Stationary Limit Order Book Dynamics (3 months ago): A very recent contribution that treats market making as a control problem in a non-stationary environment, utilizing RL to adapt to changing market regimes.
Deep Reinforcement Learning for Market Making Under a Hawkes Process-Based Limit Order Book Model (3 years ago): This research integrates DRL with Hawkes processes—a class of point processes that capture the self-exciting nature of trades and limit orders—providing a more realistic simulation environment for agent training.
Reinforcement Learning in High-frequency Market Making (a year ago): This paper provides a theoretical bridge between modern reinforcement learning theory and continuous-time stochastic control, offering a rigorous analysis of how RL agents converge to optimal market making policies.
Deep Reinforcement Learning in Cryptocurrency Market Making (6 years ago): One of the earlier explorations into using policy gradient-based DRL agents for liquidity provision in the highly volatile cryptocurrency markets.

查看全文

http://www.jsqmd.com/news/156952/