Tauroi Runs Live AI Trading Agent as Founders Warn on Execution Drag

A Live Demo of an AI Agent Trading Financial Markets With Tauroi Technologies — Ai4

Ai4

Algorithmic trading models fail in production without extensive infrastructure to counter slippage and order-flow friction, leaving the core machine learning algorithm as just five percent of the required architecture.

AI-assisted summary of A Live Demo of an AI Agent Trading Financial Markets With Tauroi Technologies — original source. Check the original for full context.

Source published:

Detailed notes

## Background Published on YouTube by the Ai4 channel on August 27, 2025, this talk features Tauroi Technologies co-founders Daniel Sonner and Avi Thaker. According to the description, the session demonstrates an autonomous AI agent trading financial markets live—raffling off the proceeds to attendees—while exploring how quantitative hedge funds leverage alternative data to engineer real-time trading systems.

## TL;DR Tauroi Technologies co-founders Avi Thaker and Daniel Sonner run an autonomous Deep Q-Network trading agent live on stage, executing a real-money position on audience-selected UnitedHealth Group (UNH). Between execution checkpoints, they outline institutional portfolio theory, illustrating why endowments prioritize alternative quantitative assets to mitigate devastating drawdowns that traditional 60/40 portfolios fail to prevent. They deconstruct market microstructure realities—including payment for order flow (PFOF) and execution slippage—and stress that core machine learning code represents only a tiny fraction of an operational trading pipeline. After exploring quantitative use cases like satellite crop forecasting and market making, they conclude the session by locking in a $99 trading gain awarded directly to an audience raffle winner.

## Atmosphere & style The session has an energetic, fast-paced, and collaborative workshop vibe, driven by lively, natural banter between Avi Thaker and Daniel Sonner as they juggle live terminal execution with slide commentary. Tying the live trading run to a cash raffle injects playful suspense and keeps the audience invested in an otherwise highly technical lecture on reinforcement learning and order book dynamics. The presentation feels grounded and refreshing, actively puncturing algorithmic trading hype by candidly confronting attendees with the unglamorous realities of slippage, market-maker advantages, and data engineering overhead.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=xzUpSPZzMys&t=4s) Tauroi Technologies demonstrates an autonomous reinforcement learning trading agent operating live on financial markets during their presentation. - [01:31](https://www.youtube.com/watch?v=xzUpSPZzMys&t=91s) While tabular Q-learning works for discrete environments, continuous and partially observable financial market states require Deep Q-Networks (DQN) to approximate value functions. - [03:35](https://www.youtube.com/watch?v=xzUpSPZzMys&t=215s) Algorithmic trading strategies often fail when transitioning from backtests to live production due to unmodeled execution drag, slippage, and exchange fees. - [05:48](https://www.youtube.com/watch?v=xzUpSPZzMys&t=348s) Institutional endowments heavily allocate to quantitative alternatives because portfolio drawdowns are mathematically asymmetrical—a 50% loss demands a 100% gain to break even. - [05:48](https://www.youtube.com/watch?v=xzUpSPZzMys&t=348s) Combining truly uncorrelated return streams can smooth portfolio volatility and virtually eliminate drawdowns while preserving terminal returns. - [08:22](https://www.youtube.com/watch?v=xzUpSPZzMys&t=502s) Traditional 60/40 equity-bond diversification is structurally fragile because rolling correlations between stocks and bonds fluctuate unpredictably and frequently turn positive during market shocks. - [08:22](https://www.youtube.com/watch?v=xzUpSPZzMys&t=502s) Payment for Order Flow (PFOF) dynamics and high-frequency market-making wholesalers introduce significant structural execution disadvantages for retail algorithmic traders. - [10:55](https://www.youtube.com/watch?v=xzUpSPZzMys&t=655s) In production quantitative trading systems, core machine learning code accounts for only a minor fraction of the architecture, with the majority of engineering dedicated to data pipelines, monitoring, and execution harnesses. - [10:55](https://www.youtube.com/watch?v=xzUpSPZzMys&t=655s) Standalone conventional technical analysis indicators lack predictive edge in modern efficient markets unless integrated into rigorous multi-modal statistical models. - [13:18](https://www.youtube.com/watch?v=xzUpSPZzMys&t=798s) Quantitative AI agents provide high utility in specialized domain applications, including cross-exchange arbitrage, order-book market making, election event prediction, and alternative data forecasting. - [16:31](https://www.youtube.com/watch?v=xzUpSPZzMys&t=991s) Deploying larger AI models in live trading creates critical engineering trade-offs between predictive capacity and execution latency.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=xzUpSPZzMys&t=4s) Introductions, Live Trading Demo Setup, and Raffle

Avi Thaker and Daniel Sonner, co-founders of Tauroi Technologies, kick off their session in Las Vegas with an ambitious plan: running an autonomous AI trading agent live during the talk using real capital, while raffling off the proceeds to the audience. Introducing their backgrounds, Avi highlights his industry experience spanning Microsoft and GSK, several patents, passing financial licensing exams, and co-running a hedge fund alongside Daniel. Daniel introduces his background as a systems engineer at Pure Storage, where he also earned patents, adding the personal fun fact that he is a World Bridge Champion.

To make the session memorable, the speakers project a QR code to enter an audience raffle for the trading profits. Avi initially mislabels the mechanism as an auction before Daniel immediately chimes in to clarify it is a raffle. Under the rules of the giveaway, attendees are guaranteed to win at least $20 even if the algorithm ends up losing money. Avi jokes about the state of his personal finances in the Golden State, noting that because he lives in California and sits in a 50% tax bracket, half the profits have to cover taxes, but the remaining upside goes directly to the winner. They urge the audience to scan the QR code right away before it is taken down, emphasizing that the talk should be an interactive, open conversation where attendees should freely interrupt with questions.

- **Avi Thaker**: Co-founder of Tauroi Technologies; former Microsoft and GSK; holds industry patents and financial licensing credentials. - **Daniel Sonner**: Co-founder of Tauroi Technologies; former systems engineer at Pure Storage with engineering patents; Bridge World Champion. - **Raffle terms**: Attendees win at least a $20 floor regardless of trading performance, with net positive proceeds distributed after a 50% California tax deduction.

---

[01:31](https://www.youtube.com/watch?v=xzUpSPZzMys&t=91s) Autonomous Agents and Reinforcement Learning from Games to Markets

The speakers transition into defining what an AI agent actually is in practical terms: an autonomous software entity capable of planning, executing actions, and leveraging external tools. Agents can range from broad, multi-purpose systems to narrow, specialized constructs designed for a singular operational scope. For the day's presentation, they focus specifically on a narrowly scoped agent engineered to interact directly with financial market order flows.

To illustrate how an agent learns, Avi draws an analogy to classic video games like Pac-Man. In reinforcement learning environments, an agent optimizes its behavior against an explicit reward function—earning positive rewards for eating dots and consuming ghosts, while suffering penalties or declining scores when idling. In simple, discrete state spaces, this can be handled via tabular Q-learning, where the agent maintains a lookup table mapping states to optimal actions. However, financial markets present near-infinite, partially observable continuous states. As a result, practical trading agents require Deep Q-Networks (DQN) to approximate complex value functions across unpredictable environments.

``` +-------------------------------------------------------------------+ | REINFORCEMENT LEARNING LOOP | | | | +---------------+ State / Market Data +------------+ | | | |----------------------------->| | | | | Environment | | Agent | | | | (Stock Order | Reward / P&L Change | (Deep Q- | | | | Book) |----------------------------->| Network) | | | | |<-----------------------------| | | | +---------------+ Action: Buy/Sell/Hold +------------+ | +-------------------------------------------------------------------+ ```

- **Core agent capabilities**: Autonomous planning, environmental tool usage, and execution across varying breadths of scope. - **Q-learning vs. Deep Q-learning**: Tabular Q-learning explicitly maps discrete states to actions, whereas Deep Q-learning uses neural networks to approximate functions across massive market state spaces. - **Action space for the demo agent**: Buy, sell, or maintain a flat position (wait).

---

[03:35](https://www.youtube.com/watch?v=xzUpSPZzMys&t=215s) Live Trading Demo: Executing UNH and Factoring in Slippage

Before firing up the software, Avi emphasizes a strict disclaimer: "Please don't try this at home. This is not financial advice and this is a demo." Even though Tauroi provides an open-source GitHub repository containing the baseline architecture, executing automated systems against live order books carries substantial financial risk. Avi calls on the audience to pick any stock from a displayed list. An attendee calls out UnitedHealth Group (UNH), prompting the team to configure the agent to pull all pre-trade market data and initialize execution.

As the script begins cycling through real-time checks, the agent initially decides to hold its position flat while awaiting favorable conditions. An account authentication and permissions hurdle crops up on the primary funded brokerage account, prompting them to pivot the terminal to a secondary demo account that mirrors live market conditions. The model issues an order and buys UNH at $259.70. Avi immediately directs attention to the transaction diagnostics on screen, specifically the cost of slippage. He notes that amateur algorithmic traders frequently overlook slippage, execution drag, and exchange exchange fees, which can quickly turn a paper-profitable model into a net-negative strategy. The team leaves the agent executing in the background to accumulate P&L while switching back to the slide deck.

- **Audience selection**: UnitedHealth Group (ticker: UNH). - **Initial execution**: Purchase executed at $259.70. - **Critical trading reality**: Model backtests frequently fail in production because they underestimate slippage and transactional friction.

---

[05:48](https://www.youtube.com/watch?v=xzUpSPZzMys&t=348s) Why Alternatives Matter: The Endowment Model and Drawdown Math

Daniel takes over to explain why sophisticated institutional investors prioritize quantitative, alternative strategies. He presents asset allocation trends from major university endowments over recent decades. Historically, endowments held heavy allocations in traditional US public equities, but over time, institutions like the Yale University endowment pioneered a dramatic shift: sharply reducing domestic equity exposure in favor of alternatives, including private equity and AI-powered quantitative hedge funds.

``` Institutional Portfolio Allocation Shift Over Time:

Traditional Mix: [===================== US Equities =====================][ Alternatives ] | v Endowment Model: [== US Equities ==][================= Alternatives ==================] ```

The underlying math driving this structural reallocation comes down to the destructive nature of portfolio drawdowns. As Daniel explains, overall annualized returns mean very little if the interim path is unbearable. If a portfolio suffers a 50% drawdown, it requires a 100% gain simply to recover to even—a process that typically takes 5 to 10 years of grueling market weathering.

Daniel illustrates this with a comparative model: - **Asset 1 (Blue Line)**: The S&P 500, delivering solid long-term returns but punctuated by deep, painful cyclical crashes. - **Asset 2 (Yellow Line)**: A hypothetical uncorrelated strategy that moves inversely to public equities, achieving the identical net return over time but enduring its own substantial drawdowns. - **Combined Portfolio (Red Line)**: Splitting capital 50/50 between the two uncorrelated streams yields the exact same ultimate terminal wealth, but completely smooths out the volatility curve, virtually eliminating drawdowns.

---

[08:22](https://www.youtube.com/watch?v=xzUpSPZzMys&t=502s) Diversification Pitfalls and the Reality of Market Microstructure

Daniel notes that retail investors often attempt to engineer this smoothing effect through a classic 60/40 equity and bond portfolio. However, historical rolling correlation data between the S&P 500 and US aggregate bonds reveals that the relationship is notoriously unstable: bonds are sometimes positively correlated, sometimes uncorrelated, and sometimes negatively correlated with equities. When equities crash during inflationary or rate-hiking shocks, bonds often fall in tandem, destroying the defensive cushion. True alternative strategies seek systematic, structural uncorrelation rather than relying on shifting asset-class relationships.

Turning to the structural hurdles facing newcomers, Avi addresses the difficulty of breaking into the market as an independent retail trader. Whether aspiring quants choose the buy side (hedge funds), the sell side (investment banks), or proprietary shops, hands-on systems experience is paramount. He warns aspiring algorithmic developers that market plumbing is stacked against the naive retail trader, illustrating this with Robinhood's September 2024 Payment for Order Flow (PFOF) revenue figures. Market makers like Citadel Securities, Two Sigma, Virtu Financial, and Wolverine Trading pay hundreds of millions of dollars for retail retail order flow, monetizing retail retail flow through market making and internalizing retail trades. Any algorithmic agent operating in this retail retail tier faces immediate structural headwinds through indirect execution costs.

- **Instability of traditional hedging**: Bond-equity correlations fluctuate unpredictably, failing to provide reliable drawdown protection during joint market downturns. - **PFOF wholesalers**: High-frequency market-making firms (Citadel Securities, Two Sigma, Virtu, Wolverine) purchase order flow to capture fractional bid-ask spreads. - **Institutional execution drag**: Retail algorithmic agents must account for being front-run or disfavored by internalized order routing mechanisms.

---

[10:55](https://www.youtube.com/watch?v=xzUpSPZzMys&t=655s) Market Efficiency, Systems Engineering, and Trading Realities

Avi reviews the Efficient Market Hypothesis (EMH), reminding the audience that while markets are largely efficient and reflect all available information, select institutional powerhouses consistently extract alpha. Firms like Citadel and Virtu achieve multi-year stretches with almost zero losing trading days, but they do so via superior low-latency infrastructure, proprietary data feeds, and privileged order-flow visibility. For an independent model to survive, it must possess verifiable statistical edges rather than relying on random variance. Avi points out that across large populations of traders, sheer luck will make a handful look like geniuses: "If you had a bunch of monkeys throw darts or randomly selected, some people were just going to win." He references a famous speculator who blew up multiple times yet claimed his success was a 700-standard-deviation statistical anomaly.

``` +-------------------------------------------------------------------+ | REAL-WORLD ML TRADING SYSTEM ARCHITECTURE | | | | +--------------------+ +--------------------+ +-------------+ | | | Data Collection | | Feature Engineering| | Logging | | | | & Cleaning | | & Pipelines | | & Monitoring| | | +--------------------+ +--------------------+ +-------------+ | | | | | | | +------------+ +-----+ +-------------+ | | | | | | | v v v | | +---------------------------+ | | | [ ML / AI Model Core (5%) ]| | | +---------------------------+ | | ^ ^ ^ | | +------------+ +-----+ +-------------+ | | | | | | | +--------------------+ +--------------------+ +-------------+ | | | Risk Management | | Order Execution | | Backtesting | | | | & Sizing Rules | | & Routing Engine | | Validation | | | +--------------------+ +--------------------+ +-------------+ | +-------------------------------------------------------------------+ ```

Citing a well-known Google research paper on the hidden technical debt in machine learning, Avi emphasizes that core ML algorithms represent only a tiny fraction of an operational quantitative trading system. The vast majority of production engineering consists of data collection, feature engineering pipelines, logging, backtesting harnesses, and monitoring frameworks. Aspiring practitioners should focus on building robust end-to-end data systems rather than fixating exclusively on algorithmic novelties. Furthermore, he warns that traditional technical indicators (such as RSI or moving average crossovers) are virtually useless in isolation against professional competition unless integrated into rigorous, multi-modal statistical engines.

- **Google ML architecture insight**: Machine learning code is a minor component surrounded by vast data processing, verification, and monitoring infrastructure. - **Technical analysis caution**: Standard off-the-shelf technical indicators carry zero standalone predictive edge in modern high-efficiency markets. - **Statistical significance**: Validating a trading agent requires hundreds of out-of-sample trades to separate real alpha from random statistical survival.

---

[13:18](https://www.youtube.com/watch?v=xzUpSPZzMys&t=798s) Domain-Specific Financial Agents and Live Trading Results

Daniel outlines four domain-specific applications where quantitative AI agents can be deployed across modern markets:

1. **Cross-Exchange Arbitrage**: Exploiting transient price discrepancies for identical assets across distinct venues. If an asset like Bitcoin trades for lower on Exchange A than on Exchange B, an agent can instantly buy on Venue A and sell on Venue B, pocketing the risk-free spread after factoring in network transfer fees. 2. **Market Making**: Providing continuous liquidity to both sides of the order book. Daniel explains the mechanics using a simple order book example:

``` +-------------------------------------------------------------------+ | MARKET MAKING ORDER SPREAD | | | | BID (Buy Orders) ASK (Sell Orders)| | Best Bid: $99.99 <--- [Market Maker Spread] ---> Best Ask: $100.00 | | | * Market maker simultaneously quotes $99.99 Bid and $100.00 Ask | | * Seller hits the bid at $99.99; buyer takes the ask at $100.00 | | * Result: $0.01 profit captured per round-trip share | +-------------------------------------------------------------------+ ```

3. **Event & Election Impact Prediction**: Tauroi published a research paper detailing a Deep Neural Network (DNN) trained to forecast political election outcomes and subsequent one-day market reactions. Rather than merely forecasting the victor, the agent identifies how asset classes react the day after results are certified. 4. **Satellite & Alternative Data Forecasting**: In another published paper, the team used overhead satellite imagery of agricultural wheat fields combined with historic commodity futures data. A computer vision and time-series model forecasted crop yields and future wheat pricing, allowing the agent to position long or short commodity contracts ahead of official government harvest reports.

The speakers switch back to the live terminal to inspect their ongoing run. The trading agent completed its execution cycle on the UNH position, locking in a total profit of $99. Avi announces that the raffle winner will receive the full $99 profit as an Amazon gift card, delivering on their promise to return real cash to the audience while reiterating that the open-source script remains an educational starting point.

[16:31](https://www.youtube.com/watch?v=xzUpSPZzMys&t=991s) Wrapping Up: Model Sizing, Production Latency, and Concluding Remarks

The speakers conclude their presentation by emphasizing that the setup shown was intentionally kept lightweight as a proof-of-concept demonstration. While practitioners have the flexibility to integrate "much larger models" into the architecture, taking such systems into a live production environment introduces critical engineering trade-offs—most notably managing execution latencies and system responsiveness. Noting that they have reached the end of their allotted time, the presenters invite the audience to connect with them afterwards for further discussion and offline questions, wrapping up the session by thanking everyone for attending.

## Notable quotes > [03:35](https://www.youtube.com/watch?v=xzUpSPZzMys&t=215s) Please don't try this at home. This is not financial advice and this is a demo. — Avi Thaker

> [10:55](https://www.youtube.com/watch?v=xzUpSPZzMys&t=655s) If you had a bunch of monkeys throw darts or randomly selected, some people were just going to win. — Avi Thaker

> [16:31](https://www.youtube.com/watch?v=xzUpSPZzMys&t=991s) much larger models — Speakers

## Mentioned Ai4, Avi Thaker, Citadel Securities, Daniel Sonner, GlaxoSmithKline (GSK), Google, Microsoft, Pac-Man, Pure Storage, Robinhood, Tauroi Technologies, Two Sigma, UnitedHealth Group (UNH), Virtu Financial, Wolverine Trading, Yale University

## Video description Daniel Sonner, Co-Founder, Tauroi Technologies Avi Thaker, Co-Founder, Tauroi Technologies

Abstracts

This talk will showcase an AI agent trading markets live, and will raffle off 50% of the winnings! Explore how hedge funds use alternative data to build realtime agents to trade financial markets!

[Source](https://www.youtube.com/watch?v=xzUpSPZzMys)

中文简报

Tauroi现场演示AI交易智能体，创始人警示执行拖累 若缺乏应对滑点和订单流摩擦的大量基础设施，算法交易模型在实盘生产中往往会失效，核心机器学习算法仅占所需架构的5%。 Tauroi Technologies联合创始人Avi Thaker和Daniel Sonner在演讲现场实盘运行了一个自主Deep Q-Network交易智能体，对观众选定的UnitedHealth Group（UNH）执行了真金白银的建仓交易。在执行检查点的间隙，他们概述了机构投资组合理论，阐释了为何捐赠基金会优先配置另类量化资产，以缓解传统60/40投资组合无法防范的毁灭性回撤。他们剖析了包括订单流支付（PFOF）与执行滑点在内的市场微观结构现实，并强调核心机器学习代码仅占可用交易流水线系统极小的一部分。在探讨了卫星农作物预测和做市等量化应用案例后，他们锁定了99美元的交易收益，并将该收益直接奖励给了一位现场抽奖中奖的观众，从而结束了这场演示。 Tauroi Technologies 联合创始人现场演示基于强化学习的自主交易 Agent，实盘买入联合国健康（UNH）并最终锁定 99 美元利润，借此展示滑点与执行摩擦对算法策略的直接影响。 (3:35) 针对金融市场连续且部分可观测的状态空间，交易 Agent 采用深度 Q 网络（DQN）逼近价值函数，以克服传统表格 Q-learning 仅能处理离散简单环境的局限。 (1:31) 机构捐赠基金转向另类量化策略的核心在于回撤的数学非对称性——亏损 50% 需要 100% 的涨幅才能回本；通过引入真正不相关的收益流可平滑波动并大幅规避深度回撤。 (5:48) 传统 60/40 股债组合的相关性在危机时常不可靠地转为正向，加上高频做市商购买订单流（PFOF）的微观结构，使独立散户量化交易者在执行层面面临显著的劣势阻力。 (8:22) 在生产级量化系统中，核心机器学习代码仅占系统工程的极小部分（约 5%），绝大多数开发精力必须投入到数据管道、特征工程、回测验证与实时监控等基础设施中。 (10:55)

Reading summary

During a live demonstration for Ai4, Tauroi Technologies co-founders Avi Thaker and Daniel Sonner argued that building successful autonomous trading agents requires mastering market microstructure and extensive systems engineering rather than relying solely on isolated machine learning models. Reinforcement Learning and Live Execution Realities (1:31) Thaker and Sonner illustrated autonomous trading using reinforcement learning, contrasting tabular Q-learning in simple discrete environments like Pac-Man with Deep Q-Networks (DQNs) required to approximate complex continuous states in financial markets. In their demo, a narrowly scoped agent was tasked with choosing between buying, selling, or remaining flat. To demonstrate execution, the presenters deployed an agent to trade UnitedHealth Group (UNH). Following an account authentication issue on their primary account, they switched to a demo account mirroring live conditions and bought UNH at $259.70. Thaker emphasized that amateur quants frequently see paper profits vanish because they fail to account for slippage, latency, and transactional friction in live order books. Drawdown Mathematics and the Institutional Endowment Model (5:48) Explaining institutional demand for quantitative alternatives, Sonner cited the evolution of university endowments, such as Yale, which shifted away from heavy US public equity allocations into alternative investments and quantitative strategies. This reallocation is driven by the asymmetry of drawdowns: recovering from a 50% portfolio loss requires a 100% gain, which typically demands 5 to 10 years of market recovery. Sonner demonstrated that pairing public equities with a hypothetical, negatively correlated asset can achieve identical terminal returns while virtually eliminating drawdown volatility. He cautioned that traditional retail attempts to diversify via a 60/40 equity-and-bond split often fail because rolling correlations between stocks and aggregate bonds are historically unstable, causing both assets to decline simultaneously during market shocks. Market Plumbing and Engineering Overhead (8:22) Thaker warned retail developers that market infrastructure creates inherent structural headwinds. Citing Robinhood's September 2024 payment for order flow (PFOF) data, he noted that major wholesalers like Citadel Securities, Two Sigma, Virtu Financial, and Wolverine Trading pay hundreds of millions of dollars to monetize retail flow, leaving independent algorithmic traders exposed to indirect execution costs. Citing Google research on technical debt in machine learning, Thaker asserted that core models represent only roughly 5% of a functional quantitative system. The remaining 95% involves data engineering, feature pipelines, execution routing, and monitoring. He cautioned against relying on off-the-shelf technical indicators like RSI, adding that true statistical significance requires validating hundreds of out-of-sample trades to distinguish sustainable edge from survivorship bias or random chance. Quantitative Agent Applications Across Alternative Datasets (13:18) Sonner outlined four distinct functional architectures for trading agents: cross-exchange arbitrage to exploit fleeting price spreads across venues, continuous market making by providing bidirectional liquidity across bid-ask spreads, event-driven election impact forecasting using deep neural networks to project post-certification market reactions, and satellite-based commodity modeling. In one published paper discussed by the founders, computer vision models evaluated overhead satellite imagery of agricultural fields alongside historical futures data to forecast crop yields and price wheat contracts ahead of official government reports. The live UNH demo concluded by locking in a $99 gain, which was distributed to an audience raffle winner as an Amazon gift card. The presenters noted that their lightweight demonstration model sacrificed complexity for immediate demonstration purposes, cautioning that scaling to larger architectures in production introduces critical latency trade-offs that can compromise live execution responsiveness.

精炼阅读摘要

Tauroi Technologies联合创始人Avi Thaker与Daniel Sonner指出，在金融市场中构建自主AI交易代理时，机器学习模型本身仅占系统极小部分，工程成败关键在于底层数据流、执行滑点控制以及真实的非相关性风险管理。 强化学习与DQN在连续市场状态中的应用 (1:31) 演讲者将AI代理定义为具备规划、行动执行并能调用外部工具的自主软件实体。与经典视频游戏（如吃豆人）中离散状态空间的表格型Q-learning（Tabular Q-learning）不同，金融市场具有近乎无限且部分可观测的连续状态，必须依赖深度Q网络（Deep Q-Networks, DQN）来近似复杂的价值函数。代理的动作空间通常包括买入、卖出或空仓等待，其目标是根据市场订单簿状态和盈亏反馈持续优化行为。 现场交易演示与滑点及执行摩擦的现实风险 (3:35) 在现场演示中，团队根据现场观众提议选择了联合健康集团（UNH）。由于主账户出现身份验证与权限问题，团队切换至反映真实市场条件的演示账户，模型以259.70美元执行买入。Avi Thaker强调，初学者构建的模型往往在回测中表现优异，但在实盘中因低估滑点（Slippage）、执行延迟及交易所费用而转为净亏损；该演示最终完成平仓并锁定99美元利润，但仅作为概念验证，不构成投资建议。 回撤数学原理与传统分散化投资的局限 (5:48) Daniel Sonner通过大学捐赠基金（如耶鲁大学模式）的资产配置变迁指出，长期机构资金正大幅减少对美国传统股票的依赖，转向包含量化对冲基金的另类资产。其核心驱动力在于回撤控制：当投资组合遭遇50%的深度回撤时，需要100%的收益才能回本，通常需耗费5至10年。Daniel解释，仅依靠传统60/40股债组合存在隐患，因为历史滚动数据显示股债相关性高度不稳定，通胀或加息周期中二者常同步下跌；机构投资者必须寻求结构上真正不相关的策略以平滑波动曲线。 市场微观结构壁垒与系统工程的技术负债 (8:22) 独立零售算法交易者面临严峻的微观结构阻力。Avi Thaker引用Robinhood在2024年9月的订单流支付（PFOF）数据指出，Citadel Securities、Two Sigma、Virtu Financial和Wolverine Trading等高频做市商斥巨资购买散户订单流并在内部撮合，使散户算法处于执行劣势。尽管极少数顶尖机构能实现长年几乎无亏损日，但这依赖于极低延迟的基础设施和特权数据流，普通投资者仅凭随机方差也可能短期表现优异。此外，根据Google关于机器学习隐性技术负债的研究，核心ML模型仅占实盘量化系统约5%，其余大量工作集中在数据清洗、特征管道、风险管理、日志监控与回测验证，孤立的技术指标（如RSI或均线交叉）在高效市场中毫无统计学优势。 金融领域AI代理的具体落地场景 (13:18) Daniel Sonner总结了四类量化AI代理的应用方向：跨交易所套利（如利用不同交易场所的瞬时价差赚取无风险利差）、做市（在订单簿双侧同时挂买单与卖单赚取买卖价差）、基于深度神经网络（DNN）的重大事件与选举次日资产价格变动预测，以及利用农田卫星图像与期货时间序列数据预测小麦产量和定价的另类数据预测模型。 演讲者总结指出，现场演示的代码仅为轻量化的教育性起点；若在实盘生产环境中集成更大规模的模型，必须在模型复杂度与执行延迟、系统响应能力之间做出严格权衡。

Detailed notes — zh

## Background 该视频由 Ai4 频道于 2025 年 8 月 27 日发布在 YouTube 平台。内容记录了 Tauroi Technologies 联合创始人 Daniel Sonner 与 Avi Thaker 在 Ai4 峰会上的演讲，展示了利用 AI Agent 在真实金融市场中进行实盘交易的全过程，并探讨了对冲基金如何利用另类数据构建实时交易智能体。 ## TL;DR 两位讲者以吃豆人强化学习为切入点，现场部署深度 Q 学习智能体对联合健康（UNH）股票进行了真金白银的实盘交易演示。随后，他们深入剖析了耶鲁捐赠基金资产配置逻辑、做市商订单流支付（PFOF）机制以及机器学习落地中数据管道与回测风控的技术现实。团队还分享了利用大选预测与农田卫星遥感图像构建另类数据交易模型的实战案例，最终实盘演示斩获 99 美元收益并兑现观众抽奖，同时强调了生产环境中平衡模型推理延迟的重要性。 ## Atmosphere & style 整场演讲充满极强的互动感与现场张力，真金白银的实盘开仓与收益抽奖迅速点燃了观众的热情。两位主讲人台风轻松幽默、配合默契，既有调侃加州税率的妙语连珠，也有拆解做市机制与市场微观结构时的清醒与犀利。现场在硬核工程探讨与击碎暴利神话的理性警示之间自如切换，营造出兼具学术严谨与实战刺激的极客氛围。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=xzUpSPZzMys&t=4s) Tauroi Technologies 联合创始人在 Ai4 大会现场演示了基于真金白银账户的 AI 交易智能体，并承诺将实盘收益的 50% 现场抽奖发放给观众。 - [01:31](https://www.youtube.com/watch?v=xzUpSPZzMys&t=91s) 由于真实金融市场的状态空间无限且不可完全观测，交易智能体需采用深度 Q 学习（Deep Q-learning）神经网络来逼近复杂的市场状态与价值函数。 - [02:51](https://www.youtube.com/watch?v=xzUpSPZzMys&t=171s) 滑点（Slippage）与交易手续费是量化交易中最容易被低估的隐形成本，微小的执行摩擦往往会对高频交易策略的实际净利润产生决定性侵蚀。 - [05:44](https://www.youtube.com/watch?v=xzUpSPZzMys&t=344s) 投资组合发生 50% 的大幅回撤需要 100% 的涨幅才能回本；耶鲁捐赠基金模式表明，引入低相关的另类策略可以在保持总回报的同时平滑净值曲线并抑制回撤。 - [09:07](https://www.youtube.com/watch?v=xzUpSPZzMys&t=547s) 独立散户交易面临极高的亏损与淘汰率，进入买方对冲基金或卖方投资银行是量化交易从业者更稳健的职业路径选择。 - [09:59](https://www.youtube.com/watch?v=xzUpSPZzMys&t=599s) 免佣券商大多依赖订单流支付（PFOF）机制将散户订单打包出售给顶级做市商套利，投资者实盘中必须预留这部分隐形成本。 - [10:55](https://www.youtube.com/watch?v=xzUpSPZzMys&t=655s) 战胜市场必须依赖真实的信息或通道优势，策略的有效性需要在海量独立交易样本中得到统计学显著性验证，以剔除纯靠运气的幸存者偏差。 - [12:04](https://www.youtube.com/watch?v=xzUpSPZzMys&t=724s) 工业级量化交易系统中核心机器学习算法仅占极小体量，绝大多数工程工作量集中在数据采集清洗、特征管道、回测风控与日志监控基础设施上。 - [12:04](https://www.youtube.com/watch?v=xzUpSPZzMys&t=724s) 单独依赖传统技术指标几乎无法构建有效策略，因为公开的技术指标规律早已被顶级量化机构充分套利殆尽。 - [13:18](https://www.youtube.com/watch?v=xzUpSPZzMys&t=798s) 垂直金融智能体的两大经典应用是跨市场无风险/低风险套利（Arbitrage），以及基于订单簿双向挂单赚取微观买卖点差的做市策略（Market Making）。 - [15:05](https://www.youtube.com/watch?v=xzUpSPZzMys&t=905s) 利用大选预测特征或卫星遥感图像等非结构化另类数据（Alternative Data），智能体能够先于传统市场参与者捕捉供需与情绪变化并执行多空策略。 - [16:06](https://www.youtube.com/watch?v=xzUpSPZzMys&t=966s) 实盘部署交易智能体必须严格权衡模型精度与推理延迟（Latency），过于庞大复杂的模型会因执行响应过慢而彻底丧失微秒级套利窗口。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=xzUpSPZzMys&t=4s) 讲者背景与现场抽奖规则 Tauroi Technologies 的两位联合创始人 Avi Thaker 与 Daniel Sonner 在拉斯维加斯 Ai4 大会现场登台亮相。Avi 首先介绍了个人背景：他拥有多项技术专利，曾在微软（Microsoft）和葛兰素史克（GSK）工作，通过了多项金融执业资格考试，目前与 Daniel 共同管理一家对冲基金。Daniel 则拥有资深的系统工程背景，此前在 Pure Storage 任职并持有相关专利；他还幽默地透露了一个个人趣事（fun fact）——他曾获得过桥牌世界冠军（Bridge World Championship）。 为了让讲座更具实战感与趣味性，两人宣布将在现场使用真实资金账户进行 AI Agent 实盘交易演示。他们还在大屏幕上展示了一个限时填写的抽奖二维码（Raffle QR code），向现场观众分发实盘收益的 50%。Avi 开玩笑解释道，因为自己住在加利福尼亚州，身处高税率区间，必须承担高达 50% 的税负；如果交易亏损，中奖观众也将获得至少 20 美元的保底奖金。Avi 强调整场演讲保持高度互动性，鼓励观众随时提问打断，“没有愚蠢的问题。我也许会给出愚蠢的回答，但我非常期待听到大家的声音。” - **Avi Thaker**：Tauroi Technologies 联合创始人，曾就职于 Microsoft、GSK，持有多项专利及金融执业认证。 - **Daniel Sonner**：Tauroi Technologies 联合创始人，曾任 Pure Storage 系统工程师，持有相关专利，桥牌世界冠军。 - **互动抽奖规则**：现场演示交易收益的 50% 赠予现场观众，若交易亏损则提供 20 美元保底奖励。 --- ### [01:31](https://www.youtube.com/watch?v=xzUpSPZzMys&t=91s) 什么是 AI Agent：从吃豆人到深度 Q 学习 Avi 深入剖析了 AI Agent（智能体）的核心定义：Agent 是一种具备自主规划、工具调用以及自主执行能力的系统。根据应用场景的复杂度，Agent 可以划分为专精于特定单一任务的窄领域智能体（narrow agents）和应对多样化复杂任务的广义智能体（broad agents）。在本次演示中，团队构建了一个高度专注且边界清晰的窄领域金融交易 Agent。 在解释 Agent 的学习机制时，Avi 以经典视频游戏《吃豆人》（Pac-Man）作为直观类比。吃豆人在吃到豆子或幽灵时会获得正向奖励（reward），而如果在迷宫中停滞不前则会遭受分数扣除的惩罚（penalty）。传统强化学习常采用 Q-learning 算法，维护一张覆盖所有状态与行动的 Q 表（Q-table）来决定下一步的最优动作。然而在真实的金融市场中，状态空间近乎无限且无法完全观测，因此必须引入深度 Q 学习（Deep Q-learning），利用神经网络来逼近复杂的市场状态与价值函数。 ```mermaid flowchart TD accTitle: 深度强化学习交易智能体架构 subgraph S1["环境与状态感知 (Environment & States)"] A["市场实时数据流"] --> B["无限维状态空间"] end subgraph S2["深度 Q 学习核心 (Deep Q-Network)"] B --> C["神经网络状态逼近"] C --> D["动作价值评估 Q(s,a)"] end subgraph S3["执行与反馈循环 (Action & Reward)"] D --> E{"执行交易决策"} E -->|"买入 / 卖出 / 观望"| F["经纪商 / 市场执行"] F -->|"收益 / 滑点惩罚"| G["反馈奖励更新网络"] G -.-> C end ``` - **动作空间（Action Space）**：买入（Buy）、卖出（Sell）、观望（Wait / Hold）。 - **奖励函数（Reward Function）**：累计获得的总美元收益（Total dollars earned），同时计入收益方差与执行惩罚。 - **代码开源**：现场提供了 GitHub 代码仓库链接供开发者研究与复现，但讲者明确强调仅作为概念验证（Demo）参考。 --- ### [02:51](https://www.youtube.com/watch?v=xzUpSPZzMys&t=171s) 现场实盘演示：交易 UNH 与滑点成本现实 Avi 现场向观众征集交易标的，并郑重向全场发出免责声明：“请勿在家中盲目尝试，这绝非投资建议，仅为演示性质。尽管我们公开了代码且它能直连市场，但用真金白银操作非常危险。”现场观众举手点名了联合健康集团（UnitedHealth Group, 股票代码：UNH）。 在将演示终端字体放大后，Agent 开始连接市场数据并进行初始化。在初始阶段，模型研判市场波动较小，选择保持空仓观望（Hold/Flat）。由于主交易账户的实时权限配置出现偶发问题，团队迅速切换至备用执行账户继续演示。程序随后在 259.7 美元的价位触发了买入指令。Avi 借此向观众重点揭示了量化交易中常被忽视的隐形成本——滑点（slippage）、交易手续费及各类通道杂费，这些摩擦成本对高频策略的实际盈亏（P&L）有着决定性影响。随后实盘程序保持在后台持续运行，以跟踪整场演讲期间的实时盈亏变化。 - **选中标的**：联合健康（UnitedHealth Group, UNH）。 - **成交点位**：买入执行价 259.7 美元。 - **交易摩擦**：强调滑点损失（Slippage Cost）与经纪商交易费用对高频决策模型的侵蚀。 --- ### [05:44](https://www.youtube.com/watch?v=xzUpSPZzMys&t=344s) 耶鲁捐赠基金模式与回撤控制：为何配置另类资产 Daniel 接过话筒，深入剖析了此类系统化投资在宏观资产配置中的核心价值。他指出，以耶鲁大学捐赠基金（Yale Endowment）为代表的顶级机构投资者，在过去数十年间显著重构了资产配置结构。历史数据显示，机构配置中传统的美国公开市场股票（US Public Equity）比例持续大幅缩减，而以对冲基金为代表的另类资产（Alternatives，其中大量采用 AI 与量化模型驱动）比例则稳步攀升。 ```mermaid flowchart LR accTitle: 投资组合非相关资产平滑回撤示意 subgraph Assets["基础资产走势"] A["标普500 (S&P 500) 高回撤"] B["非相关另类策略 高波动"] end subgraph Combined["资产组合构建 (50/50 权重)"] A & B --> C["组合净值曲线 (低回撤/平滑收益)"] end ``` Daniel 阐释了这一转变背后的核心数学逻辑：**大幅回撤极度痛苦且毁灭复利**。如果一个投资组合发生 50% 的回撤，则需要实现 100% 的涨幅才能回本；在历史均值下，弥补这一损失往往需要耗费 5 到 10 年甚至更久。通过将标普 500 指数与一个长期收益相当但彼此负相关或低相关的另类策略进行 50/50 组合构建，虽然期末的总回报率相同，但合成后的净值曲线几乎完全抹平了深度回撤。此外，针对传统利用“股票+债券”构建防守组合的做法，Daniel 援引债券与标普 500 历史相关性图表指出，股债相关性随着宏观周期剧烈漂移、极不稳定，而真正的另类投资正是为了提供持续、稳健的低相关性 Alpha。 - **回撤复原成本**：组合下跌 50% 需要 100% 的反弹方可回本，平均修复周期长达 5 至 10 年。 - **耶鲁模型启示**：机构资本大幅削减传统美股多头权重，持续加码 AI 与量化驱动的另类策略。 - **组合构建价值**：利用持续低相关资产对冲系统性风险，在保持总回报一致的前提下显著降低波动与心理痛苦。 --- ### [09:07](https://www.youtube.com/watch?v=xzUpSPZzMys&t=547s) 金融职业路径与隐藏交易费用（PFOF） 在演讲尾声阶段，Avi 梳理了进入量化与交易行业的几种典型路径。他指出，成为一名独立散户交易者（independent trader）极其艰难，绝大多数人都处于亏损状态。他更建议从业者走向机构化路径：进入买方（Buy Side）如对冲基金，或者进入卖方（Sell Side）如投资银行积累实战经验。 针对当前市场上各类宣称“轻松暴利”的 AI 交易 Agent，Avi 给出了清醒的警示：必须严密审视背后的费用结构与微观机制。他以美国免佣券商 Robinhood 2024 年 9 月的订单流支付（Payment for Order Flow, PFOF）收入为例，准备向现场观众揭开所谓“零佣金”交易背后散户所承担的隐形执行成本与做市商博弈真相。 - **从业路径选择**：独立散户交易（高淘汰率） vs. 买方对冲基金 / 卖方投行（注重经验积累）。 - **执行机制隐患**：警惕以 Robinhood 为代表的订单流支付（PFOF）机制对零售策略成交质量的潜在影响。 ### [09:59](https://www.youtube.com/watch?v=xzUpSPZzMys&t=599s) 订单流支付（PFOF）与交易执行成本的隐形成本 散户投资者在 Robinhood 等免佣金券商平台点击下单时，往往忽略了交易背后的资金流向与市场微观结构。券商并不会把订单直接送往公开交易所撮合，而是打包转卖给 Two Sigma、Virtu Financial、Citadel 以及 Wolverine Trading 等高频做市商。这些大型量化做市商通过承接这些散户订单获取巨额利润，这种机制在金融行业被称为“订单流支付”（Payment for Order Flow，简称 PFOF）。演讲者展示了做市商为每种订单类型支付的具体金额以及他们借此赚得的利润规模，并直言这个数字是极其惊人的“天文数字”（utterly crazy），而且这仅仅是一个随机抽样的时间截面，如果将其乘以 12 个月或放在市场成交量放大的时期，规模还会成倍扩张。 对于任何尝试构建交易策略的人而言，核心前提是必须假定每一笔交易都会因滑点（slippage）和执行成本而损失资金。如果不透彻理解交易策略到底在赚谁的钱、为什么能赚钱，以及能否有效规避被做市商收取的各类隐形成本，模型在实盘中就会迅速被蚕食吞没。Citadel 等行业巨头之所以能如同“印钞机”般源源不断地盈利，正是因为它们手握雄厚资金持续收购散户订单流并从中赚取微观价差。 - 券商通常将散户订单流打包路由给 Citadel、Virtu Financial、Two Sigma、Wolverine Trading 等顶级量化做市商。 - 散户在实盘交易中必须预留滑点、交易费用与执行摩擦成本，否则理论收益会在执行端迅速被抵消。 - PFOF（订单流支付）构成了现代券商与高频做市商的核心商业闭环。 --- ### [10:55](https://www.youtube.com/watch?v=xzUpSPZzMys&t=655s) 有效市场假说与统计显著性：如何区分实力与运气 有效市场假说（Efficient Market Hypothesis）主张“价格应当充分反映所有已知信息”（"prices should reflect all information"），根据这一理论，战胜市场几乎是不可能的。尽管学术界对弱式、半强式和强式有效市场假说争论不休，但现实中确实存在破局者：沃伦·巴菲特（Warren Buffett）凭借跨越数十年的超额收益证明了 Alpha 的存在；Citadel 与 Virtu Financial 等顶级量化机构更是创造了数年内“几乎没有任何亏损日”的神话。然而，它们之所以能够击败市场，是因为拥有领先于市场其他参与者获取的信息优势，或是执行速度更快、数据质量更高的交易通道。 如果开发者希望通过构建 AI 模型或量化策略来击败市场，绝不能仅仅停留在“训练一个模型然后跑去赚大钱”的盲目自信中。投资者必须在统计学层面证明自己的模型确实拥有信息优势、数据优势或算法优势。市场中存在大量单纯靠运气成功的幸存者，正如那个关于“让一群猴子掷飞镖挑选股票也会有人跑赢大盘”的经典隐喻。甚至有著名的投机家曾宣称，哪怕自己多次经历爆仓破产，却仍然自诩是正确的，其业绩在统计分布上偏离了常态分布多达 700 个标准差。想要证明策略并非纯属侥幸，唯一的途径就是通过时间检验，在海量独立交易样本中积累出具有统计学显著性的持续盈利能力。 --- ### [12:04](https://www.youtube.com/watch?v=xzUpSPZzMys&t=724s) 机器学习系统全貌：技术指标的局限与回测风控 很多人误以为量化交易的核心是设计极其前沿复杂的 AI 或深度学习算法，但现实恰恰相反。演讲者引用了 Google 一篇极具代表性的论文架构图（源自经典的《机器学习系统中隐藏的技术债》）指出：机器学习或深度学习模型代码本身，往往只是庞大工程基础设施中微不足道的一个小方块。真正占据系统绝大部分体量的是特征工程、数据采集清洗、分析验证工具、监控警报、日志记录以及合规报告流水线。对于有志于进入量化投资行业的工程师而言，在数据流水线、日志追踪或特征管道等任何一个具体模块深耕，其工业价值都远胜于执迷于“自己是否发明了最炫酷的算法”。 任何实盘系统上线之前都必须经过严密的回测检验（Backtesting）。开发者必须清晰回答核心问题：系统到底依靠什么机制赚钱？预期收益是多少？投资者的风险承受能力（Risk Tolerance）上限在哪里？系统历史上经历过怎样的最大回撤（Drawdowns）？如果实盘遇到的回撤幅度远超历史预期，系统该如何应对？与此同时，演讲者对单纯依赖传统技术指标（如各类均线、震荡指标）提出了严厉警示，直言这些指标如果单独使用“绝大多数毫无用处”（mostly useless on their own）。即便借助机器学习算法去挖掘这些技术指标，往往也难以获得真实的超额收益，因为市场上最顶尖的量化机构早已在这个赛道将所有浅层规律套利殆尽。 ```mermaid flowchart TD accTitle: 机器学习交易系统工程全貌 subgraph DataInfra["数据工程与底座"] D1["数据采集与清洗"] --> D2["特征工程管道"] D2 --> D3["元数据与分析工具"] end subgraph CoreML["核心模型层 (仅占极小体量)"] D3 --> M1["机器学习 / 深度学习算法"] end subgraph RiskProd["生产保障与风控"] M1 --> R1["回测与最大回撤评估"] R1 --> R2["日志记录与实时监控"] R2 --> R3["执行滑点与风控上报"] end ``` --- ### [13:18](https://www.youtube.com/watch?v=xzUpSPZzMys&t=798s) 金融专属智能体应用：套利与做市策略 随后，Avi Thaker 登台详细拆解了 Tauroi Technologies 如何将金融垂直领域的 AI 智能体应用到具体的策略场景中。第一类最经典的传统策略是**跨市场套利（Arbitrage）**。该策略的核心逻辑是“低买高卖”不同交易所之间的同一标的资产。以比特币等加密资产为例，如果某数字资产在 A 交易所的即时价格低于 B 交易所的挂牌价格，智能体便能在 A 交易所买入低价资产，同步在 B 交易所卖出高价资产，随后完成资产转账与对冲，从而无风险或低风险地锁定净利润差额。 第二类则是高度依赖订单流的**做市策略（Market Making）**。Avi 借助经典的限价订单簿（Order Book）展示了做市商赚取买卖价差（Bid-Ask Spread）的具体机制： - 订单簿左侧为买盘挂单（Bid），假设当前市场上最佳买价为 99.99 美元。 - 订单簿右侧为卖盘挂单（Ask），当前市场上最佳卖价为 100.00 美元。 - 做市智能体会同时挂出两笔订单：以 99.99 美元挂买单，同时以 100.00 美元挂卖单。 - 当有散户或流动性索取者提交“以市价立即卖出”的市价单时，该卖单会与做市商 99.99 美元的买单成交。 - 紧接着当另一位交易者提交“以市价立即买入”的市价单时，该买单会与做市商 100.00 美元的卖单成交。 - 只要双向市价订单大致均衡地进场，做市商就能在这两次撮合中净赚 1 美分（1 Penny）的稳定点差。通过高频智能体将这类交易重复成千上万次，即可稳定抽取订单流利润。 --- ### [15:05](https://www.youtube.com/watch?v=xzUpSPZzMys&t=905s) 另类数据驱动模型：大选预测与卫星图像分析 除了依靠微观结构的高频策略，另类数据（Alternative Data）也是智能体战胜市场的重要抓手。Tauroi Technologies 团队重点介绍了他们此前发表过学术论文且开源在 GitHub 上的两个实际研究案例： 第一个研究是**大选结果与市场联动预测**。团队训练了一个深度神经网络（DNN），汇集各类选情特征变量，不仅用于预测大选本身的胜选概率，更深入挖掘大选结果揭晓后第一天的市场联动冲击（One-day Market Impact）——即特定候选人胜选后，各类资产在次日交易日会做出何种方向与幅度的价格反馈。 第二个研究是**利用遥感卫星图像预测农产品期货**。团队抓取了小麦农田的高空卫星鸟瞰图像，结合农产品期货的历史价格波动序列，训练机器学习模型对小麦未来的供需缺口及价格走势进行前瞻预测。智能体根据预测结果直接在期货市场上执行交易操作：如果模型预判减产导致价格上行则建立多头头寸（Long）；如果预判增产价格回落则建立空头头寸（Short）。Avi 强调，这些案例旨在抛砖引玉（"get the juices flowing"），展示如何将外部非结构化现实数据转化为可执行的智能体交易策略。 ```mermaid flowchart LR accTitle: 卫星图像农产品期货预测流水线 subgraph DataIngestion["另类数据多模态采集"] S1["高空农田遥感卫星影像"] S2["小麦期货历史行情数据"] end subgraph MLPipeline["特征处理与 DNN 模型"] S1 --> P1["多模态融合模型"] S2 --> P1 P1 --> P2["小麦未来价格走势预测"] end subgraph TradeSignal["智能体执行动作"] P2 -->|"看涨信号"| T1["做多合约 (Long)"] P2 -->|"看跌信号"| T2["做空合约 (Short)"] end ``` --- ### [16:06](https://www.youtube.com/watch?v=xzUpSPZzMys&t=966s) 实盘演示收益揭晓与轻量级模型的生产考量 在演讲的最后阶段，Daniel 与 Avi 将大屏幕画面切回现场实时运行的交易智能体演示界面。经过这十余分钟的现场实时盯盘，该交易智能体在实盘环境中累计斩获了 99 美元的净利润（P&L）。Daniel 当场兑现了开场的抽奖承诺，宣布会向现场抽中幸运签的观众赠送一张 99 美元的亚马逊礼品卡。 Daniel 随后展示了后台运行的策略代码与模型架构，并特别向观众强调，今天在现场演示的仅仅是一个体量非常小巧的 Demo 模型。虽然在离线研发时工程师可以堆叠更大、更复杂的深度模型以追求更高的指标，但在真正的生产实盘环境（Production）中，团队必须时刻权衡推理延迟（Latency）与网络开销的物理现实。如果模型体积过大导致算力延迟增加，哪怕算法给出的信号再完美，也会因为无法在微秒或毫秒级时间内完成报单而彻底丧失套利空间。由于演讲时间耗尽，两位创始人表示非常乐意在台下与大家继续交流离线问答，并在掌声中结束了本次演示。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=xzUpSPZzMys&t=4s) 没有愚蠢的问题。我也许会给出愚蠢的回答，但我非常期待听到大家的声音。 — Avi Thaker > [02:51](https://www.youtube.com/watch?v=xzUpSPZzMys&t=171s) 请勿在家中盲目尝试，这绝非投资建议，仅为演示性质。尽管我们公开了代码且它能直连市场，但用真金白银操作非常危险。 — Avi Thaker > [05:44](https://www.youtube.com/watch?v=xzUpSPZzMys&t=344s) 大幅回撤极度痛苦且毁灭复利。 — Daniel Sonner > [09:59](https://www.youtube.com/watch?v=xzUpSPZzMys&t=599s) utterly crazy — Avi Thaker > [10:55](https://www.youtube.com/watch?v=xzUpSPZzMys&t=655s) prices should reflect all information — Avi Thaker > [10:55](https://www.youtube.com/watch?v=xzUpSPZzMys&t=655s) 训练一个模型然后跑去赚大钱 — Avi Thaker > [12:04](https://www.youtube.com/watch?v=xzUpSPZzMys&t=724s) mostly useless on their own — Daniel Sonner > [15:05](https://www.youtube.com/watch?v=xzUpSPZzMys&t=905s) get the juices flowing — Avi Thaker ## Mentioned Tauroi Technologies, Avi Thaker, Daniel Sonner, Ai4, Microsoft, GSK (葛兰素史克), Pure Storage, UnitedHealth Group (UNH), Yale Endowment (耶鲁大学捐赠基金), S&P 500 (标普500), Robinhood, Two Sigma, Virtu Financial, Citadel, Wolverine Trading, Warren Buffett (沃伦·巴菲特), Google, GitHub, Pac-Man (吃豆人), Amazon (亚马逊) ## Video description Daniel Sonner, Co-Founder, Tauroi Technologies Avi Thaker, Co-Founder, Tauroi Technologies Abstracts This talk will showcase an AI agent trading markets live, and will raffle off 50% of the winnings! Explore how hedge funds use alternative data to build realtime agents to trade financial markets! [Source](https://www.youtube.com/watch?v=xzUpSPZzMys)

Source quotes

[0:04] Avi Thaker: 没有愚蠢的问题。我也许会给出愚蠢的回答，但我非常期待听到大家的声音。 [2:51] Avi Thaker: 请勿在家中盲目尝试，这绝非投资建议，仅为演示性质。尽管我们公开了代码且它能直连市场，但用真金白银操作非常危险。 [5:44] Daniel Sonner: 大幅回撤极度痛苦且毁灭复利。 [9:59] Avi Thaker: utterly crazy [10:55] Avi Thaker: prices should reflect all information [10:55] Avi Thaker: 训练一个模型然后跑去赚大钱 [12:04] Daniel Sonner: mostly useless on their own [15:05] Avi Thaker: get the juices flowing

Source quotes

[3:35] Avi Thaker: Please don't try this at home. This is not financial advice and this is a demo. [10:55] Avi Thaker: If you had a bunch of monkeys throw darts or randomly selected, some people were just going to win. [16:31] Speakers: much larger models

Original transcript

[0:04] Hi everyone. Hope you're having an [0:06] amazing time in Vegas. I'm Avi. I'm [0:08] Daniel. [0:10] And we're going to talk about agents in [0:11] finance. We'll give a bit of background [0:13] on ourselves, then we'll dive into the [0:15] talk. I'm AI. I have some patents. I run [0:18] a hedge fund with Daniel. I've been in [0:20] the industry for a little bit. I was at [0:22] Microsoft, at GSK, and it's also [0:24] important to note that I passed some of [0:26] the licensing exams. Uh, I'm Daniel. My [0:29] background is I used to work at Pier [0:30] Storage doing systems engineering there. [0:32] I have some patents for my work there. [0:34] And uh, fun fact about me is I've won a [0:37] Bridge World Championship. [0:40] Um, so today we're going to have a live [0:43] demo of trading real money in in my [0:46] account. Uh, we're going to distribute [0:48] the winnings via an auction. So feel [0:51] free to sign out the QR code right here. [0:53] >> Sorry, a raffle. A raffle. Raffle. Um, [0:56] and it's going to be computed basically, [0:58] uh, someone will win at least 20 bucks. [1:01] And if we lose money, you'll win at [1:02] least 20 bucks. And because I live in [1:05] California and pay a lot of taxes, I got [1:07] to I got to pay my 50% tax tax bracket. [1:11] So feel free to scan this. This QR code [1:14] won't be up anymore, so scan it right [1:16] now. [1:17] The other thing I want to mention is [1:19] this talk is meant to be interactive. So [1:21] feel free to ask questions. Interrupt [1:23] me. No question is stupid. I might give [1:25] a stupid answer, but I'm excited to hear [1:27] what you guys have had to ask us. [1:31] All right. So, what is an agent? An [1:35] agent is something that's like [1:36] autonomous. It can plan and execute. It [1:40] can use tools. And there's varying [1:42] levels of agent. There are narrow [1:43] agents. There are broad agents. They can [1:45] have many actions that they can plan. [1:47] Uh, in our case today, we're going to [1:49] demo a very narrowcoped agent that we [1:51] think is going to be interesting and [1:52] exciting and show you how to use it. [1:56] So, how does an agent learn? I think of [1:59] agents, how they interact in video [2:01] games. A lot of the first agents [2:03] actually worked in video games. And so, [2:05] you can see here, this is Pac-Man. He [2:07] gets a reward every time he eats some [2:09] dots and every time he eats a ghost. And [2:12] you can see that oh there's a penalty of [2:13] if he's not doing anything his score [2:16] starts decreasing. You can see that [2:18] decreasing right now. And so one of [2:21] these ways that you can train an agent [2:22] to do this is you can use something [2:23] called a Q-learning algorithm. And a [2:25] Q-learning algorithm basically has a [2:27] table for all these actions and states [2:28] that you can be in. We update this table [2:30] and it says hey given this what do we do [2:32] next? We're going to do something called [2:34] deep Q because well in markets there [2:36] could be many many states that you have [2:38] and you can't see all of them and so you [2:40] have to learn to approximate them and so [2:42] agents uh interact with an environment [2:45] and we're going to show you our agent. [2:47] Any questions here before we dive into [2:48] it. [2:51] Great. Uh so what is our agent? Our [2:54] agent has an environment. Uh we're going [2:56] to select the stock. Um I'm going to [2:58] pull the audience and say pick a stock. [3:00] We're going to trade it. Um all data [3:02] will be pulled before it starts trading. [3:04] And the agent can do two things. It can [3:06] buy an order, sell an order, and it can [3:08] also wait. [3:10] Uh the reward is total dollars earned. [3:13] Um and when we do stuff, it's subject to [3:14] some variance. In this case, it's a very [3:16] simple agent, so it can just do total [3:18] dollars earned. Um there's a GitHub link [3:20] where you can actually try the agent out [3:22] yourself um and see the code that that's [3:26] being used. Again, it's a demo use case, [3:29] so just remember it's a starting point. [3:35] Um, so let's try this with real money. [3:38] Please don't try this at home. This is [3:39] not financial advice and this is a demo. [3:43] Even though we're giving you code and it [3:45] can work directly with the market, I [3:47] encourage you to not try it with real [3:49] money because it's dangerous. Um, [3:52] so now is the time to switch to this [3:54] screen and I want you guys to pick from [3:58] this list of stocks or any stocks. You [4:01] might want to make that way bigger. [4:02] >> Oh, yeah. Let me make that bigger. Um, [4:05] plus [4:05] >> command plus. [4:07] Uh, anyone in the audience, first person [4:09] I see, raise their hand. All right. You, [4:11] um, what what stock should we trade? [4:14] You United Healthcare. Uh, [4:19] >> sure. [4:21] picking off the menu. All right, so [4:23] we're going to trade United Healthcare. [4:25] Um, so what we're going to do is we are [4:28] going to run this. [4:31] And this is trading live, real money, [4:33] real account. You can see right here. [4:36] And so now we've not done anything yet. [4:38] Let's make this bigger. [4:41] And it'll start making trades. So right [4:43] now it's saying, hey, we're going to [4:45] choose hold our position to be flat. and [4:47] hopefully it's going to decide to start [4:49] making a trade soon. [4:51] And so we'll give it a minute. This will [4:52] cycle through as we give the [4:54] presentation and talk. Um, so it thinks [4:57] nothing's happening. It's not a very [4:59] exciting stock to trade right now. It's [5:02] basically saying, um, I have to enter [5:04] some permissions on there, so we'll just [5:05] use a different account. And [5:10] unfortunately, this won't be with real [5:11] dollars. Uh, but you'll still make it'll [5:13] still show you roughly the same amount [5:14] of money. All right, so this will trade [5:15] now. Um, here we go. [5:20] So, now we bought UNH at 259.7. [5:23] And right here, you're seeing the [5:25] slippage cost to make a trade. One of [5:27] the important things people don't see is [5:29] when you trade, there's a huge amount of [5:31] slippage cost, transactions, and fees [5:32] that occur while you're trading. So, I'm [5:35] going to start letting this run in the [5:36] background while we continue the [5:37] presentation, and you'll see how much [5:39] P&L will generate over time. Could we [5:41] switch back to the uh to the [5:44] presentation? So here's the code again [5:47] and we're going to talk about why this [5:48] matters. So why does this type of [5:52] investing matter? One reason it matters [5:54] is it's how endowments and invest. And [5:57] so this is a chart showing over time the [6:00] allocation that endowments make between [6:02] different various asset asset classes. [6:04] In green you can see alternatives that [6:06] would be things like hedge funds like [6:08] the types of strategies that we're [6:10] showing today. And many of those are AI [6:12] powered. And in blue you can see the US [6:14] public equity exposure. And one [6:16] interesting thing you see with this [6:18] chart is that over time there's been an [6:20] increasing exposure to alternatives and [6:21] a decreasing exposure to US equities. [6:24] This was a study of endowments. Just [6:26] looking again at one particular [6:27] endowment. Yale is one that many [6:29] endowments are modeled after. And it's a [6:32] similar sort of picture where again in [6:33] green you see various alternatives that [6:35] are being allocated to many of which are [6:37] AI powered and then in blue you see US [6:40] equity allocation has sharply decreased [6:43] over time. [6:45] So how does this affect a person's [6:47] portfolio or a person how a person might [6:49] invest? It matters a lot because draw [6:52] downs are painful. So even though you [6:54] might end up with similar overall [6:56] returns if the path that you take to get [6:59] there matters a lot. For example, if you [7:01] had a 50% draw down, it would take a [7:04] 100% gain to make that back. And on [7:06] average, that can take 5 to 10 years and [7:08] sometimes more. That's an extremely [7:11] painful 5 to 10 years to weather. So, [7:14] looking at an example here on the right, [7:16] you can see a chart where in blue [7:18] there's the S&P 500. In yellow there's a [7:23] hypothetical uncorrelated strategy. [7:25] uncorrelated because when the S&P tends [7:27] to go up, the uncorrelated strategy [7:29] tends to go down. And when the S&P goes [7:32] down, the uncorrelated strategy tends to [7:34] go up. So, it's an uncorrelated [7:36] strategy. If you look at investing in [7:38] either one of those lines, they'll end [7:40] up at the same place, the same overall [7:42] return at the end of the period, but the [7:44] path they take to get there is [7:45] different. And both the blue and the [7:47] yellow line have substantial drawd [7:49] downs, substantial periods of pain where [7:51] you're down from where you previously [7:53] were over the course of many years. With [7:56] portfolio construction, you could invest [7:58] half your money in in each of them and [8:01] then you'd end up with the red line. You [8:02] end up at the same point, the same total [8:04] return, but now the path there has [8:06] almost no draw downs. You've eliminated [8:09] a lot of your pain through portfolio [8:11] construction. So that's the real sales [8:13] point for alternatives and why people [8:16] find them an appealing investment [8:17] because you can lower your overall draw [8:19] down and your overall pain. [8:22] One way people sometimes try to get some [8:24] uncorrelation is they do something like [8:26] buy S&P 500 and buy bonds. That's the [8:29] chart on the right there where on the y- [8:30] axis you can see the correlation to the [8:33] S&P 500 and the x- axis is over time and [8:36] then in green you can see uh bonds [8:39] correlation to the S&P 500. And the main [8:42] takeaway there is it's not consistent. [8:44] Sometimes it's correlated, sometimes [8:46] it's uncorrelated, sometimes it's [8:47] somewhere in the middle. For portfolio [8:50] construction, alternatives are really [8:52] appealing because you can try to find [8:53] things that are consistently [8:55] uncorrelated with your other investments [8:57] and combine those again to get something [8:59] a bit more like the red line where you [9:01] avoid a lot of pain and a lot of draw [9:03] downs. [9:07] So now we're going to talk about how do [9:09] people get into the industry, what they [9:10] want to do. There's different cycles and [9:12] different paths that you can go towards. [9:15] Um, typically you start, do you want to [9:17] go be an unprofitable independent trader [9:20] and that's really really hard if you [9:22] want to do it. You can go to a hedge [9:24] fund, you can go on on on the buy side, [9:26] you can go to go to a bank on the sell [9:28] side and there's lots of different paths [9:29] here. Here's the way that people can [9:31] take. Um, I would encourage you to get [9:34] lots of experience because experience [9:35] tends to be one of the best ways uh to [9:37] do it and help you get started in the [9:40] area. So, the other thing I want to talk [9:42] to you about is all these agent things [9:45] sound very cool. You can make a lot of [9:47] money, but you really, really want to [9:49] think about fees. So, I'm going to pop [9:51] up how much Robin Hood made in September [9:54] 2024 [9:55] um from payment for orderflow. And Robin [9:59] Hood sends money, sends your trade to a [10:03] two sigma, a Veru, a Citadel, someone [10:07] Wolverine Trading, and they make money. [10:09] So, here's how much money in dollars [10:12] they made and how much they got per [10:14] order type. And you can see it's a [10:15] non-substantial amount. When you trade, [10:17] you should assume that you're probably [10:19] going to lose some money due to [10:20] slippage, execution costs, etc. Really, [10:23] really understand how much money you're [10:25] making, why, and any can you escape the [10:29] fees that you might pay. And this will [10:31] give you a sense of how much companies [10:34] pay for your order. And it's utterly [10:37] crazy. And this is just a random [10:38] sampling. multiply that number by 12 or [10:40] when there's more volume, etc., what [10:41] that looks like. And so, be very [10:44] careful. People make money and the the [10:47] citadels of the world, etc. can make [10:49] even more money because they're willing [10:50] to buy people's trades. That's called [10:52] payment for flow. [10:55] Um, the next thing we want to talk about [10:56] is the efficient market. Uh, a lot of [10:58] people believe the market's very [11:00] efficient. It's impossible to beat the [11:02] market. Sometimes you can if you have [11:04] information before someone else, etc. uh [11:07] look at Warren Buffett, how he's done, [11:09] Citadel, how they print money, virtue, [11:11] how they've had basically no losing [11:13] days. Uh but they do it by having [11:15] information or trades that they can do [11:16] ahead of other people. Um but the [11:18] efficient market basically says, "Hey, [11:21] prices should reflect all information." [11:22] And there's different versions of this [11:24] efficient market hypothesis. If you're [11:26] trying to beat the market and trying to [11:28] trade, you should have information or [11:30] you should have data or a model that's [11:32] better and you can prove it out. [11:35] You should also know statistically what [11:37] you're doing. And so I'm not saying you [11:38] can just train a model and say, "Hey, [11:40] let's do it. Let's make a ton of money." [11:41] But some investors are simply lucky. Uh [11:44] you want to prove out with lots of [11:45] trades over time that what you're doing [11:46] makes money. And if you had a bunch of [11:48] monkeys throw darts or randomly [11:50] selected, some people were just going to [11:52] win. Uh that's a quote by uh a very [11:55] famous speculator who said that, "Hey, [11:57] look uh even though I blew up a bunch of [11:59] times, I was still right." He was 700 [12:01] standard deviations away. [12:04] Um, and the other thing I want to talk [12:06] about is AI and ML is only one very [12:09] small portion. This is a slide from [12:11] Google. The feature engineering, the [12:13] data collection, the analysis tools, the [12:18] logging, the reporting are all huge [12:19] sections of it. And if you want a job in [12:22] the industry, doing any one of these [12:23] components is extremely valuable. It's [12:26] not just the did I invent the coolest [12:28] algorithm, etc. Do the things that you [12:31] enjoy and are useful. You can alter the [12:33] AML, but it's usually a very small [12:34] portion of all the stuff that goes into [12:36] building systems that work. [12:40] Um, when you build a system, you should [12:42] back test it. How does it make money? [12:45] How much money is it making? And what's [12:47] your risk tolerance? Does it have draw [12:49] downs, etc. And is are you comfortable [12:51] with that? What happens if those draw [12:53] downs are bigger than you expected? [12:55] And the last thing I want to touch upon [12:57] is technical indicators. They are mostly [13:00] useless on their own. you should have a [13:01] reason why. Um, here's a list of a lot [13:04] of things that people look at. Uh, you [13:06] can use machine learning and technical [13:07] indicators and even that might not give [13:08] you something. Think about the smart [13:10] people that may be doing stuff there. [13:12] So, be realistic and really spend the [13:14] time looking into what you're doing. Um, [13:18] and so we'll we'll briefly cover a [13:19] couple of domain specific agents. We'll [13:21] pop up how much P&L we made and then we [13:24] will uh end it there. Yeah, I'll just go [13:26] through uh a few types of domain [13:28] specific agents that uh might apply to [13:31] finance different sorts of uh problems [13:33] that perhaps you could apply agents to. [13:35] So, one would be a classic trading [13:37] strategy is arbitrage where you buy in [13:39] one place and you sell on another. So, [13:41] say there's two different exchanges and [13:43] they're offering the the same uh the [13:45] same product like Bitcoin or something [13:47] like that. If it's cheaper in one place [13:49] to buy and more expensive in another [13:51] place to sell, you can buy it on the [13:53] cheap exchange, sell it on the expensive [13:55] exchange, transfer the asset over, and [13:57] then you've netted a profit. [14:01] Market is another traditional strategy [14:03] where you're trying to make money off [14:05] order flow. And so an example there is [14:08] there's an order book where on the left [14:09] side you can see people offering to buy [14:12] the stock for a certain price. And right [14:14] now the best bid on that stock is [14:16] $99.99. [14:18] And then there's people offering to sell [14:20] for a different price. Uh 100 being the [14:22] best price that people are offering to [14:23] sell at to market make that stock. What [14:26] you might do would be you'd [14:27] simultaneously offer to buy for $99.99 [14:31] and sell for 100. And your hope is that [14:34] people are going to come in with market [14:35] orders saying I just want to buy or I [14:37] just want to sell the stock at whatever [14:38] the market price is. And you hope they [14:40] come in approximately evenly. And so [14:43] somebody comes in saying they want to [14:45] sell their stock. You're the guy [14:47] offering it at $99.99. So they buy from [14:49] you. Somebody else comes in and they say [14:51] they want to buy the stock. You're the [14:52] guy offering to sell it for 100. So they [14:54] come to you and then you've profited one [14:56] penny. And so the idea is you do that [14:58] trade over and over and over again. And [15:01] uh you make money just off order flow. [15:05] Another model which we have written a [15:06] paper on is election prediction. So with [15:10] that one, we trained a DNN using various [15:12] features to predict the election and [15:14] then more interestingly the one-day [15:16] market impact after the election. I who [15:19] wins the election and then what's the [15:20] market response to that election result [15:23] the day after. And if you're interested [15:25] in that, we have a GitHub and a paper on [15:27] that. [15:29] Another strategy that we wrote a paper [15:31] on is using aerial imagery. So this one [15:34] is where we took overhead satellite [15:36] images of wheat fields and we took [15:39] historical price data of crop prices and [15:42] we combined them you the machine [15:45] learning model to predict the future [15:46] price of wheat and we expressed that as [15:48] a trade either going long expecting the [15:50] price to go up or going short expecting [15:52] the price to go down. You could come up [15:54] with many other sorts of uh problems you [15:57] might want to apply agents to, but maybe [16:00] these are a few to get the juices [16:02] flowing and come up with ideas. [16:06] So, thank you guys. Um, if we can pop [16:09] back to the demo real quick. [16:13] Uh, so we made $99. So, someone in the [16:16] raffle uh I'll I'll send you over an [16:19] Amazon gift card or something with uh [16:21] $99. It's a very small and again it's a [16:24] demo model. I really want to clarify. Um [16:29] this is the this is the model right [16:31] here. [16:35] So very small. It's meant to be a demo. [16:37] You can add much larger models etc. But [16:39] then when you want to do things in [16:40] production you have to think about [16:41] latencies etc. Um I think we're at time [16:44] so we're happy to take some questions [16:46] offline. Um and thank you all for [16:49] attending.