AI 能理解困境债务吗？Harneet Kaur 的 iHeartMedia 案例分析

简要描述Brief Description

Harneet Kaur 探讨 AI 是否能够对陷入困境的信贷进行推理，并以 iHeartMedia 为例，比较模型在经过结构化推理流程前后的预测。文中将模型之间的一致意见与正确性的证据区分开来。Harneet Kaur examines whether AI can reason about distressed credit, using iHeartMedia to compare model predictions before and after a structured reasoning process. Agreement between models is distinguished from evidence of correctness.

目录Table of Contents

• 引言
  Intro
• 困境信贷的实际含义
  What distressed credit actually means
• 为什么旧版评分模型存在不足
  Why old scoring models fall short
• 理解信用意味着什么
  What it means to understand a credit
• AI 能预测接下来会发生什么吗？
  Can AI predict what happens next?
• 案例研究：iHeartMedia，以及两个模型意见不一
  Case study: iHeartMedia, and two models disagree
• 对推理的每个层次进行基准测试
  Benchmarking each layer of reasoning
• 结构化推理使模型收敛
  Structured reasoning makes the models converge
• 要点：收敛与正确性的区别
  Takeaways: convergence vs correctness

简介Intro

感谢大家来到这里。我非常期待进行这场演讲。先简单介绍一下我自己。我叫 Harneet，是 Axim Intelligence 的创始人兼首席执行官。在 Axim，我们正在为不良信贷投资打造智能层。thanks to everyone for being here. I'm so excited to be doing this talk. So, kind of quick intro on me. So, my name is Harneet. I'm the founder and CEO of Axim Intelligence. And in Axim, we're building the intelligence layer for distressed credit investing.

那么，这也是我们今天要讨论的内容。这场演讲的标题是：人工智能能理解困境信贷吗？So, also what we're going to be talking about today. and the title of this talk is can AI understand a distressed credit?

困境信贷的实际含义What distressed credit actually means

我想，如果你不在信用领域，大家心里的第一个问题就是：困境信用到底是什么意思？信用其实很简单。信用就是一家公司的债券或贷款。比如，可口可乐的债券就是一种信用资产。那么，陷入压力或困境又是什么意思呢？对吧？I think the first question on everyone's minds, if you're not you know, in the world of credit, is what does distressed credit really mean? credit is pretty simple. So, credit is just a bond or loan you know, of a corporation. The bond of Coca-Cola is a credit. now, what does it mean to be stressed or distressed? Right?

所以，如果公司存在任何潜在风险，可能导致其一是资金耗尽，二是无力偿还债务，那么就真的会出现压力。因此，每年信贷市场都会有一部分面临某种程度的压力或困境。这可能是由周期性问题造成的。比如说，建筑业市场低迷，房地产市场也低迷；也可能是由运营问题造成的。So, stress really happens if there is any sort of underlying risk in the company that would lead them to either one run out of money, two, not be able to pay back their debt. So, every portion every year some portion of the credit market faces some level of stress or distress. It can be caused by cyclical issues. So, let's say construction you know, the construction market is down, housing market is down. it can be caused by operational issues.

如果你经营一家餐饮企业，发生大规模食品召回时，也可能是由长期性的行业变革造成的。所以，如果你现在经营一家软件公司，可能正受到 AI 颠覆的影响，对吧？有些企业会面临挑战，然后很快恢复元气；有些企业面临挑战后，则会陷入困境。这里说的“困境”是什么意思？If you're a restaurant business, there's a big food recall. it can also be caused by secular disruption. So, if you're a software business today, and maybe seeing some AI disruption, right? some businesses face challenges and then come right back up. Some businesses face challenges and then get to a place of distress. What does distress mean?

其实这并不简单。破产是陷入困境后可能出现的一种结果。你也可以进行负债管理操作，这些术语我之后可以详细讲，但你只要知道这也是一种可能的情况。此外，还可以选择庭外重组或协商交换。所以，下一张幻灯片。It's actually not straightforward. Bankruptcy is one outcome of distress. You can also have a liability management exercise and these are terms I can get into them, but just know that's another thing that could happen and then or you can have an out-of-court restructuring negotiated exchange are a couple of the options. So, next slide.

为什么旧版评分模型存在不足Why old scoring models fall short

所以，正如你所看到的，这并没有那么简单，对吧？过去，判断财务困境的分析问题相当非此即彼：你可以预测一家公司会不会破产，也可以预测它会不会违约。如今，情况复杂得多了。因此，许多传统的机器学习方法已经跟不上了；而许多评分模型在十年、十五年甚至二十年前可能相当重要。最典型的就是奥特曼 Z-score：面对每月、每季度乃至每年都在变化的情况，以及透明度低得多的市场，它如今显得非常静态。So, as you saw, it's not that simple, right? Historically, the engineering question for distress was pretty binary. So, you can predict a bankruptcy either you predict it goes to a default or it doesn't. Today, that's become a lot more complex. so, a lot of these traditional machine learning methodologies just haven't kept up and a lot of the scoring that was pretty important maybe 10, 15, 20 years ago. the big one is the Altman Z-score has become now very static in terms of a situation that changes, you know, sometimes monthly, quarterly, yearly, and also is in a much more opaque market.

所以，如果你一直关注信贷市场，就会知道，私人信贷的兴起以及其他一些领域的发展，导致了定价缺失。另一方面，理解和分析困境信贷有一种非常定性的方法。作为困境信贷投资者，你可以分析财务状况，了解贷款人一方的情况，阅读信贷文件。So, if you've ever been following credit, the rise of private credit and some of these other areas have caused, you know, a lack of pricing. On the other hand, you have a very qualitative method of understanding and analyzing distressed credit, which is as a distressed credit investor, you can you know, analyze financials, understand what's going on in the lender side, read the credit docs.

基本上，你知道，就是普通的对冲基金投资者或私募股权投资者为了了解一笔信贷会做的所有事情。这个过程非常依赖人工，显然很难规模化，对吧？现在的问题是，你知道，创办 Axim 的部分原因，以及我思考过的一个问题是：AI 在理解一笔陷入困境的信贷，以及你可以采取的各种路径方面，究竟能做到什么程度？你能否将这个本质上相当复杂的流程系统化，对吧？Basically, you know, all the things a normal, you know, hedge fund investor or private equity investor would do to understand a credit. This process is super manual, very hard to scale, obviously. Right? Now, the question is, you know, part of starting Axim and a question I thought about is how far can AI get at understanding a distressed a distressed credit and the type of paths you can take. and can you systematize what is effectively a pretty complex process, right?

理解学分的含义What it means to understand a credit

那么，现在理解一笔困境债务意味着什么呢？对吧？作为投资者，我们要设法回答很多问题。比如从财务角度会考虑哪些事情，对吧？比如，流动性状况如何？资本结构是什么样的？So, now what does it mean to understand a distressed credit, right? Like as an investor, we have a lot of questions we're trying to answer. the things that you think about from a financial side, right? What is you know, what is liquidity? What is the capital structure?

然后你还会遇到一些法律方面的问题。比如说，债务在结构上的优先级如何？对吧？你们的优先级是更高还是更低？然后还有运营方面的问题，比如说，你们如何理解这门生意？And then you have questions from the legal side. How, you know, what's the structural priority of the debt? Right? Like do you have higher priority or lower priority? And then you have operational issues, like okay, how do you understand the business?

营收是多少？根据你对业务基本面的了解，也就是基本面分析，明天的营收会增加还是减少？对吧？目前我们在市场上看到的情况是，如果你关注 Rogo、Harvey 或其他一些大型 AI 公司，就会发现它们所做的，确实是对各种不同的问题进行基准测试，对吧？What are the revenues? Are the revenues going to go up or down tomorrow based on fundamental you know, fundamental analysis of the business. Right? Right now, what we've seen in the market, you know, if you're following Rogo or Harvey or some of these, you know, big AI companies, what they've done is really think about benchmarking all these different questions, right?

比如，AI 在回答“债务中哪一部分在结构上具有优先受偿权”这个问题时表现如何？然后再进行严格的基准测试，并把它——你知道——称作“大金融基准”。Like how good is AI answering what part of the debt is structurally senior? and then kind of do rigorous benchmarking and make that you know, call it the big finance bench.

AI 能预测接下来会发生什么吗？Can AI predict what happens next?

我想换个角度稍微思考一下这个问题，对吧？因为在我看来，真正理解信贷，尤其是困境信贷，关键在于能不能预测接下来会发生什么，对吧？我能不能弄清楚它未来可能走哪条路径，并得出相应的概率分布？I wanted to think about it slightly differently, right? Because the way I thought about it is the true understanding, right, of credit, and particularly distressed credit, is can you predict what happens next? Right? Can I figure out and have a probability distribution of what path this takes in the future?

案例研究：iHeartMedia，以及两个模型意见不一Case study: iHeartMedia, and two models disagree

在这项练习中，我选择了 iHeartMedia。这家公司于 2018 年申请破产保护，也经历过一番折腾。2024 年，它又进行了一笔困境交易。所以，这是一个正在发生的案例，就发生在今天。For this exercise, I chose iHeartMedia, which is a company that filed for a bankruptcy in 2018. So, it's you know, it's been through the rigmarole. it also went and did another stressed transaction back in 2024. So, this is a live situation. This is happening today.

没人能真正说清楚未来会怎样，对吧？也不存在什么标准答案。但我做的就是把这个问题丢给最聪明的那些模型。我用的是 Max Intelligence 上的 Astra 6，还有 Opus 5.5。一个特别有意思的结论是，它们的看法并不一致。No one has a real, you know, answer for what's the future, right? there's no ground truth. But what I did was let's throw it into the smartest, you know, models. I did Astra 6 on Max Intelligence and Opus 5.5. And what was a really interesting kind of conclusion is that they disagreed.

也就是说，在给定数据结构或背景的情况下，他们对未来各种情况发生的概率看法不同，对吧？所以，如果你看这里的协商重组，Astra认为这是破产时的最佳选择，Opus也认为这是最佳选择，对吧？Like they disagreed on what the probabilities of the future would be given a structure or context of the data, right? And so, if you look here on negotiated restructuring, Astra thought that was the best option for a bankruptcy, Opus thought that was the best option, right?

那么，你怎么判断哪个对、哪个错呢？你没法判断，因为我不知道概率是多少，对吧？我不知道接下来会发生什么。我不知道会不会破产，也不知道会不会进行庭外重组。Now, how do you say which is right or wrong? You can't cuz I don't know what the probabilities are, right? I don't know what's going to happen. I don't know if it's going to be a bankruptcy. I don't know if it's going to be a out of court restructuring.

我不知道这会不会是一次再融资。我会尽量作出最好的判断。我会说我倾向于 Astra 6，但你知道，我们都是人。这是个挺有意思的结论，对吧？就像，等等，他们意见不一致。I don't know if it's going to be a refinancing. I have my best judgment on things. I would say I lean towards Astra 6, but you know, we're all human. And it's a kind of an interesting conclusion, right? Like, okay, wait they disagree.

对推理的每一层进行基准测试Benchmarking each layer of reasoning

我想到的第二件事是：好吧，我该如何至少改进对他们流程的评估，对吧？这样，我就可以看看自己是否同意这些不同的假设。基本上，你们已经建立了这样一种基准，也就是对每一层进行基准测试。所以，显然，LLM 在每一层都可能出错，对吧？就信息提取这一层而言，我会说，对于上市公司来说，这个问题基本上已经解决了。The second thing I thought about is Okay, how do I improve at least an assessment of their process, right? So, I can look at whether I agree or disagree with the different assumptions. and so, basically you've created this sort of benchmark where you know, we benchmark each layer. So, obviously these are at each layer the LLM could mess up, right? On the extraction layer, I would say with public companies it's pretty it's pretty solved.

我想说，面对私人贷款，以及 Excel 文档和 PDF，这仍然可能是个挑战，对吧？接下来还有对账或推导这一层。抽取就是拿到 EBITDA。推导则是：好，现在你有了 EBITDA 和债务，就可以算出杠杆率。然后还需要做出判断。I would say with private loans and when you have like Excel documents and PDFs, that can that can still be a challenge, right? Then you have the reconciliation or derivation layer. So, extraction would be you got EBITDA. derivation would be, okay, now you have EBITDA and you have debt, now you can have leverage. and then you have a judgment.

好了，现在你就有了判断的筹码。这是一笔困境信贷吗？好。所以，你已经有了判断层，对吧？然后，你就可以开始了。好，如果你能评估每一层，对吧？到了预测层时，你已经确认数据点没问题、核对层没问题，也确认有某种合理的判断，对吧？Okay, now you have the leverage. Is this a distressed credit or not? Okay. So, So, you got the judgment layer, right? And then, then you can start, okay, if you can assess each layer, right? Once you get to the prediction layer, you've assured the data points are good, you've assured that the reconciliation layer is good, you've assured that there's some sort of reasonable judgment, right?

然后，一旦你确认了这些事情，就可以进入预测层了。预测层其实就是你作为一名人类分析师，思考所有这些步骤的方式，对吧？如果你的判断是建立在一个糟糕的数据点上的，那就不会是个好判断，事情就是这样，对吧？And then, once you've assured those things, then you can get to the prediction layer, which is really how, you know, as a human analyst, you think about all of these steps, right? If you build a judgment on a bad data point, that's not going to be a good judgment, and that's that's just the way this works, right?

结构化推理使模型趋于收敛Structured reasoning makes the models converge

好，哎呀，开始吧。我们发现的情况其实是——这也许并不令人意外，但我觉得挺有意思——一旦我们真正建立起结构化的推理流程，两种智能，也就是这两个模型，对于业务未来可能性的判断实际上就趋于一致了，对吧？所以，现在不再是 Opus 认为破产概率为 62%、Astra 认为是 25%，而是两个模型都认为破产概率实际上只有四分之一，也就是 25%；并且两个模型都同意，对这家企业来说，协商重组是更有可能的路径。So, Oop. Here we go. So, basically what we found, and maybe this is not surprising, but thought it was interesting, is that once we actually created the structured reasoning process, you actually had both the intelligence, both the models actually converge in terms of what they believed the future probabilities of the business are going to be, right? So, now instead of saying 62% bankruptcy with Opus and 25% bankruptcy with Astra, now both these models are saying there's actually only a quarter or 25% chance of bankruptcy and both models agreed that the negotiated restructuring path is actually the more likely path of the business.

要点：收敛性与正确性Takeaways: convergence vs correctness

这里最重要的结论是什么？首先，我会说，面对不可验证的问题，如果给模型提供和人类相同的结构化、组织好的上下文，它们有预测优势吗？你必须说没有，对吧？What are the big takeaways here? one I would say the first takeaway is one given non-verifiable questions, right? With the same structured context, organized context as a human. I mean, do these models have a predictive edge? And you have to say no, right?

因为它们实际上如此，即便智能体也是人，对吧？在某种意义上，我就不深入讨论这里的争议了。第二，结构化的推理过程可以让两个模型，或者说，不同的智能体，最终对未来形成基本相同的看法，并对一家企业及其可能的发展路径形成相似的理解。但那真的是正确的路径吗？我们不知道，对吧？而且，只有在最真实的衡量标准下，这才成立，对吧？否则，任何形式的历史分析都会造成某种前视偏差。Because they are in effect, even agents are human, right? In in some sense of the word and I won't get into the controversy there. the second is but a structured reasoning process can get the two models to or, you know, different intelligences to converge on effectively a same view of the future and effectively a similar understanding of a business and the path it could take. but is that the correct path? We don't know, right? and is this is really only at the truest measure, right? Cuz otherwise any sort of historical analysis would create some sort of look ahead bias.

但我觉得，真正有意思的一点是，我一直在想，这种收敛也相当有意思，对吧？因为如果你能系统地构建一个过程，让它最终收敛到相同的概率结果，那实际上就是市场中的价格。所以我觉得这是一个挺有意思的启示。这就是最后一点。然后，我再简单介绍一下自己。but I think the really interesting thing that I've been thinking about is the convergence is pretty interesting, too, right? Because if you can systematically build a process that actually converges on the same probabilistic outcomes. That is effectively what is a price in a market. and so I I thought that was an interesting takeaway. so that's the last thing. And then just a quick kind of spiel on me.

那么，正如我刚才说的，我叫 Harneet，正在打造 Axim。简单介绍一下我的背景：我在金融行业工作了 10 年，最近几年在一家专注于困境投资的私募股权基金／对冲基金担任困境信贷和特殊机会投资人。如果你对 Axim 感兴趣，这是二维码。我要说的就这些，谢谢。So, my name's, like I said, Harneet. building Axim. kind of background on me. I spent 10 years in finance. the last few as a distressed credit and special situations investor at a distressed PE / hedge fund. and then this is the QR code if you're interested in Axim. and that's all I had. Thank you.