债券交易员担心人工智能是有道理的——人工智能确实会抢走他们的饭碗，但只会抢走一部分，而且不会一蹴而就。
Bond traders are right to be worried about artificial intelligence, it’s coming for their job — but only some of it, and not all at once.
汇丰控股有限公司企业及机构银行业务负责人迈克尔·罗伯茨表示，虽然人工智能代理可能在几年内就能在华尔街银行为债券交易，但经验丰富的人类仍需“参与其中”，以防止他们的数字同事失控。人们对市场动荡期间债券机器人表现不佳的记忆犹新。市场参与者表示，它们现在正在变得更好，与此同时，招聘经理们正在增聘精通人工智能的初级交易员和更多量化分析师来协助运行这些系统。
While AI agents could be trading bonds at Wall Street banks within a couple of years, experienced humans will still need to be “in the loop” to prevent their digital colleagues from going haywire, according to Michael Roberts, who heads corporate and institutional banking at HSBC Holdings Plc. Memories are still fresh about how bond bots fell short during periods of market unrest. They’re getting better now, and meanwhile hiring managers are adding AI-native junior traders and more quants to help run them, market participants said.
巴克莱银行主题固定收益、货币和大宗商品研究负责人佐尔尼察·托多罗娃在谈到人工智能机器人执行更多交易时表示：“根据我们与客户的所有对话，每个人都在为此做准备。这是一个缓慢发展的过程，但它肯定会发生。”对于美国公司债券市场而言，利害关系重大。根据克里西尔联盟格林威治公司的数据，美国公司债券市场在通常的一天里有价值620亿美元的证券换手。
“Based on all the conversations that we’ve had with clients, everybody’s preparing for this,” said Zornitsa Todorova, who leads thematic fixed-income currency and commodities research at Barclays, referring to AI bots executing more trades. “It’s a slow burner, but it’s definitely going to happen.”The stakes are huge for the US corporate bond market, where $62 billion of securities change hands on a typical day, according to data from Crisil Coalition Greenwich.
巴克莱银行最终估计，在未来几年里，这可能会推动美国投资级债券的电子交易量从目前的45%上升到65%，高收益债务交易量则会翻倍至50%。
Ultimately, Barclays estimates, this could push up electronic trading of US investment-grade bonds to 65% of volume in coming years from 45% currently, with high-yield debt trades doubling to 50%.
市场参与者表示，目前，早期采用者主要将人工智能代理用于信息搜集、分析数据或复杂的契约，以及提炼新闻头条，以帮助做出交易前决策。人工智能代理还可以帮助构建筛选和可视化工具，并实现内部任务自动化。
For now, early adopters are mostly using AI agents for information gathering, analyzing data or complex covenants and distilling news headlines to help make pre-trade calls, according to market participants. AI agents can also help build screening and visualization tools and for automating internal tasks.
Tradeweb市场公司是一家电子交易市场，在85个国家拥有3万多名客户。Tradeweb全球市场战略与解决方案负责人伊齐·康林表示，该公司正从寻求人工智能协助的客户那里获得“大量兴趣”。这包括交易前和交易后分析、帮助交易员比较不同协议的成本、为特定交易推荐交易商，或建议将哪些债券纳入组合交易。
Tradeweb Markets Inc., an electronic-trading marketplace that counts more than 3,000 clients in 85 countries, is getting “a ton of interest” from clients looking for AI assistance, said Izzy Conlin, head of strategy and solutions for global markets at Tradeweb. This includes pre-trade and post-trade analysis, helping traders compare costs across different protocols, recommending dealers for specific trades or suggesting bonds to include in a portfolio trade.
Tradeweb已经是美国公司债券的第一大电子交易平台，它正通过以约19.7亿美元收购排名第三的特鲁米德控股公司来押注更大的交易量。即便如此，康林在一次采访中表示：“我仍然无法想象没有人类交易员的世界。如果你试图非常迅速地改变交易商和投资者10年来习惯的做法，那是行不通的。”电子交易场所正通过将人工智能嵌入用户界面和工作流程来推动人工智能的采用。
Tradeweb, already the No. 1 e-trading platform for US corporate bonds, is betting on more volume by purchasing No. 3 Trumid Holdings for about $1.97 billion. Even so, “I still don’t see a world where you don’t have human traders,” Conlin said in an interview. “If you try and change what dealers and investors have done over 10 years really quickly, it doesn’t work.”Electronic trading venues are driving AI adoption by embedding it into the user interfaces and workflows.
今年6月，LTX在其BondGPT应用程序中增加了代理功能。该工具被称为“代理式BondGPT”，可以采取预定义的操作，例如创建交易凭证，甚至代表人类交易员发起交易。
In June, LTX added agentic capabilities to its BondGPT application. Dubbed “Agentic BondGPT,” the tool can take predefined actions like creating trade tickets or even launching trades on behalf of human traders.
人工智能代理还可以通过挑出价格差异或陷入困境且债务已经重组的公司的繁复文件中的隐藏价值，提醒人类注意信用市场晦涩角落里的机会。
An AI agent can also alert humans to opportunities in obscure corners of the credit market by sorting out price discrepancies or hidden value in the byzantine paperwork of distressed companies whose debts have been restructured.
目前，风险最高的美国公司债券（评级为CCC的债券）的平均利差自9月下旬以来一直超过1000个基点，这一水平通常意味着违约、重组或损失的高概率。信用投资者和交易员正在要求“代理式BondGPT”在例如某只CCC级债券的交易价格与其前一日收盘价相差超过50个基点时向他们发出警报。
Right now, the average spread on the riskiest US corporate bonds — notes rated CCC — has been more than 1,000 basis points since late September, a level that typically implies high probability of default, restructuring or loss. Credit investors and traders are asking Agentic BondGPT to, for example, alert them when a CCC bond trades more than 50 basis points wide from its prior day close.
其他提示还包括标记交易价格比前一天高或低一个点的垃圾债券，或者当出现针对评级至少为BBB且利差为85个基点或更宽的10年期工业债券、且报价达500万美元以上的出价时，自动填充交易凭证。
Other prompts include flagging junk bonds trading one point higher or lower than the previous day or populating a trade ticket when there’s an offer that’s $5 million-plus on a 10-year industrial bond rated at least BBB with a spread of 85 basis points or wider.
LTX首席执行官吉姆·科维亚特科夫斯基表示：“人工智能代理与在交易台聘请人类交易助理没有区别。”
“AI agents are no different to hiring a human trading assistant on the desk,” said LTX Chief Executive Officer Jim Kwiatkowski.
交易员们一致认为，此类人工智能工具可以极大地节省时间，使他们能够将几天的工作量缩短到几个小时完成。尽管如此，输出结果仍需要双重核查，而且在大多数情况下，人类并不信任人工智能来执行交易。交易员表示，在某些情况下，合规规则甚至完全禁止他们使用人工智能代理。
Traders agree that such AI tools are a huge time saver, allowing them to perform several days worth of tasks into a few hours. Still, the output needs to be double-checked, and humans for the most part aren’t trusting AI to execute trades. In some cases, traders say, compliance rules bar them from using AI agents altogether.
巴克莱银行指出，一个理论上的风险是，随着越来越多的人工智能代理日益对相同的数据和市场信号作出反应，它们可能会同时采取相同的行动并引发“阶段性波动”——换句话说，就是一场撼动市场广阔领域的痉挛。例如，如果多个人工智能模型都检测到某个发行人的信用状况恶化，它们可能会同时试图削减风险敞口，从而导致一波抛售潮，如果没有人类干预，这可能会放大价格走势。
One theoretical risk is that as more AI agents increasingly react to the same data and market signals, they could all act the same way simultaneously and trigger “episodic volatility,” according to Barclays — in other words, a convulsion that shakes a broad swath of the market. For example, if multiple AI models all detect worsening credit conditions for an issuer, they may all attempt to cut exposure at the same time, resulting in a wave of sell orders that could amplify price moves if humans don’t intervene.
詹姆斯·汉德森多部门和公司信用全球负责人约翰·劳埃德表示：“我们将其视为增强而非取代人类决策的工具。”
“We view it as a tool that augments human decision-making rather than replaces it,” said John Lloyd, global head of multisector and corporate credit at Janus Henderson.