朱塞佩·帕莱奥洛戈 - 多管理人对冲基金与深入思考简单事物（第七季第11集）

大家好，我是Corey Hofstein。如果你一直在收听《Flirting with Models》，就知道我专注于重新思考投资组合构建，我想亲自邀请你参加一个正在做这件事的活动。10月8日，我们将在芝加哥Cboe Global Markets举办回报叠加研讨会。这是一场为期一天的线下深度探讨资本效率策略的活动，我们邀请了达美航空首席投资官Jonathan Glidden、One River的Patrick Casley，以及加拿大养老金计划系统策略组董事总经理Mark Horbul等演讲嘉宾。

Hey, everyone. Corey Hofstein here. If you've been tuning in to Flirting with Models, you know I'm all about rethinking portfolio construction, and I wanna personally invite you to an event that's doing just that. On October 8, we're hosting the return stacking symposium at Cboe Global Markets in Chicago. It's a one day in person deep dive into capital efficient strategies, and we're featuring speakers like Jonathan Glidden, CIO of Delta Airlines, Patrick Casley from One River, and Mark Horbul, managing director of the systematic strategies group at Canada Pension Plan.

这是你直接听取领导便携阿尔法和回报叠加策略的机构配置者见解的机会。但名额有限，请前往returnstacked.com/symposium了解更多信息并注册。期待在那里见到你。

This is your chance to hear directly from the institutional allocators leading the charge on portable alpha and return stacking. But space is limited, so head over to returnstacked.com/symposium to learn more and register. Hope to see you there.

三，

Three,

二，一。让我们开始吧。大家好，欢迎收听。我是Corey Hofstein，这里是《Flirting with Models》，这档播客节目揭开量化策略背后的人为因素。Corey Hofstein是Newfound Research的联合创始人兼首席投资官。

two, 1. Let's jam. Hello, and welcome, everyone. I'm Corey Hofstein, and this is flirting with models, the podcast that pulls back the curtain to discover the human factor behind the quantitative strategy. Corey Hofstein is the cofounder and chief investment officer of Newfound Research.

由于行业法规，他不会在播客中讨论任何Newfound Research的基金。所有播客参与者的观点仅代表其个人意见，并不反映Newfound Research的观点。本播客仅供信息参考，不应作为投资决策的依据。Newfound Research的客户可能持有本播客讨论的证券头寸。更多信息，请访问thinknewfound.com。

Due to industry regulations, he will not discuss any of Newfound Research's funds on this podcast. All opinions expressed by podcast participants are solely their own opinion and do not reflect the opinion of Newfound Research. This podcast is for informational purposes only and should not be relied upon as a basis for investment decisions. Clients of Newfound Research may maintain positions and securities discussed in this podcast. For more information, visit thinknewfound.com.

在本期节目中，我与Giuseppe Paliologo（他喜欢被叫做GAPI）进行了对话。目前正处于花园休假期的GAPI，此前曾在Citadel从事风险与量化分析工作，担任Millennium的企业风险主管，最近则担任HRT的风险管理主管。我们首先讨论了在多经理对冲基金中，量化研究员实际做什么。作为对基本面PM的半支持角色，Gappy解释了投资经理覆盖、因子对冲和内部阿尔法捕获如何协同工作，以帮助最大化公司盈亏。然后我们讨论了因子研究和投资组合构建的广阔领域，Gappy分享了他的一些坚定观点，包括因子应如何构建以及如何运用。

In this episode, I speak with Giuseppe Paliologo or GAPI as he likes to be called. Currently on garden leave, GAPI has previously worked in risk and quantitative analytics at Citadel as head of enterprise risk at Millennium and most recently as head of risk management at HRT. We begin the conversation with the discussion as to what a quant researcher actually does at a multi manager hedge fund. As a semi support role to fundamental PMs, Gappy explains how portfolio manager coverage factor hedging and internal alpha capture can all work together to help maximize firm P and L. We then discuss the broad field of factor research and portfolio construction where Gappy shares some of his strongly held views both on how factors should be constructed as well as how they should be utilized.

相关话题包括回报与特征、混合与整合阿尔法信号、单期与多期优化，以及线性与非线性模型。请享受我与Giuseppe Paliologo的对话。Gappy，欢迎来到《Flirting with Models》。能在你公司间花园休假期邀请到你真是荣幸，你曾在一些真正的巨头公司工作，为这期播客带来了非常宝贵的经验。今天能有你做客，我无比高兴。

Topics here include returns versus characteristics, mixing versus integrating alpha signals, single versus multi period optimization and linear versus nonlinear models. Please enjoy my conversation with Giuseppe Paliologo. Gappy, welcome to Flirting with Models. This is a real privilege to catch you while you are on garden leave between firms and you've worked at some real mega names and bring some really incredible experience to this podcast. So could not be more delighted to have you today.

非常感谢。

Thank you so much.

嗯，谢谢你，科里。我只是想说，我听你的播客已经很多很多年了。所以感觉我已经认识你了。

Well, thanks to you, Corey. I just wanted to say that I have been listening to your podcast for years and years. So I feel like I already know you.

我很感激。那么让我们开始吧，先为那些可能没有关注你在推特上迅速成名的人铺垫一下背景，这样大家就能了解你的经验以及你在这里能带来什么。

I appreciate that. So let's get started with just setting the table for folks who maybe haven't been following your meteoric rise to fame on Twitter at least with a bit of your background so people can know what your experience is and what you're bringing to the table here.

我的名字是朱塞佩·帕利奥洛戈。我的昵称是GAPPY，g a p p y。我在这个行业已经工作了大约十四年。在此之前，我是IBM Research在约克镇高地的一名应用数学家。后来我转行进入了金融领域。

My name is Giuseppe Paliologo. I go by the nickname GAPPY, g a p p y. I have been in the industry for about fourteen years. Before then, I was an applied mathematician working for IBM Research in Yorktown Heights. I transitioned into finance.

首先，我去了Axioma，这是一家因子模型和优化服务提供商。后来我被一位猎头联系，为Citadel工作。所以我被Citadel聘用。我在那里工作过两次。我还在Millennium担任过企业风险主管，在HRT担任首任风险主管。

First, I moved to Axioma, which is a factor model and optimization provider. I got called by a recruiter for Citadel. So I got hired by Citadel. I worked there twice. I worked at Millennium as head of enterprise risk, HRT as the first head of risk.

在这期间，我还在一家相对较小的对冲基金担任量化投资组合经理。我于11月离开了HRT，并将在11月加入Aliasny资产管理公司。这是公开信息。我想说明的是，我所说的一切当然与我过去和未来的雇主无关。我写了一本书，叫做《高级投资组合管理》。

In between, I was also a quantitative PM for a relatively small hedge fund. I did leave HRT in November and I will join Aliasny Asset Management in November. This is public information. And I just want to mention that of course everything that I say has nothing to do with my previous and future employers. I wrote a book called Advanced Portfolio Management.

关于这本书，我其实有个有趣的故事。《高级投资组合管理》实际上就是简单的投资组合管理。人们被这个标题吸引，高级投资组合经理。而我正在写第二本书，叫做《量化投资的要素》，这才是真正的高级投资组合管理。所以它是为工具和模型的制造者以及投资组合优化者准备的。

I have actually a funny story about this. So advanced portfolio management is actually simple portfolio management. People get caught into this title, advanced portfolio manager. And I am writing a second book called The Elements of Quantitative Investing, which is instead truly advanced portfolio management. So it's meant for the makers of tools and models and portfolio optimization.

关于《高级投资组合管理》这本书，我确实是以为基本面投资组合经理为核心撰写的。这是一本专门献给他们、旨在帮助他们日常工作的书。

Whereas advanced portfolio management, I really wrote with fundamental portfolio managers in mind. It's a book dedicated to them aimed at helping them in their everyday life.

有点像冰岛和格陵兰岛的名字混淆问题——两本书的名字正好反了。读完你的第一本书以及第二本书的草稿后，我实在无法更强烈地推荐它们。正如你所说，《高级投资组合管理》并不是面向最顶尖的量化专家的，它真正服务的是基本面投资组合经理，但书中涵盖了一些精彩的基础知识，而且你的新书也毫不回避数学内容。

Sort of like a Iceland Greenland problem. You got the names backwards on the books. Having read both your first book as well as a draft of your second book, I cannot recommend them enough. To your point, advanced portfolio management is not going to be for the most advanced quant. It truly is for the fundamental portfolio manager, but it goes over some wonderful basics and your new book does not shy away from the math.

我认为量化从业者也会非常喜欢它——至少从我读到的草稿来看是这样，而且我相信现在肯定已经完善了很多。我期待能买到一本，也许你还能给我寄一本签名版放在我的办公桌上。

And I think quant practitioners will really enjoy it from at least the draft I read and I'm sure it's been far improved from there. I look forward to picking up a copy and maybe I can get you to send me a signed copy for my desk.

当然可以。

Of course.

你提到了自己丰富的经验，其实第一个问题就有人在我宣布邀请你上播客后联系我——一位推特网友说，他最想问的是（我觉得这是个很棒的问题）：你在那么多人们梦寐以求的公司工作过，像Citadel、Millennium和HRT这些公司之间有哪些独特之处？

You mentioned your breadth of experience and one of the first questions, actually someone reached out to me after I mentioned I was having you on the podcast, someone from Twitter said, well, the first thing I would want to ask, I think this was a great question, was you have worked at so many different firms that people aspire to work at. What are some of the idiosyncrasies between your citadels and your millenniums and your HRTs?

我引用其中一家公司的联合负责人（就不提名字了）曾经告诉我的话：Citadel像新加坡，中央集权管理，效率极高，一切井井有条；Millennium则像美国，因为它是联邦制的。

I'll actually quote a co head of one of these firms without mentioning his name, who once told me, Citadel is Singapore. It's centrally managed. It's very efficiently managed. Everything is clean and working. Millennium is The United States because it's federated.

如果你想说得直白点，也可以说是各自为政。它有点混乱，规模庞大，但两种模式都能成功运作。所以我会沿用这个比喻，并补充说HRT有点像瓦坎达。

If you want, you could say siloed. It's messy. It's big. But things work out in both models. And so I would take that metaphor and I would add that HRT is a little bit like Wakanda.

所以HRT规模较小，某种意义上更酷。我认为它的人力资本、人员素质可能是我工作过的公司中最高的。而且他们的做事方式与其他两家公司截然不同。

So HRT is smaller, in a sense it's cooler. The human capital, the quality of people I think is probably the highest among the firms that I've worked at. And they do things in very different ways than the other two.

我们会避免让你陷入太多麻烦，也不提及其他事情。我很喜欢那个比喻。那么，我们来谈谈你担任过的角色。我认为当我们使用'量化'这个词时，它的含义范围非常广泛。每当有年轻人来找我说'我对成为量化分析师感兴趣'时，我都会说'你得具体告诉我这对你意味着什么'，因为买方股票量化与卖方衍生品量化非常不同，高频量化又完全是另一个领域。

We'll keep you from getting into too much trouble and mention anything else. I love the metaphor there. Well, let's talk about the roles that you've served in. I think when we use this word quant, there's such a big breadth of what it can mean to be a quant. And whenever I have someone young approach me and say, I'm interested in being a quant, I'm like, you're gonna have to tell me more what that means to you because buy side equity quant is very different than sell side derivatives quant is very different than the high frequency quant.

你在这些公司担任过所谓的'量化研究员'角色。能否从最高层面描述一下这个角色到底是什么？它包含哪些内容？具体职责是什么？

You've served in this role of what's called quote a quant researcher at these firms. And can you maybe describe at the highest level what that role actually is? What does it entail? What are the responsibilities?

作为介绍，我先后担任过两个角色。我在Citadel对冲基金两次担任量化研究员（QR），然后在Millennium的HRT担任风险岗位，现在即将回到BAM担任量化研究员。

As a way of introduction, so I have served alternatively in two roles. I was QR in Citadel twice in the hedge fund. Then I was risk in Millennium at HRT. I'm going back to QR in BAM. Okay.

所以我算是在这两个角色间切换。这两个角色有相似之处，但也有根本区别。总的来说，量化研究员通常是买方提供的量化服务。如果是卖方，你会称之为Stratz或Desquant等，那是不同的角色。在买方，量化研究员通常提供中心化服务，为公司做中心化服务，包括对冲、内部阿尔法捕获（这个我们稍后可以讨论）、因子模型和定制因子，以及投资组合管理咨询或覆盖服务，随便怎么称呼。

So I kind of oscillate between the two. The two roles have similarities, but also fundamental differences. In general, I would say QR is meant to be in general quantitative services on the buy side. Otherwise, you would say somebody's Stratz or a Desquant or something like this, but it's a different role on the sell side. On the buy side, typically a QR person serves central services, does central services for the firm and those include hedging, internal alpha capture, which we can discuss a little bit later, factor models and custom factors, and then portfolio management advisory or coverage or engagement, call it whatever.

但基本上就是帮助投资组合经理完成日常工作，理解他们的业绩表现并将他们的想法变现。简而言之，我认为这就是量化研究员的职责。有时还附带执行服务、核心流动性账簿角色，为业务主管或对冲基金负责人提供风险分配给业务内各团队的建议。通常，将风险分配给整个业务的职能属于首席风险官或其办公室。总的来说，量化研究员背后的激励原则是通过可重复的方法帮助公司更好地变现想法，从而最大化公司的损益。

But you basically help the portfolio managers in their daily job of understanding their performance and monetizing their ideas. That is in a nutshell, I think what QR is. There is also sometimes attached to this, execution services, essential liquidity book role, advising the business heads or the head of the hedge fund in allocating risk to individual teams within a business. Typically the role of the function of allocating risk to the entire business belongs to the CRO or the office of the CRO. In general, the inspiring principle behind QR is that QR is meant to maximize the P and L of the firm by helping the firm monetize, better their ideas through repeatable methods.

同时也旨在帮助投资组合经理提升。所以这是一种咨询角色，让他们变得更好。而风险岗位则相反。量化研究员就像足球队中的进攻职能，而风险岗位通常就像运动队中的防守职能。

And also it's meant to help portfolio managers improve. So it's a sort of advisory role to make them better. Risk on the other side. So QR is like the attack function in a football team. And risk is like the defense function in a sport team in general.

所以，首先你要确保合伙人的风险偏好与公司实际承担的风险相一致，同时还要详细了解各个策略的具体运作，防止公司陷入最坏情况，或者至少让最坏情况变得可承受，并开发衡量风险的方法等等。因此，风险管理也非常、非常依赖量化。在我看来，某种意义上，风险管理是整个公司中最具智慧含量的角色。风险管理这个叫法其实不太准确，它应该被称为关于公司投资的理性决策。

So you want to make sure that the risk appetite, first of all, of the partners is aligned with the risk that the firm is actually taking, but also understands in detail what individual strategies do and prevents the firm from incurring into the worst case, or at least makes the worst case tolerable, and develops the methods to measure risk and so on. So risk management is also very, very heavily quantitative. In a sense, risk management is the most intellectual role in the whole firm in my view. Like risk management is a misnomer. Risk management should be called rational decision making about investment in a firm.

首席风险官确实是一个非常重要的人物，尤其是在买方。卖方和买方在这些职能上存在根本性差异。但我想说，一个健康良好的对冲基金拥有一个有效且有力的风险管理职能。这是针对对冲基金而言。至于自营交易公司方面，我的经验没有那么丰富，我唯一的经验来自HRT。

The CRO is truly a very important person, especially on the buy side. So the sell side and the buy side have fundamental differences with respect to these functions. But I would say this, a good healthy hedge fund has a good and healthy risk management function that matters, that has bite. This is for hedge funds. Now I do not have as an extensive experience on the prop trading firm side, so my only experience is with HRT.

我想再次强调，虽然不深入探讨HRT的细节以及这个地方有多么出色和独特，但在HRT，量化研究和量化态度真正渗透到整个公司。我当时在风险部门，但那里的每个人在掌握所有基础知识方面都远超平均水平。所以那里的情况有所不同。

I would say that again without getting into details of HRT and how great and unique this place is, but in HRT I think quantitative research and a quantitative attitude is really underlying the whole firm. So I was in risk there, but everybody is way above average literate in terms of knowing all the basics. So it's different there.

我想对听众说，你们一定要系好安全带。这可能会成为我录制过的最长的一期节目，因为你刚才谈到的所有内容，我们都将在本期节目中深入探讨，包括那些不同的角色和职能。这期节目非常令人兴奋。我想从投资组合经理覆盖这个概念开始，因为在我与其他量化人员交流时，发现这个话题并不常见。也许这更像是一个内部秘密，当人们想到典型的量化研究员时，这个角色并不常被提及。

I wanna say for listeners, you should definitely buckle up. This might end up being the longest episode I've ever recorded because everything that you just went into, we're gonna explore in this episode, all those different roles and functions. This is an exciting one. And I wanna start with this idea of portfolio manager coverage because this isn't one that I find comes up a lot when I'm talking to other quants. Maybe this is more an internally held secret and it's not a role I hear a lot about when people think about a typical quant researcher.

那么，你能分享一下当你谈到投资组合经理覆盖这个概念时，具体的工作范围是什么吗？

So can you share what the scope of work is when you talk about that portfolio manager coverage concept?

我认为'覆盖'并不是一个通用术语。我相信它最初至少是Citadel特有的命名方式。但一般来说，你也可以称之为投资组合管理咨询或量化支持等等。PM覆盖人员是做什么的？他们协助完成QR的两项核心任务。

I think that coverage is not a universal term. I believe that it's mostly originally at least a Citadel specific nomenclature. But in general you could call it also a portfolio management advisory or quantitative support, whatever. What does a PM coverage person do? They do help on the two mandates of QR.

即货币化方面和咨询或分析方面的工作。我们来稍微列举一下具体任务。首先有一个非常基础的事情，就是在大多数大型公司中——不仅限于平台或pod shop，我认为目前任何多经理对冲基金（不一定非得是平台，可能是5个PM的地方，甚至只有1个PM的地方）——都存在一个量化框架。但总的来说，大多数市场中性对冲基金都采用风险模型、因子模型来进行绩效归因、风险分解等工作。

So the monetize side of things and the advice or analysis side of things. Let's list a little bit the tasks. First of all there is a very basic thing which is there is a quantitative framework in most large firms, not only platforms or pod shops, but I would say at this point in any multi manager hedge fund, which is not necessarily a platform, it could be a 5PM place. It could be even a 1PM place. But in general, most market neutral hedge funds employ risk models, factor models to performance attribution, risk decomposition and the like.

所以这已经成为一种常态。但它伴随着包袱，并不简单，尤其当投资组合经理来自不同背景时。想象一下，一位非常擅长选股的基金经理曾在机构资产管理公司或家族办公室做多头策略。他们来到当铺（指对冲基金），会经历文化冲击。因为虽然他们能理性思考主动头寸，但在纯多头组合背景下操作与现在整个组合市场中性、主动头寸规模至少是原来10倍的情况截然不同。

So this has become a sort of normal thing. And it's coming with baggage, it's not straightforward, especially when a portfolio manager is coming from a different experience. So imagine you have a very talented stock picker who worked for an institutional asset manager or for a family office long only. They come into a pawnshop and it's a culture shock. Because yes, they rationally can think about their active positions, but it's a very different thing to do it in the context of a long only portfolio versus well now my entire portfolio is market neutral and my active positions are 10 times at minimum what they used to be.

这会彻底扰乱你的思维，你不知道该怎么做，不知道应该交易多少。因此解释模型和培训基金经理可能是最基础的任务。然后从这里开始，你会进行某种高级风险分析，向基金经理说明他们在哪些地方承担风险。你还会运行绩效归因，除了标准绩效归因外，还可以做更高级的绩效归因分析。

It really messes up with your head and you don't know what to do, you don't know how much you should trade. So explaining the model and training the PMs is maybe the zeroth order of magnitude task. Then from here what you do is you do some type of advanced risk analysis. So you can explain to a PM where are they taking their risks. You run also a performance attribution, there is something more than the standard performance attribution that you can do, you can do more advanced performance attribution.

你还会开发这些方法，可能是与量化研究团队中不直接面对基金经理的其他部门合作。然后你解决定制化问题。具体是什么问题我不确定。一般来说，比如你有因子模型。因子模型通常在国家级别上有明确定义和优化。

You develop also these methods, jointly maybe with other parts of the QR team that are not PM facing. And then you solve bespoke problems. What can be these problems, I don't know. In general for example, you have factor models. Factor models are typically well defined and well tuned at the country level.

也有覆盖美国和欧洲或全球的全球因子模型。但每次提升一个层级，你都会损失一些分辨率。这就变成了使用美国模型还是全球模型的权衡取舍，有得必有失。如果有人只覆盖美国市场，那没问题。

There are also global factor models that cover US and Europe or the entire world. But every time that you go up a level, you're going to lose something in the resolution power. It becomes a trade off whether to use US versus global. You gain and you lose something. If somebody has only US, for example coverage, that's okay.

但如果他们同时覆盖美国和欧洲，就会出现例如市场暴露相关的问题，对吧？什么是市场暴露？如何管理？所以有各种各样的问题。哦，模型中的因子载荷不正确，你需要解决这个问题。

But if they have US and Europe then there are problems for example related to market exposures, right? What is market exposure? How you manage that? So there are all sorts of questions. Oh the loadings in a model are not correct, so you want to address that.

当然还有出现大幅回撤的情况。这时要去找出回撤的真正根源。通常你无法用因子模型解释大部分回撤。所以大多数回撤原因在于我们所谓的特异损益（idiosyncratic P&L），这不是系统性损益，也不是市场范围的损益，而是一种更高阶的系统性原因。

And then of course there are cases when there are large drawdowns. So go and find out what the source of the drawdown really is. Typically you cannot explain a lot of the drawdowns in a factor model. So most of the drawdown causes are in what we call idiosyncratic P and L, which is not systematic P and L, it's not market wide P and L. It's kind of a higher order, let's call it higher order systematic causes.

因此需要付出这些努力。然后你会从基金经理那里听到很多问题，保持接地气的沟通。覆盖人员会把这些问题反馈给做因子风险模型的研究人员，或者做内部阿尔法捕获的团队等等，让他们来解决。所以这是非常重要的工作。从某种意义上说，量化研究中的所有工作都是覆盖工作的下游。

And so there is that effort. And then there are lots of problems that you hear from a PM, you have your ear on the ground. And then the coverage person will feed these problems back to the researchers that do factor risk models, or they do internal alpha capture or whatever so that they solve that. So it's a very important thing. In a sense, everything in QR is downstream of coverage.

如果没有覆盖，纯研究和损益（P&L）生产者之间就真的没有任何联系。因此，做好这一点非常非常重要。成为好的倾听者、好的沟通者，具备扎实的量化训练，拥有严谨的研究流程至关重要，因为你实际上一直在进行非常短期的研究项目，你既不想欺骗自己，也不想欺骗投资组合经理。这其中有着非常强的受托责任。在某种程度上，你是投资组合经理最好的朋友。

Like if there is no coverage, there is really no link between pure research and the P and L producers. So it's very, very important to do it well. It's very important to be good listeners, good communicators, have like a sound quantitative training, to have a disciplined research process because you do in a way very short term research projects all the time and you don't want to fool yourself or the portfolio managers. There is a very strong fiduciary aspect to this. In a way you are the portfolio manager's best friend.

这从某种意义上说就是覆盖的意义。

That is in a sense what coverage is.

你提到的一个非常有趣的数据集是许多这类公司都拥有的历史交易信息，既包括你所覆盖的PM（投资组合经理）的信息，也包括跨时间和跨PM的横截面数据。你在我们预通话中提到，你可以利用这些数据来帮助理解PM的行为，并敦促他们有时承担更多风险，有时减少风险，帮助他们更好地理解自身行为，从而成为更敏锐的交易者。我希望你或许能详细说明一下，这些数据支持了哪些有助于提升PM绩效的有用研究。

One of the really interesting data sets that you mentioned that a lot of these firms have is historical trading information, both about the PM that you're covering, but going back over time and cross sectionally across PMs. You mentioned on our pre call that you can use this data to help understand the behavior of a PM and urge them to take more risk at times, less risk, help them understand their own behavior better to become sharper traders. I was hoping you maybe you could expand on what sort of useful studies that data allows in support of PM performance.

你让我想到了一个重要的竞争优势，尤其是一个平台型公司相对于小型对冲基金的优势。平台带来的是数据的深度和广度，实际上是供量化研究员（QR）分析的交易数据。总的来说，对于交易频率不是非常高（比初创团队或初创公司慢）的投资组合经理而言，我们讨论的换手率范围大概是从零（实际上在对冲基金里没人换手率极低）到每年十五、二十次，甚至更高。对于非高频交易团队来说，拥有日终数据就足够了。

You brought to mind one important competitive advantage that especially a platform has over a smaller hedge fund. What a platform brings is a depth and a breadth of data, of actually trading data for a QR person to analyze. In general, when it comes to portfolio managers that are not trading at a very high frequency, slower than a startup for a team. So we are talking about turnover that goes from let's say zero, nobody has a very low turnover in a hedge fund, from zero to let's say fifteen, twenty times a year, even more actually. Having end of day data is sufficient for non standard teams.

这些通常是日终头寸数据。如今你也可以获得盘中头寸数据，在某些情况下还可以获得PM级别的订单级信息。大多数时候这些不被使用，因为日终数据足以查看所谓的头寸损益（P&L）——即用收盘到收盘的回报率乘以日终头寸数据得到的结果——这往往能较好地模拟实际的会计损益（P&L）。所以，其差异你可以称之为交易损益（P&L）或盘中阿尔法（alpha）。优秀的对冲基金拥有近20年的PM历史数据，并且在任何给定时间都有5到100名PM。

These are typically end of day position data. You can also get nowadays intraday position data, you can get order level information at the PM level in some cases. Most times these are not used because end of day data it's okay to look so called position P and L, which is what you get if you multiply the close to close returns with the end of day position data, tends to mimic the actual accounting P and L relatively well. So the difference you could say is trading P and L or intraday alpha. Good hedge funds have almost twenty years of historical data on PMs, and they have 5,100 PMs at any given time.

你拥有的这些数据不是在PM级别，而是在分析师级别。这确实形成了一个相当庞大的数据集，使你能够跨越时间、跨越PM、跨越行业来归纳行为模式，这是其他任何人都没有的。这一点非常重要。那么你还有什么？这差不多是入场的基本条件了。

You have this data not at the PM level, you have this data at the analyst level. It really gets a pretty significant dataset that allows you to generalize behavior across time, across PMs, across sectors that nobody else has. That's very important. Then what else do you have? It's kind of table stakes.

所有规模尚可的对冲基金和自营交易公司在数据和另类数据方面都有很大的预算。你能获得高质量的回报数据、全球的盘中回报和日终回报数据，拥有优质的证券主数据（sec master），以及丰富的另类结构化数据。所以你可能也能获得非结构化数据，但我会说，将非结构化数据转化为结构化数据的角色，更多地属于数据科学团队，而非量化研究员（QR）。你的基本数据（primitives）基本上是可以用来创建因子模型、定制因子以及进行各种时间序列分析等的大型结构化数据表。

All decently sized hedge funds and prop trading firms have a large budget for data, for alternative data. You get good quality returns, you get intraday returns, end of day returns worldwide, you have a good sec master, and you have a wealth of alternative structured data. So you can also probably get in unstructured data, but I would say the role of transforming unstructured data into structured one, it belongs more to a data science team than to a QR person. Your primitives are basically large tables of structured data that you can use to create factor models, custom factors, and various time series analysis and so on.

在你我交谈之前，我就读过你的著作《高级投资组合管理》。回想起来，我现在觉得这本书有点像一封写给与你共事过的基本面投资经理的情书，或者某种程度上是出于一种挫败感，希望能加速他们的学习进程。我很好奇，在你与这些投资经理的所有合作中，你认为量化研究得出的哪个结果对他们来说最反直觉，最难接受或整合到他们的流程中。

I read your book Advanced Portfolio Management before you and I ever spoke. In hindsight, I now see that book a little bit maybe as a love letter to the fundamental PMs you've worked with or perhaps somewhat in frustration hoping you can get them accelerated in their learning. I'm curious in all your work with these PMs, what do you think the most unintuitive result arising from the QR research is for them to accept or integrate into their process.

随着APM（高级投资组合经理）在对其业绩进行越来越复杂分析和理解的旅程中前进，他们每一步都会感到惊讶。所以我认为，对因子框架相对陌生的APM会觉得这很令人意外。通常这种惊讶是这样的：嘿，今天我在英伟达上赚了一百万美元，但我在英伟达的特异性盈亏却是负的。我确实赚钱了，但为什么我在英伟达甚至整个投资组合的盈亏是负的呢？

As APM progresses in the journey toward more and more sophisticated analysis and understanding of their performance, they get surprised at every turn. So I think that APM who is relatively new to a factor framework will find it surprising. Typically this is the surprise. Hey, today I made a million dollars in Nvidia, but my idiosyncratic p and l in Nvidia is negative. I have actually made money, but how come my p and l in Nvidia and maybe for my whole portfolio is negative?

通常提出这个问题时心情并不愉快。我确实赚钱了。我的意思是，那是我的钱。哦，顺便说一句，他们会说，我在英伟达上赚了钱，市场是平的，但我的EDO盈亏仍然是负的。怎么会这样？

Typically this is asked not in a happy mood. I did make money. I mean, that's my money. Oh, and by the way, they say, I made money in Nvidia, the market is flat, and still my EDO P and L is negative. How come?

然后你必须解释，你必须展示：但是看，英伟达所在的某个行业今天赚了很多钱。而且英伟达是超级多头动量股。动量因子今天表现极佳，把这两件事结合起来，这解释的比你只在英伟达上经历的盈亏要多。所以剩下的部分就是负的了。这有点反直觉。

And then you have to explain, you have to show, Well, but look, some industry that NVIDIA is in has been making a lot of money today. And NVIDIA is super long momentum. Momentum had a field day to put these two things together, and that's explaining more than the p and l that you experienced in NVIDIA. So the rest is negative. This is a little bit unintuitive.

如果我们稍微延伸一下，在所谓的特异性空间中思考在开始时有点困难，特别是如果你来自纯多头背景。在任何空间中思考都是反直觉的，但我认为从长远来看，这对投资经理非常赋能，因为他们学会了利用这一点来获得优势。在某种意义上，他们意识到自己脑海中曾经存在一些混淆变量，现在不再有了。所以他们有了这种关注点分离。我不关心市场，至少在一阶程度上不关心（当然我们可以讨论并非完全不关心），但我不再关心曾经困扰我的许多事情。

And if we extend this a little bit, thinking in so called idiosyncratic space is a little bit hard in the beginning, especially if you're coming from a long only background. Thinking in any space is unintuitive, but I think in the long run is very empowering to PMs because they learn to use this to their advantage. And in a sense, they realize that they had confounding variables in their mind that are not there any longer. So they have this separation of concern. I don't care about the market, at least to a first order, but we can discuss not completely, but I don't care about lots of things that were on my mind.

现在我可以关注，比如说，在EDIUS空间中的业绩差异。另一件更高级但属实的事情是，每个投资经理都相信自己拥有惊人的头寸规模调整技能。所以他们非常擅长调整头寸大小。每个人都知道，我认为这现在是常识了，你的绝大部分盈亏来自于站在股票赌注的正确一边，而不是来自于对某只股票仓位过大或过小。我可以理解人们为什么相信这一点。

And now I can look at, say, divergence in performance in EDIUS space. Another thing that is a little bit more advanced but true is every PM believes that they have amazing sizing skills. So that they're really good at sizing their positions. Everybody knows, I think it's common wisdom at this point, that the vast majority of your P and L is coming from being on the right side of the bet of a stock, not from being overly large in a stock, overly small. I can rationalize why people believe this.

我将尝试用非常、非常定性的术语来解释。在投资经理的横截面中，盈亏是重尾分布的。我的意思是，当你查看一位投资经理年终的盈亏时，通常倾向于在各个股票名称上均匀分布或相当均匀地分布。然而，当盈亏是非常大的正数或非常大的负数时，你会看到各个股票贡献的分布变得非常偏斜。基本上，少数几只股票对那一年的盈亏做出了不成比例的贡献。

I'll try to explain it in very, very qualitative terms. In the cross section of a portfolio manager, P and L is heavy tailed. By this I mean that when you look at the P and L at the end of the year for a PM, typically tends to be equally distributed or fairly equally distributed across names. However, when the P and L is very large positive or very large negative, you will see that the distribution of contribution across stocks becomes very skewed. And basically a handful of stocks contribute disproportionately to the P and L of that year.

所以，重尾分布的一个决定性特征是，当总和很大时，渐近地看，贡献主要来自一两个项，即总和中的最大项。这是事实，在投资组合经理（PM）内部成立，甚至在平台中跨PM也成立。当平台由于某些深层统计原因赚很多钱时，有一个业务对整个损益（P&L）贡献了40%。这是一个非常显著的事实，当你看到这种情况，别人赚这么多钱或你自己有一个丰收年时，你会将其与规模配置联系起来。

So this is really the defining feature of a heavy tail distribution that when the sum is large, the contribution asymptotically is coming from one or two terms, from the largest term in the sum. This is true and it's true within a PM, and it's true even across PMs in a platform. When a platform is making a lot of money for some deep statistical reason, there is one business that is contributing 40% to the whole P and L. This is a very conspicuous fact when you see this, other people making this amount of money or yourself having a killer year. You associate this to sizing.

显然，是我的顶级头寸赚了钱。所以你想做的是，没错，你想超配。但我认为这有点行为偏差，某种程度上是对的，但这取决于你赚了很多钱。在大多数情况下，投资组合经理并没有赚到那种钱。因此他们认为自己有规模配置的技巧。

Clearly, it's my top name that made money. So what you want to do is, yeah, you want to oversize. But I think this is a little bit of a behavioral bias, a little bit it's true, but it's conditional on you making a lot of money. In most cases, a portfolio manager is not making that kind of money. And so they think that they have sizing skill.

我相信这是一个直观的结果，PM需要一点时间来接受它，但QR人员也需要一点时间意识到，在某些条件下，超配对PM来说是正确的事情。我想转向讨论一些关于

I believe this is sort of an intuitive result that takes a PM a little bit of time to accept it, but it's also taking a little bit of time to QR people to realize that sometimes oversizing in certain conditions is the right thing to do for PMs. I wanna transition to talking about some of

对冲角色的问题，我想从一个可能基本但重要的问题开始，根据你的经验，在多经理人公司中，因子对冲实际上是如何进行的？你对公司中的每个经理都应用相同的因子模型吗？它们是不同的因子模型吗？你在PM级别进行对冲，还是全部汇总后在整体级别进行对冲？我希望你能分享一些操作细节。

the hedging role, and I wanna start with maybe what is a basic fundamental question, but I think an important one, which is how is factor hedging actually performed in a multi manager shop in your experience? Are you applying the same factor model to every manager in the shop? Are they different factor models? Are you hedging at the PM level or is it all getting rolled up and hedging at the aggregate level? I was hoping you could share some of the operations there.

让我先稍微定义一下这里的术语，因为有时有些模糊。在多空股票投资组合的背景下，对冲与对冲衍生品组合有些不同。有一些高层次的相似之处，但又不完全相同。相似之处在于，你基本上有一个核心的头寸组合，无论是衍生品还是现金头寸，这是给定的，组合是给定的。你必须添加一个叠加组合，使得整体组合在某种视角下，比如风险视角下，是理想的。

Let me define a little bit the terms here because sometimes there's a bit of ambiguity. In the context of a long short equity book, hedging is a little bit different let's say if you are hedging derivatives book. There are some high level similarities but at the same time not exactly the same. What is similar is that you basically have a core portfolio of positions, be them derivatives or just cash positions, that's a given, the portfolio is given. And you have to add an overlay portfolio that makes the aggregate portfolio desirable under certain lens, under a risk lens.

这就是相似之处。不同之处在于，在股票组合的情况下，你在一个组织中进行，这带来了很多额外的复杂性。而且我会说，对99%的人来说，对冲就是市场对冲。市场对冲是直接的。细节到位，很好理解。

So that's what's similar. What's different is that in the case of an equity portfolio, you do it in an organization and that's generating a lot of additional complexity. And also I would say that for 99% of people hedging is market hedging. Market hedging is straightforward. The details to the mark, well understood.

它们相当准确。你基本上计算你的对冲比率。对冲的交易成本相对较低，然后你就完成了。也有一些细节。你应该对冲到期货吗？

They're quite accurate. You basically compute your hedge ratio. The transaction costs of the hedging are relatively low, and you're done. There are details as well. Should you hedge to a future?

我应该使用SPY吗？还是应该用别的？完美。但使用因子模型时，是有因子的。首先，交易因子比交易市场更复杂、更昂贵且更不纯粹。

Should I use spy? Should I use something else? Perfect. But with a factor model, have factors. First of all, trading factors is more complicated and more expensive and dirtier than trading the market.

因子投资组合本身是市场中真实系统性风险的代理。所以这是个纯数学问题。我们来谈谈组织层面的问题。首先，有些公司根本不进行对冲。即便是非常大的公司，他们也不做对冲。

The factor portfolios themselves are proxies for the true systematic risk in the market. So that's a pure mathematical issue. Let's talk about the organizational issue. First of all, some firms just don't do hedging. Even very large ones, they don't do hedging.

他们做的是我们可以称之为内部对冲的方式，即你给投资经理一个授权，让其因子风险保持在这些界限内，然后他们管理自己的投资组合。他们基本上利用整个覆盖范围的股票池来维持在该限制内。所以我把问题推给了投资经理们，祝他们好运。这样做的问题在于，很多时候即使投资经理们很纯粹，他们也只是在自己层面进行对冲，往往对因子有相同的方向性暴露。

What they do is something that we could call internal hedging, which is you give a PM a mandate, stay within this bounds of your factor risk, and then they will manage their portfolio. They basically are using their whole coverage universe of stocks to keep within that limit. So I am pushing the problem onto the PMs and good luck with it. The problem with this is that sometimes a lot of PMs, even if they are very pure, they are hedging at their own level. They tend to have the same directional exposures to factors.

因此风险呈线性复合，而特质风险大致按投资经理人数的平方根复合。到某个点时，如果你有足够多的投资经理，你的动量暴露会变得如此普遍，以至于你必须采取行动，这时你还是得进行对冲。那么你该怎么做？首先你能做的是——这里我只谈组织层面，不讨论同样要求很高的数学部分——你可以在公司层面创建一个对冲账簿，包含与公司其他部分相反方向的动量暴露。这样当动量亏钱时，公司对冲账簿就能赚钱。

And so the risk compounds linearly, whereas the idiosyncratic risk compounds more or less like the square root of the number of PMs. At some point, if you have enough PMs, your momentum exposure becomes so prevalent that you have to do something, and now you have to hedge anyway. So what do you do? The first thing you can do is, and I'm talking only about organization here, we're not talking about the math which is also demanding, but you can create a hedge book at the firm level that contains, let's say, a momentum exposure in the opposite direction to the rest of the firm. What happens is that if momentum is losing money, the firm hedge book makes money.

但如果动量赚钱，公司就会亏钱。所以公司承担了对冲账簿的风险。那你能怎么办？你可以将其下推一级，通过授权将动量对冲分配给各个投资经理。这是一种可能性。

But if Momentum is making money, the firm is losing money. So the firm is taking the risk of that hedge book. So what can you do? You can push it one level down by mandating that you are distributing basically the momentum hedge to the individual PMs. That's a possibility.

你可以选择一些中间方案，让投资经理转移部分由公司层面持有的动量对冲。你可以完全取消公司对冲，让投资经理购买公司为他们创建的动量对冲，或者他们甚至可以去主经纪商那里购买动量因子对冲——我不推荐这样做。有些ETF是动量因子的，但都是不好的动量因子。我认为一个好的量化研究团队应该内部进行对冲，但只是说存在多种配置方式。此外，你还可以进行一些纯主观的、非系统性的、非程序化的对冲，由公司管理层决定。

You can choose some intermediate solution where you let PMs transfer some of the momentum hedge that is held at the firm level. You could remove the firm hedge altogether and let the PMs buy a momentum hedge that the firm will create for them or maybe they can even go to a prime and buy a momentum factor hedge, I don't recommend that. There are ETFs that are momentum factors, are bad momentum factors. I think that a good QR team should do hedging internally, but just to say there are a variety of configurations. And then what else can you do is you can also have some purely discretionary, not systematic, not programmatic hedging that is decided by the management of the firm.

这是常见的问题。是对冲还是阿尔法？当它变成主观对冲时，两者兼而有之，但那是另一种可能性。所以这些大致是对冲的配置方式。

Is the usual problem. Is it hedging or is it alpha? When it become discretionary hedging, it's a little bit of both, but that's another possibility. So these are broadly the configurations for hedging.

到目前为止，我们在这场对话中是以一种柏拉图式的抽象方式讨论因子的。我们假设这些因子存在，并且到目前为止我们对这些因子的定义是正确的。但在实际实施时，归根结底，特质性P&L或阿尔法，或者无论你怎么称呼它，最终都将由你的因子贝塔来定义。我很想知道，你如何解释特质性风险不一定是纯粹的阿尔法这一观点，并且如果没有一个事前完全覆盖的因子集，你最终可能只是杠杆化了你尚未考虑到的因子风险。

So far in this conversation we've spoken about factors in sort of a platonic abstract way. That they exist and we assume so far that our definitions of these factors are correct. But when it comes to actual implementation, at the end of the day, idiosyncratic P and L or alpha or whatever you want to call it, is ultimately going to be defined by your factor betas. I'd love to know how do you account for the idea that idiosyncratic risk isn't necessarily pure alpha and without a ex ante, you know you've got a fully spanning factor set, you might end up just levering up a factor risk you haven't accounted for yet.

因子模型实际上同时做两件事。那就是定价。所以想象一下，这些因子没有任何回报。因子不是异常现象。你不会投资于因子，但你在你的投资宇宙的回报中考虑了因子。

A factor model does really two things at the same time. That's pricing. So imagine the factors don't have any return. So the factors are not anomalies. You don't invest in factors, but you account for factors in the returns of your universe.

你用因子进行定价，但你也处理异常现象。你想要捕捉异常现象。两者你都需要。没有定价，你就无法进行异常建模。所以因子模型扮演着这种双重角色。

You do pricing with factors, but you also do anomalies. You want to capture anomalies. You need both. You cannot do anomalies modeling without having pricing. So the factor model serves this two fold role.

除此之外，与大多数经济学术文献不同，因子模型还预测风险，即为你的回报创建一个协方差矩阵。同时处理许多、许多事情。那么阿尔法是什么？从数学上讲，它是有明确定义的，但在商业上定义就不那么明确了。真正意义上的阿尔法，讽刺地说，就是你的有限合伙人愿意为之付费的任何东西。

In addition to that, a factor model, as opposed to most of the economic academic literature, a factor model also predicts risk, which is creating a covariance matrix for your returns. Many, many things at the same time. And what is alpha? Mathematically it's kind of well defined, but commercially it's less well defined. What's really alpha is, cynically alpha is whatever your limited partners are willing to pay for.

你可能会说，如果我投资动量，那就是阿尔法。科克伦有一句名言，出自一篇我认为名为《贴现率》的论文。文中他与一位对冲基金经理对话。我怀疑这位对冲基金经理可能是克利夫·阿斯内斯。所以如果他在听这个，如果他能说是或不是我，我将非常感激。

You could say, well, if I invest in momentum, it's alpha. There is a famous quote by Cochrane, which is a paper called Discount Rates, I think. It talks to a hedge fund manager. I suspect that the hedge fund manager could be Cliff Asness. So if he's listening to this, I'd really appreciate if he says, no, no, it was me or it wasn't me.

但基本上，因为他们总部在芝加哥，那位对冲基金经理说，是的，这些是因子，但我知道如何交易它们。科克伦说，没有阿尔法。只有你知道的因子和你知道的因子。有你了解得更好的，也有你不了解的。从这个意义上说，你可以说没有阿尔法。

But basically, because they're based in Chicago, the hedge fund manager says, yes, these are factors, but I know how to trade them. Cochran says, there is no alpha. There are factors that you know and factors that you know. There is better that you know and better that you don't know. In this respect, you could say there is no alpha.

在某种程度上，一切都是因子，尤其是对于策略而言。他们交易的很多东西实际上都是普遍存在的因子。好吧。很好。但也不完全正确。

To some extent, everything is a factor, like especially for strategies. A lot of what they trade is really pervasive factors. Okay. Great. But also not true.

还有一些真正纯粹的阿尔法策略。单一名称、非常非常狭窄的主题。那么因子模型的范畴是什么？投资组合经理应该为什么获得报酬？如果他们驾驭了时间或动量因子，应该为此获得报酬吗？

There are also things that are really pure alpha. Single names, very, very narrow themes. And so what is in the scope of a factor model? And what should be a PM paid for? If they tame time or momentum, should they be paid for that?

也许应该。如果他们只是做多动量因子，应该为此获得报酬吗？也许不应该。这个问题确实没有标准答案。这在一定程度上都是可协商的，也是一个优秀量化研究团队设计或制定正确激励机制的一部分，以激励投资组合经理构建他们的投资组合。

Maybe yes. If they are just long momentum, should they be paid for that? Maybe not. There is really no right answer on this. It's all partially negotiable and it's also all part of a good QR team to design or to devise the right incentives for PMs to create their portfolios.

不过我想说的是，对于投资组合经理和量化研究人员而言，智慧的首要标志是对模型本质不完美的谦逊认知。有时候会出现模型未能捕捉但明显普遍存在的损失，这说明模型不够好。而有时候盈亏中也存在因子效应，但因子模型未能捕捉到。你知道它存在，但很难精确界定。

I would say this though, the number one sign of wisdom, I think for both portfolio managers and quantitative researchers, is a level of humility in understanding that models are by definition imperfect. Sometimes there are losses that are not captured by a model, but they are clearly pervasive. So the model is not good. And sometimes there are P and L that is also factor, but it's not captured by the factors. You know it's there, but it's very difficult to pin it down.

模型无法捕捉它。这非常困难。人们往往将因子模型视为既定事实。市面上有许多供应商销售的因子模型等等。但因子模型远非已解决的问题。

The model doesn't capture it. It's very hard. People treat factor models as a given. There are many factor models sold by vendors and so on. Factor models are not a solved problem.

我们离解决这个问题还非常遥远。这是我的反向观点。

We are very far from it. That's my contrarian belief.

我们肯定会深入探讨因子模型。不过在进入正题之前，我想简要谈谈内部阿尔法捕获这个概念。因子模型将是个长篇讨论，而这个话题相对简短，所以我想先解决这个部分。我想谈谈量化研究的第三项职能——内部阿尔法捕获。

We're definitely gonna dive into factor models quite a bit. Before we go there though, I wanna briefly touch on this idea of internal alpha capture. Factor models are gonna be a long discussion. This one, a little less so, so I figured I'd get this one out of the way. I wanna talk about this third part of the QR role, which is internal alpha capture.

在预通话中您曾说过：'一旦解决了业绩问题，持续面临的挑战就是如何配置资金'。您能解释一下这句话的含义吗？

And on our pre call, you said, quote, once you solve the performance problem, the ongoing problem is how to deploy capital. Can you explain what you meant by that?

我认为这有点像生活中的情况。在生活中，你首先解决性能问题，然后解决扩展问题。首先你要明白自己擅长什么，然后你想最大化你的超能力，就像你希望在自己做的事情上取得最大程度的成功。某种程度上，对冲基金也面临同样的问题。

I think it's a little bit like in life. First in life, you solve a performance problem and then you solve a scaling problem. First you want to understand that you're good at something. And then you want to maximize your superpower, like you want to succeed the most amount possible at what you do. In a sense hedge funds face the same problem.

他们起步时，谁知道他们好不好。他们实际上会优化流程。当他们表现出色时，确实会有很高的夏普比率。我认为，任何足够优秀的人，在达到一定规模时都能赚钱。

They start and who knows whether they're good or not. They actually refine their processes. They're good. They actually have a high shark. Everybody who's sufficiently good, I think, at some level of size can make money.

对某些人来说，这可能是非常大胆的说法。如果他们足够优秀，就能赚到一些钱。困难的部分不是用高夏普比率赚点钱，而是用足够高的夏普比率赚大钱。有很多自营交易公司的夏普比率达到10，有些策略的夏普比率甚至达到30。

This is a very maybe bold statement to someone. If they're good enough, you can make some money. The difficult part is not to make some money with a high sharp, it's to make a lot of money with sufficiently high sharp. There are lots of prop trading firms that have a sharp of 10. There are strategies of prop trading firm that have a sharp of 30.

谈论夏普比率30甚至都没有意义。但如果我一年赚5000万美元，谁在乎呢？问题是要以夏普比率2一年赚150亿美元。这难度无限大。而这就是赚钱的地方，这就是你想做的，作为对冲基金或任何你想实现的目标，你希望为你的有限合伙人赚最多的钱。

That doesn't even make sense to talk about a sharp of 30. But if I make $50,000,000 a year, who cares? The problem is to make $15,000,000,000 a year with a sharp of two. It's infinitely harder. And that's where the money is, that's what you wanna do, that's what actually as a hedge fund or whatever you want to achieve, you want to make the most money for your limited partners.

所以这实际上是最正确的衡量标准。如何做到这一点？通过学习如何扩展。从商业角度来看，学习如何扩展意味着知道如何复制模型并为不同资产类别创造相同的激励机制。如何构建资产类别策略非常困难，但可以学习，也有人掌握了它。

So that's actually the most correct metric. How do you do this? By learning how to scale. From a business perspective, learning how to scale means know how to reproduce the model and generate the same incentives for different asset classes. How to build an asset class strategy is really hard, but can be learned and somebody masters it.

然后在某个资产类别内，你想最大化该资产类别的容量，特别是对于基本面股票。如果你没注意到，但钱真的会在任何层面上扰乱你的心智。你赚了很多钱，可能就开始吸可卡因。你在回撤中损失很多钱，疯狂的事情就会发生。比如你开始出现身体抽搐。

And then within an asset class, you want to maximize the capacity of that asset class, specifically for fundamental equities. In case you didn't notice, but money really messes up with your head at any level. You make a lot of money, probably you start sniffing coke. You lose a lot of money in a drawdown and crazy things happen. Like you start developing physical tics.

我见过这种情况发生。他们失眠。这发生在我身上，比如你损失500万美元就已经很糟糕了。正因为如此，投资组合经理有时只愿意承担一定量的风险。这是他们的操作特性。

I've seen this happen. They lose their sleep. It happened to me, like you lose $5,000,000 it's already terrible. Because of this, portfolio managers sometimes are comfortable running only a certain amount of risk. That's their operating characteristic.

另外，有时候，因为他们是人类并以人类的方式进行交易，他们并非以最优方式交易，因此他们的能力有限。他们不能轻易地实现增长。他们开始看到价格受到自己行为的影响。在这种情况下，如果一家公司成功了，通常他们总是拥有比能轻易部署的更多的资金。你遇到了成功诅咒。

Also sometimes, because they're humans and trade in a human manner, they are not trading optimally and so their capacity is limited. They cannot really grow that easily. They start seeing the price being affected by their own actions. In cases like this, if a firm is successful, typically they always have more money than what they can easily deploy. You have a success curse.

有时候，你知道，你拿了太多钱，这对你的回报产生了负面影响。因此，内部阿尔法捕获是在投资组合经理的核心投资组合之上叠加另一个投资组合的过程，其设计方式高效，能够将适当的风险分配给各个阿尔法来源。它消除了原始阿尔法来源在交易中的行为偏差，以最优方式进行交易，并部署额外资本，为公司产生额外的损益。这就是所谓的内部阿尔法捕获。为什么是内部的？

And sometimes, you know, you're taking too much money and it's affecting negatively your returns. And so internal alpha capture is the process of overlaying yet another portfolio to the core portfolios of the portfolio managers designed in an efficient way so that it allocates the right risk to the individual sources of alphas. It removes the behavioral biases in trading of the original sources of alphas, trades that optimally, and deploy the additional capital and generates additional P and L for the firm. So that's what's called internal alpha capture. Why it's internal?

因为，这里就有点主观了，但我认为内部阿尔法捕获团队的使命，正如其名称所示，是强制性地只使用内部阿尔法。如果你将内部阿尔法与外部信号混合在一起，那么我认为问题就变得有点模糊了。一般来说，你应该有一个分离原则，即分离定理。如果你有外部信号，欢迎你启动自己的低频系统性外部阿尔法捕获账簿。

Because, and this is where it gets a little bit subjective, but I think that a mandate of an internal alpha capture team, as the term says, is you are by mandate using only your internal alphas. If you are stacking commingling your internal alphas with external signals, then I think that the problem becomes a little bit more opaque. And in general, you should have a separation thing, the separation theorem. If you have external signals, you're welcome to start your own low frequency systematic external alpha capture book.

当你考虑你活动的投资回报率（ROI）时，从公司整体损益的角度看，投资组合经理覆盖的ROI与内部阿尔法捕获的ROI相比如何？

When you think about the ROI of your activities, how does the ROI of PM coverage compare to the ROI of internal alpha capture when you're looking at firm wide P and L?

我在这里会非常主观。就损益而言，不考虑边际，而是考虑绝对值，内部基础设施具有难以置信的价值。在某些情况下，它几乎可以使基本面业务的损益翻倍，也许最多翻倍。即使只增加30%，也很多了。平均而言，这是一个巨大的贡献。

I'm going to be very subjective here. In terms of p and l, not thinking at the margin, but thinking absolute terms, internal infrastructure has an incredible value. It can, in some cases, almost double the P and L of a fundamental business, maybe tops double. Even if it's increasing it by 30%, it's a lot. It's on average is an incredible contribution.

而且这是一个非常具有挑战性的问题。就绝对值的投资回报率而言，与其他选择相比，它是巨大的。一旦你达到了阿尔法捕获的运营效率，边际改进就变得非常困难。很难再挤出那额外的半个百分点。这变得很有挑战性。

And it's a very challenging problem. In terms of ROI in absolute terms, compared to the alternative, it's huge. Once you get to the operating efficiency of alpha capture, it becomes very hard to improve at the margin. It becomes really hard to find that extra juice to make another half percent. It becomes challenging.

仍然存在很多问题。世界一直在变化，诸如此类。在边际上，一旦你处于运营效率的极限，这就变成了一个难题。与无人看管的APM（助理投资组合经理？此处原文APM可能为笔误或特定缩写，按字面保留）相比，覆盖的投资回报率较低。

There are still lots of problems. The world changes all the time and whatnot. At the margin, once you are at the operating efficiency envelope. That's becoming a difficult problem. Coverage has a lower ROI compared to APM left unattended.

你进来，稍微改进一点，告诉别人不要这样做，避免明显的错误。但覆盖的好处在于它是一个持续的学习过程。覆盖没有上限。没有人解决了投资问题，也没有人解决了覆盖问题。总有很多随着时间的推移而复合的改进空间。

You come in, you improve a little bit, you tell somebody don't do this, don't do obvious mistakes. But the benefit of coverage is that it's an ongoing learning process. There is not a maximum for coverage. Nobody has solved the investment problem and nobody has solved the coverage problem. There is always a lot of improvement that compounds over time.

与过去完全忽视因子的传统投资经理相比，我认为覆盖整体上在风险调整后表现上额外增加了约50%。并且这种提升将在未来持续下去。覆盖在沟通和扩展内部制造等问题方面具有难以量化的投资回报率。现在说说对冲。对冲不会以任何方式增加盈亏，但它能提高策略的夏普比率。

At this point compared to the old I think PM stock picker who was completely oblivious to factors, coverage as a whole has added also like another 50% in risk adjusted performance. And it will keep going for the indefinite future. Coverage has a very difficult to quantify ROI in terms of communicating and extending the problem for internal manufacturing and whatnot. Now hedging. Hedging doesn't increase P and L in any ways, but it improves the Sharpe ratio of the strategy.

提高多少？从理论上说可能是50%，这很巨大，到更可实现的，我不知道，20%。你也可以将其视为投资回报率，因为如果公司从风险角度思考，如果我的夏普比率提高了20%，我就能承担更多风险。我将风险降低20%，规模扩大20%，盈亏就能增加20%。所以这也非常非常有价值。

By how much? From I would say theoretical maybe 50%, which is huge, to a more achievable, I don't know, 20%. You can consider that also ROI because if the firm thinks in terms of risk, if I have a 20% higher sharp, I can take more risk. I reduce my risk by 20%, I scale up by 20%, and I have 20% more P and L. So it's also very, very valuable.

它们都是相辅相成的，因为如果我能部署更多风险，那么我就面临阿尔法捕获的问题。我需要以某种方式部署这些风险。因此很难精确量化，但内部阿尔法捕获确实能为公司的盈亏做出很大贡献。不过这三者在不同方面都很重要。

And they all come together because if I can deploy more risk, then I have an alpha capture problem. I need to deploy this risk in some way. So it's difficult to put exactly a number on it except that internal alpha capture can really contribute a lot to the P and L of a firm. But all three are important in different ways.

好的。让我们深入探讨因子研究的核心内容。我们有很多要讨论，但也许先从基础开始：经常有人说，在因子模型中，学者使用收益率，而从业者使用特征。我知道你对这一点有强烈看法。为什么特征是一种更优越的创建因子的方法？

Alright. Let's get into the meat and potatoes of factor research here. We got a lot to cover, but let's start with maybe a bit of a basic one here, which is that it's often said that in factor models, academics use return where practitioners use characteristics. I know this is something you feel strongly about. Why are characteristics a superior means of creating factors?

为不熟悉因子模型或股票的听众提供一些背景，我认为你的听众中实际上有很多衍生品领域的人。你们很可能看过Fama French 1993年的论文或一些更近的提及Fama French的论文。Fama French有一种非常特殊的构建因子模型的方法，这实际上源自Fama Macbeth 1973年的另一篇论文。那些论文采用股票收益率与因子代理之间的时间序列协方差。Corey所指的是，如果你打开Barra或Axioma模型的手册，你会看到的不是股票收益率与捕捉市净率的因子投资组合的协方差。

As a background for the listener who is not steeped into factor models or equity, I think there are lots of people who are actually derivatives people among your audience. You have most likely seen the Fama French '93 paper or some more recent paper that's mentioning Fama French. Fama French has a very specific way of constructing a factor model that actually originates in another paper by Fama Macbeth from 1973. Those papers employ time series covariances between the stock returns and a factor proxy. What Corey was referring to is that instead if you open the manual of Barra or Axioma model, you will see not the covariance of the stock returns to a factor portfolio that captures price to book.

它不是市净率因子投资组合与英伟达收益率之间的相关性。它实际上是该公司的账面市值比特征。或者我可以用盈利价格比。那是另一个度量。为什么特征优于构建Fama French模型的协方差方法？

It's not the correlation between a price to book factor portfolio and Nvidia's return. It's actually the book to price value characteristic of that firm. Or I could use earnings to price. That's another measure. Why is characteristics beating covariances built the Fama French model?

首先，你可以辩称特征不是历史性的，根据定义它们实际上可以是前瞻性的。它们可以包括在我的收益价格比中获得共识。事实上，它们确实如此。特征不是长期时间序列协方差，后者根据定义是向后看的。它们或多或少捕捉了股票在某个时间点的特征。

You could argue, first of all, characteristics are not historical and they can actually be by definition forward looking. They could include earning consensus in my earnings to price. And in fact, will do it. Characteristics are not long term time series covariances that are by definition backward looking. They capture more or less the point in time features of the stock.

这也非常直观。我试图像在回归问题中那样快照所有相关特征。我想知道这就是公司当前的状态。特征并不明确描述公司在时间序列因子风险敞口方面的风险。它们真正旨在根据公司当前的外观来解释当前横截面中的公司回报。

It's also very intuitive. I am trying to snapshot all the relevant features like in a regression problem. I want to know this is the firm right now. Characteristics do not explicitly describe the risk of a firm in terms of risk exposure to a time series factor. They're really meant to explain the returns of the firm right now in the cross section based on what they look like right now.

还有一个学术验证，我认为主要是由肯特·丹尼尔大约二十年前发起的。他和合著者，比如纳迪亚·丹尼尔和迪特曼，比较了协方差与特征。你可以看到，如果我已有协方差，然后在同一模型中再加入特征，我是否获得了额外的解释力？或者反过来呢？特征往往更优越。

There is also an academic, if you want validation, which is mostly initiated I think by Kent Daniel, basically about twenty years ago. He and co authors, think Nadia Daniel and Dietman, compared covariances versus characteristics. And you can see if I have covariances and then I do characteristics afterwards in the same model, do I have additional explanatory power? Or how about the other way around? Characteristics tend to be superior.

特征也往往易于集成到大型因子模型中。当你拥有多个协方差时，存在结构性问题，在基于协方差的因子模型中，创建这些代理投资组合会变得稍微复杂一些。总的来说，特征在实践中往往表现相当好，并且非常灵活。

Characteristics also tend to be easy to integrate in large factor models. There are structural problems when you have multiple covariances, multiple models in a covariance based factor model becomes a little bit more complicated to create this proxy portfolios. In general, characteristics tend to perform quite well in practice and to be very flexible.

我想把你之前的回答联系回我关于因子跨越和异质性风险定义的问题。你在回答中谈到了账面价格比和收益价格比作为定义因子的可能特征，这两者我通常会说属于价值因子。但你也可以考虑像动量这样的东西，比如三个月动量或十二个月动量。或者你可以看特定行业内的动量，比如医疗保健动量。天真地说，这些选择将为因子定义不同的结果。

I want to tie back to an answer you had before to my question about factor spanning and this definition of idiosyncratic risk. In your answer there you spoke both about book to price and earnings to price as possible characteristics to define a factor, both of which I would say typically are characterised under the value factor. But you could also consider something like momentum where you've got three month momentum or you've got twelve month momentum. Or you could look at momentum within a specific sector, healthcare momentum. And naively these choices are going to define different outcomes for the factor.

我很好奇，从你的角度来看，这些具体选择对于你试图通过因子对冲或因子模型达到的最终结果有多重要？

And I'm curious in your perspective, how important are those specific choices to the ultimate end result of what you're trying to get at with factor hedging or factor models in general?

对于因子对冲，它们确实非常重要。所以让我们先看看这些特征的起源。它们来自哪里？我会说商业供应商包含这些特征是因为它们被发表了。它们出现在某篇论文中。

For factor hedging, they do matter a lot. So let's look first at the genesis of this characters. Where are they coming from? And I would say that commercial vendors include those because they get published. They appear in some paper.

它们出现在论文中的方式之所以被证明合理，是因为每个单独的特征都是针对标准模型（通常是Fama French模型）独立测试的，仅作为一种定价异常，即该因子能解释超出市场收益的额外超额收益。账面市值比出现在最初的Fama French三因子模型中，但之后你可能会说，哦，我还有盈利价格比、盈利波动率、盈利能力。它们都是不同的异常现象。所以一个被称为价值因子，另一个被称为盈利因子，再一个被称为质量因子，等等。现在问题来了。

And the way they are appearing in a paper and they're justified there is because each individual characteristic is tested in isolation against a standard model which is the Fama French model, typically only as a pricing anomaly in the sense that this factor explains some additional excess return in excess of the market. Book to price, it's in the original Fama French three factor model, but then you can say, oh, I have also earning to price, I have earnings variation, I have profitability. And they are all different anomalies. So one will be called value, another one will be called profitability, another one will be called quality, and so on. Now problems.

首先，它们的产生方式有点可疑，因为这有点像每次向一个三变量回归问题添加一个变量，但从未同时考察所有这些变量在模型中的存在。这可能是因为学术界通过这种方式发表论文并获得终身教职，但这是一个问题。所以当你把它们放在一起时，它们经常相互解释，存在一定程度的冗余。然后有时你会观察到，这些特征本身在任何给定时间点在股票之间往往是相关的。这就是所谓的共线性问题。

First of all, the way that they come about, it's a little bit suspicious because it's a little bit like adding one variable at the time to a three variable regression problem, but never looking at the presence of all these variables at the same time in a model. And because this is the way academics publish paper and get tenure probably, but this is one problem. So when you put them together, they end up very often explaining each other, there is a level of redundancy. And then sometimes you observe that the characteristics themselves at any given point in time across the stocks tend to be correlated. So that's called a collinearity problem.

如果你还记得线性回归，当你有两个共线性预测变量时，你的回归结果会变得不稳定，因为回归参数的误差会随着这些特征的共线性程度而膨胀。在金融领域，当误差膨胀时，膨胀发生在因子本身的误差估计上。这种误差会传播到这些因子投资组合的纯度或缺乏特质风险中。所以我认为因子的选择很重要。这对对冲尤其重要，因为你开始需要对冲那些在某种意义上非常嘈杂的因子。

If you remember your linear regression, when you have two collinear predictors, the outcome of your regression becomes unstable because the error around the parameters of your regression becomes inflated as a function of the collinearity of these characteristics. And when it becomes inflated in finance, the inflation happens on the error estimates of the factors themselves. And this error propagates into the purity or the lack of idiosyncratic risk into these factor portfolios. So I would say that the choice of factors is important. It's important for hedging specifically also because you start having to hedge to factors that are really very noisy in a sense.

供应商有一种不正当的激励，总是向模型添加因子或追逐当前潮流。在某些时候，Barra引入了一些阿尔法因子。我不会对具体选择发表意见，但如果供应商推出一个10因子模型后就止步不前，那在商业上可能是自杀行为。这也是QR主题花费大量时间开发定制因子模型的原因之一。因为它们是根据公司需求定制的，并且在某种意义上更具学术严谨性。

And there is a perverse incentive by vendors to add factors to models all the time or to come with the current fashion. At some point Barra introduced some alpha factors. I won't opine on the specific choices but definitely if the vendors came up with a 10 factor model and that's it, that would be probably commercially suicidal. And this is one of the reasons why QR themes spend a lot of time developing custom factor models. Because they are tailored to the needs of their firm and because in a sense they are more intellectually rigorous.

在讨论因子研究时，经常出现的一个词是'正交化'。这尤其在从业者中经常被提及。它是什么？为什么如此重要？为什么它对通过因子视角查看投资组合的基本面投资经理相关？

One of the words that comes up all the time when talking about factor research is this phrase orthogonalization. And that comes up especially among practitioners. What is it? Why is it so important? Why is it relevant to fundamental PMs who are looking at their portfolio through this factor lens?

这个答案需要稍微回忆一下基本回归知识。除此之外，我认为很多对量化投资或数据科学感兴趣的听众，可能在他们的书库中有Friedman、Hastie和Tibshirani的《统计学习基础》，这是一本非常好的书。虽然有点旧，但仍然很棒。书中有一个很棒的章节我推荐阅读，关于将多元回归作为一系列单变量回归来处理。简而言之，你可以一次性进行多元回归，或者想象你有一种技术或求解器，每次只进行一个单变量回归。

This answer requires a little bit of remembering of basic regression. Aside from that, I think that a lot of your listeners who are interested in quant investing or interested in data science, probably in their library they have the elements of statistical learning by Friedman, Hastian, Tibshirani, which is a very nice book. It's kind of old, but it's still very nice. And there is a beautiful section there which I recommend, which is multivariate regression as a sequence of univariate regressions. In a nutshell, you could do a multivariate regression in one shot, or imagine that you have a technology or a solver that does one univariate regression at a time.

所以你可以做的是：首先将我的y变量对一个预测变量进行回归，然后取残差，即第一次回归的误差。然后我有第二个预测变量，我做的有点反直觉的是，将第二个预测变量对第一个预测变量进行回归。并取这个回归的残差。所以我实际上做了两次回归：y变量对第一个预测变量，然后第二个预测变量对第一个预测变量。第二步是将y对第一个预测变量的残差对第二个预测变量对第一个预测变量的残差进行回归。

So what you can do is I regress my y variable against one predictor, then I take the residual, so the error from this first regression. Then I have a second predictor, and what I do is counter intuitively, I regress the second predictor against the first predictor. And I take the residual of this regression. So I do really two regressions, y variables versus first predictor, and then second predictor against first predictor. And then second step is I do regress the residual from y versus first predictor against the residual of second predictor versus first predictor.

基本上，你现在有两个残差项，将一个对另一个进行回归。将第二个预测变量对第一个预测变量进行回归的直观理解是：我提取第二个预测变量中无法被第一个预测变量解释的那部分。这被称为正交化。它具有许多优良特性，其中之一是我无需重新估计第一个预测变量的系数。

Basically now you have two residuals that you regress one against the other. Regressing the second predictor against the first predictor has the intuition that I take that part of the second predictor that I cannot explain by the first predictor. That's called orthogonalization. It has many beautiful properties. Among them is the fact I do not need to re estimate the coefficient of the first predictor.

我可以加入第二个预测变量、第三个预测变量，依此类推。每一个基本上都是在回归中添加我无法用第一个预测变量解释的部分。这就是为什么它被称为正交化。它的作用是帮助你理解我引入的任何额外变量对回报的贡献是什么。这既好又糟糕。

I can add the second predictor and the third predictor and so on. And each one of them is basically adding what I can't explain from the first predictor to the regression. So that's why it's called orthogonalization. And what it does is helping you understand what is the contribution to returns from any additional variable that I have introduced. And that's nice and terrible at the same time.

好的一面是，例如，我不必重新估计一个因子模型，它能告诉我真正能额外理解什么，即添加一个新变量。糟糕的一面是，你经常会发现，如果我进行正交化，许多看似有价值的东西实际上在因子层面并没有价值。它们不是好因子。这是坏事吗？既是也不是。

It's nice because I don't have to re estimate a factor model, for example, and it's telling me what I really can understand in addition, adding a new variable. It's terrible because you will find often that if I orthogonalize, a lot of things that seemed to have value don't have actually value in terms of factors. They are not good factors. Is that a bad thing? Yes and no.

你不能将这些变量作为因子加入，但也许它们作为主题仍然具有一定的预测价值，或者在某些时间点非常、非常有预测力，等等。但这就是正交化的本质，它基本上是试图理解、估计一个变量在现有因子测试基础上的额外预测价值。这实际上是一个智力过程。许多通常不在专业领域的人不这样做，因为他们无法接触因子模型。

You cannot add those variables as factors, but maybe they still have some predicted value in terms of being themes, or maybe they are very, very predictive at some points in time, and so on. But this is what orthogonalization is, it's basically trying to understand, estimate the additional prediction value of a variable against the test bed of existing factors. It's an intellectual process really. Many people who typically are not in the professional realm don't do it because they do not have access to a factor model.

你在因子研究领域已有相当长的时间。你目前正在润色你的新书，该书相当广泛地涵盖了因子研究。我最近在Twitter上看到，你正在进行一项关于因子研究的大型文献综述。关于因子构建，或许还有因子模型本身，你有什么坚定的观点，认为你的同行们可能不一定认同？

You've been in the field of factor research for quite some time. You're currently polishing your newest book, which covers factor research quite extensively. I saw recently on Twitter, you were doing a large literature review of factor research. What strongly held views do you have about factor construction and maybe factor models in general that you don't think would necessarily be shared by your peers?

我会重申我之前提到的几点，但这值得重复。首先，因子模型是一个已解决的问题。这很无聊。我从非常资深的人那里听到过这种说法，也从面试求职者时听到过。

I will reiterate a couple of things that I mentioned before, but it's worth repeating. So the first is factor models are a solved problem. That's boring. I have heard this from very senior people. I've heard this from people I interviewed for jobs.

构建一个因子模型能有多难？你有你的输入，你做横截面回归，你计算因子回报，你用移动平均做指数加权。通常当我听到这些时，我认为我正在面试的人并不知道自己不知道什么。实际上，一知半解是非常危险的。我们真的不知道一个好的因子模型应该是什么样子。

How difficult could it be to make a factor model? You have your inputs, you do the cross sectional regression, you do your factor returns, you do exponentially with the moving average. Typically when I hear this, I think that the person that I'm interviewing doesn't know what they don't know. Actually it's a very dangerous thing to have just a little knowledge. We really don't know what a good factor model should be.

挑战来自于供应商的激励机制和历史因素，因为一旦你建立了一个因子模型，就很难撤回或大幅改变。这是一个非常技术性的问题，也是一个深层次的问题。它不像做PCA（主成分分析）那么简单，远非一个已解决的问题。

There are challenges that are coming from incentives for vendors, history, because once you have established a factor model, it's very difficult to recant or to change dramatically. It's a very technical problem. So it's a deep problem. It's not like, oh, you do a PCA or something. It's far from being a solved problem.

我还要说，这不是一个可以忽视的问题。你不能简单地说‘让我们用深度学习来解决它’。它在概念上是必需的，而且很困难，因此还有很多工作要做。另一个常见的观点是因子越多越好。这很有趣，因为在某种程度上你可以认为更多因子可能更好。

And I would also say it's not a problem that you can ignore. You cannot say well let's throw deep learning at it. It's conceptually needed and it's difficult so there is still a lot of work to do. Another common wisdom is more factors are always better. This is interesting because in a way you could argue more factors could be better.

有些论文说实际上可以有无限个因子。但如果你是为了风险管理和投资组合构建而做因子，实际上在因子数量上是有物理限制的。你可以有无限的预测变量来捕捉异常，这很好，但我认为现实中不可能有无限的风险预测变量。关于拥挤交易（crowding）的讨论很多。拥挤是对冲基金的两个主要风险来源之一。

So there are some papers that say you could actually have infinite factors. But if you want to do factors for risk management and portfolio construction, you do have actually a physical limit in the number of factors. You could have infinite predictors, that's great, to capture anomalies, but I don't think that realistically you can have infinite predictors of risk. There is a lot of talk about crowding. Crowding is one of really the two major sources of risk for hedge funds.

我的意思是，你有市场风险，你会对冲这个，但还有拥挤风险，这基本上是重叠的，每个人都有一致的想法，每个人都被迫在同一时间平仓，有时被称为流动性螺旋。拥挤通常被作为因子添加到因子模型中，这没问题。我不是说这是错的。但拥挤不是一个因子。为什么它不是因子？

I mean, you have market risk, you hedge to this, but then there is crowding, which is basically the overlapping, having everybody the same ideas and everybody being forced to liquidate your book at the same time, sometimes called liquidity spirals. Crowding is typically added as a factor in factor models, which is okay. I'm not saying that it's the wrong thing to do. But crowding is not a factor. Why it's not a factor?

因为它实际上是一个内生事件。它不是来自外部，也不是日复一日重复发生的事情。拥挤有点像动量，动量也可能是一个非常异常的因素，因为它就像一个动态系统，由明显相互作用的代理生成，达成共识。这是一个模仿过程，我们互相复制，押注相同的股票，自我创造成功。

Because really it's an endogenous event. It's not coming from the outside. It's not something that repeats day after day. Crowding is a little bit like momentum, which is also probably a very anomalous factor in the sense that it's like a dynamical system is generated by agents clearly interacting with each other, having consensus. And there is a mimetic process where we copy each other and we bet on the same stocks and we self create our success.

然后我们也自我制造分崩离析。这个过程如何形成是复杂的。这是另一个未解决的问题，但你不能只是说‘我添加一个拥挤因子，它大多数时候会赚钱，然后偶尔我会亏很多钱’。作为一个因子，这并不那么有用。

And then we also self create our distraction. And how this process comes to be, it's complicated. It's another open problem, but you cannot just say, I'm adding a crowding factor, which is gonna make money most of the time. And then once in a while, I am losing a lot of money. That's not that useful as a factor.

作为一个因子，它也没有那么有洞察力。因子建模仍然是一个广阔的开放领域。更多因子更好。拥挤只是你添加到因子模型中的一个因子。

It's not that insightful as a factor either. Factor modeling is still a wide open field. More factors are better. Crowding is just a factor that you added to a factor model.

我记得某些听众会记得这个论坛，有一个叫Nuclear Finance的论坛在2000年，也许是2006年或2011年左右非常流行，算是它的鼎盛时期。我记得在那里读到过一个理论，有人提出像Citadel、Millennium和HRT这样的公司都采用因子模型，比如大家都采用Barra模型，导致了这些拥挤效应。随着时间的推移，我听到过同样的说法，但还加入了其他因素，比如现在资本向这些大公司集中，以至于能够运营的小公司越来越少，它们无法竞争。因此，资本集中到这些更少的公司中，它们可能使用相似的因子模型。

So I remember certain listeners are gonna remember this forum, but there was a forum called nuclear finance that was very prevalent back in 2000 maybe 2006 or 2011 it was sort of in its heyday. And I remember reading a theory there where someone posited that the adoption of factor models across firms like Citadel and Millennium and HRT by adopting all the same factor model, everyone adopting the Barra model, for example, was leading to these crowding effects. And over time, I've heard that same thing mentioned, but with other things added in such as you now have a consolidation of capital towards these major firms so that there's fewer smaller firms that are able to operate anymore. They can't compete. And so you get a consolidation of capital into these fewer firms using potentially similar factor models.

然后随着人员从Millennium跳槽到Citadel，还会出现一些知识产权交叉。这实际上导致了因子模型知识产权的进一步趋同。所有这些都增加了未来发生拥挤和量化地震式事件的风险。这只是一个吓人的故事吗？你对此有什么看法？

And then you get some IP crossover as people move from Millennium to Citadel. And so that actually leads to increased convergence of the factor model IP. All of this leading to an increased risk of crowding and Quantquake esque like events going forward. Is it just a boogeyman? What are your thoughts there?

让我们把你的问题或陈述拆分一下。我认为这里面有多个问题。第一个是，这个领域是否存在某种拥挤？首先，在某种程度上这是可以量化的，因为你可以观察拥挤特征随时间的变化。实际上，我不确定你是否一定能在拥挤特征中看到这一点。

Let's split the question or the statements in your broader question. So I read this as multiple questions. So the first is, is there some crowding in the space? First of to some extent this can be quantified because you can look at the crowding characteristic maybe over time. Actually, I'm not sure that you would see this necessarily in the crowding characteristic.

然而，这就像 obscenity（淫秽内容）一样，当你看到时就能感觉到。有一种身处拥挤空间的感觉。但我的观点是，这个问题的答案是肯定的。某些空间确实变得越来越拥挤。

However, it's like obscenity. You feel it when you see it. There is a sense of being in a crowded space. But the answer to this question is, in my view, yes. Some spaces are becoming more crowded.

我稍后会尝试证明这一点。还有第二个问题。第二个问题是，但是每个人都在做同样的事情，比如使用相同的因子模型。这正在产生拥挤。对此，我强烈认为这是胡说八道。

I'll try to justify this later. There's a second question. The second question is, but everybody is doing, for example, the same thing with factor models. And this is generating crowding. To this, I have a strong view that it's BS.

这是一种人们不假思索就重复的陈词滥调。如果你建立一个相对简单的模型，来分析每个参与者都使用完全相同的因子模型如何可能导致阿尔法拥挤，你会得不出任何结论。完全没有任何关系。实际上，它可能产生阿尔法的分散。但总的来说，我可以在某个时候证明这并不会产生拥挤。

It's the kind of common wisdom that people repeat mindlessly without thinking too carefully about the statement. If you do a relatively simple model of how having exactly the same factor model for every participant could generate crowding in alphas, you will come to nothing. There is absolutely nothing. It could actually generate dispersion across alphas. But in general, I can show at some point that that's not generating crowding.

它并没有集中不同参与者之间的阿尔法向量。这不是一个强烈的观点，但每个人都以相同的方式、使用相同的因子模型管理风险正在产生拥挤的说法是不正确的。

It's not concentrating the vector of alphas across participants. Not strongly held view, not true that everybody managing risk the same way to the same factor models is generating crowding.

是否存在这样一种可能的论点：尽管它不必然在阿尔法策略中造成拥挤，但可能会在这些贝塔的对冲操作中造成拥挤？

Is there a potential argument that even though it doesn't necessarily create crowding in the alphas, it can create crowding in the hedging of those betas?

我不这么认为。但我可以告诉你——这其实很精妙——如果你回顾几年前的情况，当时使用的因子模型通常会在月初第一天发布新的基本面特征。这些模型需要一两天时间验证，然后才会在对冲基金投入生产。接着你就获得了更新后的因子载荷，这意味着你的风险敞口已经更新，可能超出风险边界，需要进行对冲操作。

I don't think so. But I can tell you, and that's actually beautiful, that if you looked a few years ago and you looked at the factor models typically used to be published with the new fundamental characteristics, They used to be published the first day of the month. Then typically they get validated, it takes them up one or two days, and then they get into production at hedge funds. And then you have updated loadings. So you have updated exposures and you are outside of your bound and you need to hedge.

假设在每月第三或第四天，由于模型检测到这些边界突破，对冲操作就不得不发生。有时候你会看到某些因子在月初出现大幅波动，这些其实就是对冲账户的操作痕迹。随着风险模型更新频率提高，以及现代对冲账户交易策略的精细化发展（现在不再是每月只对冲一次），这种效应已经消失了。

Let's say on the third or fourth day of the month, hedging happened, had to happen because the model generated these breaches. And sometimes you could see that certain factors had large movements at the beginning of the month. These were the hedge books. That effect has disappeared because the factor models, the risk models are updated more frequently and also the way the hedge books are traded nowadays it's a little bit more sophisticated. You don't hedge once a month.

在过去，人们确实每月只对冲一次，但现在不再如此。严格来说我不认为这属于拥挤交易，但从某种意义上看，确实是由因子模型引发的。可以说部分如此。现在我们回到阿尔法策略本身——问题不在风险模型，但我认为确实存在更严重的拥挤现象。

In the good old days people were hedging once a month not happening. I don't think it's really strictly speaking was crowding, but in a sense, yes, it was caused by the factor models. I would say partially yes. We go back to the alphas right now, it's not the risk models. However, I think there is more crowding.

不过我认为原因有所不同，这更像是个实证问题。只需观察行业现状：对冲行业已经 consolidation。虽然公司数量大致不变，管理规模也相近，但集中度完全不同。LCH的研究显示，过去三年对冲基金38%的盈亏来自三家公司——Citadel、Millennium和...（注：原文不完整）。自对冲基金行业诞生至今，前20家公司创造了行业总盈亏的19%，这充分说明行业集中度之高且仍在提升。

And the reason though, I think it's different and it's like an empirical reason. Just look around, the hedging industry has consolidated. In terms of number of firms, it's kind of the same AUM, concentration in terms of size, not the same at all. There was this LCH study that showed that 38% of the P and L in the past three years in hedge funds was generated by three firms, 38%, Citadel Millennium and the Inception of the hedge fund industry to date, the top 20 firms have generated 19% of the total P and L generated in industry ever. So this tells you how concentrated the industry is and is getting more concentrated.

另一方面，在供给端还出现两个新现象：一是被动投资占比从百分之十几激增到40%——这很惊人，40%的管理规模都是被动投资。在1987年的黄金时代，Jeremy Stein的论文指出48%的资金流来自零售端。

And then on the opposite side, like on the supply side, you have two more effects. One is the growth of passive, which ranges from the low teens to more like 40%. It's crazy. 40 of AUM is passive. In the good old days, in '87, there is a paper by Jeremy Stein that says 48% of flow was retail.

如今零售占比曾经是10%，现在可能15%，已大不如前。被动投资曾经是零，现在接近40%。这就形成一个小规模资金池，由极少数玩家进行投资。那么这个工具的最后一环是什么呢？

Nowadays it used to be 10, now it's maybe 15. It's a shadow of its former self. Passive used to be zero, it's almost 40%. So now you have a very small pool of capital where a very small number of players invests. And then what is the last leg of this tool?

2008年2月之后，来自卖方的基本面投资者常规供应基本枯竭，因为自营交易和资产管理业务缩减，银行资产负债表收缩。现在行业内出现了一种不健康的基金经理流动现象，他们从一个平台跳到另一个对冲基金。因此他们都彼此认识并相互交流。这导致公司间形成高度共识。这就是我对为何会出现拥挤交易、更多共识现象，以及拥挤因子表现更剧烈的纯经验性解释。

After 02/2008, the regular supply of fundamental investors that was coming from the sell side basically dried up because prop trading and asset management was reduced, the balance sheet of banks was reduced. Now you have an unholy turnover inside the industry of PMs that go from one platform to the other to another hedge fund. And so they all know each other and they talk to each other. And this generates a lot of consensus across firms. So this is my purely anecdotal explanation of why you do see crowding, you do see more consensus, you do see crowding as a factor behaving in more violent ways.

有时候这很棒。当大家都在成长型行业有相同想法时很棒。你们互相买入对方的头寸。这越来越像抢椅子游戏有时候。其中一

And sometimes it's great. It's great when everybody has the same ideas in a growing sector. You buy up each other's position. This is more and more resembling a little bit of a game of musical chairs sometimes. One of

我们这次对话尚未涉及的一个领域是投资组合构建。在学术研究甚至一些实践应用中，存在默认简单按alpha信号排序的做法。这是量化研究可以帮助投资经理通过优化等方法创建更风险高效组合的领域吗？

the areas we haven't touched upon yet in this conversation is portfolio construction. And in academic research and even in some practitioner implementations, there's this default to naively sorting on alpha signals. Is this an area where QR can help portfolio managers create more risk efficient portfolios through something like optimization?

构建投资组合的一种方法（这种方法没有太大问题，完全可以理解）是：我心中有一个特征指标，按这个特征对股票排序。比如说，我对所有总统候选人的言论进行文本分析，从他们的纲领、政策平台等可以了解他们偏爱哪些行业，又打算监管哪些行业。在大选前，假设我是一家卖方机构、银行。我会创建一个组合：做多特朗普偏好的前50家公司，做空后50家，对哈里斯的偏好也做同样处理。

One way to construct a portfolio, and there's nothing too wrong with it, it's perfectly understandable, is I have a characteristic in mind. I sort the stocks according to this characteristic. Let's say for example that I do a text analysis of all the utterances of presidential candidates, And I can get an idea from their programs, platforms, whatnot of what sectors they are favoring and what sectors they are instead wanting to regulate in the future. Before the election, I am say a sell side firm, a bank. I will produce a portfolio with long the top 50 firms that Trump favors and short the bottom 50, and I do the same for Harry's.

好的。这样我就通过排序创建了我的投资组合。然后你作为银行可以说：猜怎么着？我还可以帮你交易这些组合。这样大家都开心。

Okay. So I have created my portfolios from sorts. And then you can say, well, as a bank, guess what? I can also help you trade those. So we are all happy.

可以理解。但这是创建特朗普vs哈里斯组合的好方法吗？我会说不是。为什么不是？有多个原因。

Understandable. Is this a good way of producing a Trump versus Harris portfolio? I would say no. And why it isn't? Because of a number of things.

首先，你只使用了股票宇宙的一个子集。你可以有一个基数度量，比如从-1到1连续变化。为什么要把自己限制在，比如说40%的宇宙范围内？为什么要对数据做二分处理？统计学的一个常见智慧是：尽可能不要对数据进行二分处理。

First of all, you're only using a subset of the universe. You can have a measure that is cardinal measure, so it goes from continuously from minus one to one. Why restrict yourself to, let's say, 40% of the universe? Why dichotomize your data? One of the common wisdom in statistics is do not dichotomize data if you can help it.

第二个问题是，你给我的这个投资组合非常不干净。这个组合会有动量敞口、价值敞口、行业敞口。所以我希望通过使用因子模型来净化这个组合。第三个反对意见是，这个组合对公司的风险不够敏感。如果我做多，投资比如英伟达和IBM，它们的风险特征截然不同。

The second thing that's wrong with this is that you're giving me a very dirty portfolio. This portfolio will have momentum exposures, it will have value exposures, sector exposure. So I'd like to purify this portfolio by using a factor model. The third objection is that the portfolio is sort of insensitive to the risk of the firm. If I am on the long side and I'm investing in, let's say, Nvidia and IBM, they have massively different risk characteristics.

在它们内部，情况也一样。这简直是对人性的犯罪。基于排序构建的投资组合存在很多问题。而这正是经过深思熟虑的量化研究和组合构建过程能够提供帮助的地方之一。

Within them, the same. It's a crime against humanity. There are many things that portfolios from sorts do wrong. And that's one of the things where a thoughtful QR and portfolio construction process can help.

对我来说，投资组合构建中另一个仍有争议的领域是如何结合阿尔法信号的问题。我的意思是，大约在2015、2016年有很多关于这个的论文，核心问题是：是在投资组合中混合阿尔法信号更好，还是将信号整合在一起，形成一个统一的信号，然后基于此优化你的投资组合更好。所以，混合方法基本上是为每个阿尔法信号构建一个投资组合，然后合并这些组合；而整合方法则是将阿尔法分数结合在一起。我很好奇你在这个辩论中站在哪一边。

Another area of portfolio construction that to me is still an unsettled debate is the question of how to combine alpha signals. I mean, there were a lot of papers written about this in like 2015, 2016 with the big question of is it better to mix alpha signals in a portfolio or integrate the signals together to come up with a single unified signal from which to then optimize your portfolio. So a mixed approach basically builds a portfolio for each alpha signal and then combines those portfolios and then an integrated approach combines the alpha scores together. Curious where you come down in that debate.

我在Reddit上做过一个问答，有人问了同样的问题。这是一个如此美妙而深刻的问题，我无法完全公正地回答。所以，与其直接回答，我更想重新阐述这个问题，并补充一点我力所能及的微小贡献。但现实是，让人们理解这个问题的重要性比给出GAPI的立场更有价值。你可以用多种方式表述这些问题。

I have a Reddit AMA, and somebody asked the same question. It's such a beautiful and deep question that I cannot do it justice. So what I would like to do instead of answering the question, I'm going to restate the question and add a little bit of minimal contribution like I can give. But the reality is, it's better to understand for people how important this question is than to have GAPI's stake on it. You can formulate these questions in multiple ways.

所以，我认为你最初问的是，假设我有20位医疗保健领域的投资组合经理。我的意思是，这其实是一种异常情况。他们对同一个股票池有20套阿尔法信号。他们在训练自己的投资组合。而我作为对冲基金经理可能会说。

So one is I think what you're asking originally, is, let's say I have 20 portfolio managers in healthcare. I mean that's an aberration. They have 20 sets of alphas for the same universe. They're training their portfolios. And I could say, well, I am the hedge fund manager.

我不在乎。把你们的投资组合给我，比如说，我会把它们乘以二，这就是我的内部阿尔法捕获账簿。我直接复制它们。把你们的头寸给我。我不在乎你们怎么想。

I don't care. Give me the portfolios, and I'm going to, let's say, multiply them by two, and there is my internal alpha capture book. I'm just gonna replicate them. Give me your positions. I don't care how you think.

我会选择那种方式。另一种观点是，不，把你们的阿尔法信号给我，然后我会对这些阿尔法做一些神奇的处理。接着，我会基于这些阿尔法创建一个投资组合。正确的做法是什么？我可以告诉你，在一些理想化的玩具模型中，假设每个人都是不变性优化者，包括对冲基金和个体投资组合经理，再加上一些额外条件。

I'm gonna go that way. An alternative view is, no, give me your alphas, and then I'm going to do some magic on the alphas. And then I am going to create a portfolio based on these alphas. What's the right thing to do? I can tell you that under some toy, high idealized conditions, that everybody is an invariance optimizer, both the hedge fund and the individual portfolio managers, some additional conditions.

实际上，在我即将出版的新书中有一个章节专门讨论这个问题。你可以证明这些投资组合将会完全相同，这某种程度上令人安心，因为事实上，将投资组合而非信号进行组合是非常常见的做法，因为它更简单、更具商业可行性、更实用，而这告诉你，它验证了你正在做正确的事情。如果你对阿尔法做了各种神奇的处理，也不会得到更好的结果。不过在实践中，这是一个玩具模型。而且我认为可以放心地说，如果你获得了阿尔法，然后至少高效地进行交易，你不会比直接采用投资组合做得更差。

I have actually a section in my upcoming book on this. You can actually show that the portfolios are going to be exactly the same, which is kind of reassuring because as a matter of fact, the idea of combining portfolios instead of signals is very common, because it's much easier, it's much more commercial, practical, and this is telling you, it's validating you're doing the right thing. If you did all magic on alphas, you wouldn't get better. In practice though, it's a toy model. And in practice, I think it's safe to say that if you get the alphas, and then at the very least you trade them efficiently, you cannot do worse than just taking the portfolios.

你只能通过试错等方式来进行。这就是为什么内部因子实际上是一个如此困难的问题。这是一个问题。让我们稍微抽象一下。我有阿尔法，好吧，来自20位投资组合经理。

You just have to do it in trial and error and so on. That's why internal fact actually is such a difficult problem. That's one question. Let's abstract it a little bit. I have alphas, okay, from 20 portfolio managers.

现在这些阿尔法是真正的阿尔法。你知道有些人一年交易20次，有些人一年交易5次。这些不是相同的阿尔法。它们有不同的信息系数。一个来自优秀的投资经理，另一个有不同的广度。

And now these alphas are real alphas. You know somebody is trading 20 times a year, and somebody is trading five times a year. These are not the same alphas. And they have different information coefficient. One is from a great PM, one different breath.

它们具有不同的特征。想象一下，我只想将我认为具有高预测能力的一天期阿尔法与一些具有很好预测能力的一月期阿尔法结合起来。我如何将其组合成一个单一的阿尔法？我应该这样做吗？在同一时间跨度和跨多个时间跨度的阿尔法组合问题真的非常重要，因为它会影响你的投资组合——在没有交易成本的情况下，谁在乎呢？

They have different characteristics. Imagine that I just want to combine alphas that I think have a high predictive power one day out with some alphas that have a great predictive power one month out. How do I combine this into a single alpha? Should I do it or not? The problem of alpha combination at the same horizon and across multiple horizons is really, really important because then it affects your portfolio in the absence of transaction costs, who cares?

当你实际支付价格时，你就真的需要为你的阿尔法拥有正确的前瞻曲线。有一家叫做Jacob's Levy的公司，他们有一篇古老的论文，叫做“单一阿尔法法则”。他们基本上将所有他们的论点整合成每个资产的一个阿尔法。我认为这说明了这样做应该是正确的。这是一个非常困难的问题。

When you actually pay a price, then you want to really have the right forward curve for your alphas. There is a firm called Jacob's Levy and they have this old paper, the law of one alpha. They basically integrate all their thesis into one alpha per asset. I think it sends the reason that that should be the right thing to do. It's a very difficult problem.

我只知道，当我在2010年代中期研究这个问题时，辩论双方都有非常聪明的人。

All I know is that when I was doing research on this problem back in the mid two thousand tens, you had very high intellect people on both sides of the debate.

公平地说，还有另一篇论文，是高盛的一些人在2018年发表的，他们表明信号更好。但为什么他们显示信号更好？因为他们通过排序构建投资组合。如果你通过排序构建投资组合，所有赌注都失效了，但这首先就不是你想要的方式。而且他们只有多头投资组合。

So there is another paper, in fairness, there is a paper by some Goldman Sachs people from 2018, and they show that signals is better. But why do they show that signal is better? Because they do portfolios from sorts. If you do portfolios from sorts, all bets are off, but that's not the way you want to do it in the first place. And they had long only portfolios.

是的。我想我发给你过我写的一个简短研究报告，如果你通过排序构建投资组合，即使选择相同数量的股票，最终也会得到不同的集中度。所以这个问题有很多复杂性，是个非常有趣的问题。关于创建这个阿尔法远期曲线的想法，让我想到了单期优化与设计多期优化投资组合的概念。有很多关于多期优化的研究论文。

Yeah. I think I sent you a quick research note I had written where if you do portfolios from sorts, you actually end up with different concentrations if you're selecting the same number of stocks. And so there's a lot of complications to this question and it's a fascinating question. On this idea of creating this forward curve of alphas, it brings to mind for me this idea of single period optimization versus designing a portfolio with multi period optimization. And there's a lot of research papers around multi period optimization.

我发现在与从业者交谈时，我们最终都倾向于默认使用单期优化。从你的角度来看，你认为多期优化只是一种有趣的研究好奇心，还是认为人们实际上可以在实践中使用它？

I tend to find when talking to practitioners, we all just sort of end up defaulting to single period optimization. From your perspective, do you think multi period is just sort of this interesting research curiosity or do you think it's something that folks can actually use in practice?

我觉得随着对话的进行，我会逐渐降低外交障碍，变得越来越不外交，更加直言不讳地表达观点。让我们总结一下这个问题。首先，阿尔法具有期限结构。交易是历史依赖的。我可以今天交易一点，明天交易一点，等等，也可以一次性全部交易。

I think as the conversation goes on, I kind of lower my diplomatic barriers and I get more and more undiplomatic, more openly opinionated. Let's summarize the problem. Alphas have a term structure, number one. Trading is history depending. I could trade a little bit today and tomorrow and so on, I could trade all in one shot.

我通常不觉得这些论文很有用。有一些更好的方法近似，但在绝大多数情况下，例如，他们会优化终端状态。所以这更像是卖方公司的执行问题。但再次说明，即使对此来说也太理想化了。所以忘掉学术文献吧，因为与实践存在很大的脱节。

I don't find these papers usually very useful. There are some better approximations of the approach, but in the vast majority, for example, they will optimize for a terminal state. So it's more of an execution problem by a sell side firm. But again, it's even too idealized for that. So forget the academic literature, because there is a big disconnect with practice.

你基本上是在单期优化中调整参数，通过参数调优，实际上解决的本质上是一个多期优化问题。这类似于在用卡尔曼滤波器估计线性系统状态时，你不需要解决一个超级复杂的问题，而是对先前状态和当前观测值进行移动平均或加权平均。本着同样的精神（即使不是同样的数学形式），通过单期优化你可能正在做正确的事情。但并不总是如此。

You basically change your parameters in your single period, you tune your parameters, and you're solving essentially a multi period optimization. Not unlike when you estimate the state in a linear system with say carbon filter, you don't need to solve a super complicated problem. You take essentially a moving average or a weighted average of your previous state and your current observations. In the same spirit, if not the same mass, you could be doing the right thing by doing single period optimization. But not always.

所以我的总结观点是：目前来看，多期优化显然是正确的方法。但实践中也可能大多数场景最终都归结为解决单期优化问题。那样做也没问题。

So my view in summary is right now is that multi period optimization is obviously the right thing to do. It's also possible that in practice, most scenarios boil down to solving a single period optimization. You'll be fine that way.

我们对话中有一个没有明说但很重要的背景是，我们主要讨论的是股票。当我们谈论因子时，通常也是指股票。我很好奇你的看法：你认为这些模型、技术、观点，以及我们在量化研究角色、投资组合经理覆盖等方面讨论的所有内容，是否适用于其他资产类别？还是这些方法只适用于股票？

Part of the conversation here that's gone a little bit unspoken is that we're predominantly talking about equities. When we talk about factors, typically it's talking about equities. Curious as to your thoughts. Do you think these models, techniques, views, everything we've talked about in the quant research role, portfolio manager coverage, all that stuff extends to other asset classes, or is this something that is unique to equities?

我确实相信它在股票之外也有很好的应用。不仅仅是相信，我知道确实如此。在股票之外也存在采用因子模型的成功策略。我想补充几点：第一点可能是老生常谈，但你知道有时候你明知某些做法是错误的，却还是会去做。

I do believe that it has good applications outside of equities. Not only believe, I know that it does. There are successful strategies outside of equities that employ factor models. I would add a couple of things. The first one is elderly wisdom maybe, but you know sometimes you know that some things that you do are wrong but you do them anyway.

如果我对债券做因子模型，这显然是不对的。如果我做一个包含债券、期货、可能还有信用债和互换的不伦不类的组合，这在理论上真的不是你应该做的事。但为什么即使知道收益不是平稳的，你还是要这样做？因为你需要一个风险度量。跨资产类别时，唯一能共享的就是收益。

If I do a factor model on bonds, clearly it's the wrong thing to do. And if I do an unholy panel of bonds, futures, and maybe credit bonds and swaps, the thing should be on paper really not the thing you want to do. But why do you do it anyway even if you know that the returns are not stationary? You do that because you need a measure of risk. Across asset classes, the only thing that you can share is returns.

但我不进一步询问。这么说吧。因子的概念

But I don't inquire further. Let's put it this way. Concepts of factor

模型已经有五十多年的历史了。这些并不是真正的新想法。你认为未来研究的沃土在哪里？人们应该把时间花在什么地方？

models go back fifty plus years now. These aren't really new ideas. Where do you think the fertile ground is for research going forward? Where should people be spending their time?

有一个明显的答案，那就是我们拥有这些变革性技术，基础设施技术，以及有时以通用逼近器形式出现的数学技术。这项技术不是深度学习。这项技术是过度参数化的大规模模型，可以在未来运行在百万个GPU之类的东西上。目前最先进的是几万个。我们正在移植一项旧技术。

There is the obvious answer, which is that we have this transformative technologies, infrastructure technology, and sometimes mathematical technologies in the form of universal approximators. The technology is not deep learning. The technology is overparameterized, large scale models that can run on, in the future will be a million GPUs or something like that. Right now state of the art is a few tens of thousands. We are porting an old technology.

我的意思是，CUDA真的不是为这个设计的。一旦我们能够扩展它，使其能够承载能量负载和数量负载，我们并不确切知道这会带来什么，对吧？但这项技术正在到位。不尝试利用它来识别更好的风险和回报因子模型将是犯罪。怎么做？

I mean really CUDA is not designed for this. Once we can scale this so that it can carry the energy load and the number load, we don't know exactly what this will do, right? But the technology is being put in place. It would be criminal not to try to use it to identify better factor models for risk and for returns. How?

另一个非常有趣的领域是——人们从五十年前就开始讨论，但至今未能兑现其承诺——即基于代理的或试图建立投资者相互博弈的微观结构模型。因为即使从非常宏观、粗略的层面理解这些过程如何发生，也将帮助我们理解拥挤交易现象。这对投资很重要。我认为对监管机构也很重要。学术界有些论文写得不太好，因为我觉得学者们有时对现实关注不够。

The other area that's so interesting is, and people talk about it since also fifty years ago, and it has not delivered on its promises, is in general agent based or trying to have a microstructural model of how investors play against each other. Because understanding even to a very broad level, very rough level, how these things happen will help us understand crowding. So this is important for investing. I would say it's also important for regulators. There are papers written by academics that unfortunately are not very good because I think academics don't pay enough attention to reality sometimes.

总体而言，基于代理的拥挤建模是一个被严重低估的研究领域。顺便说一下，谈到古老的技术——神经网络比因子模型历史更悠久。它们已经经历了两轮炒作周期，但现实是真正发展不过从2006年至今。我的意思是，这项非常古老的技术其实真正发力也就是过去十年左右。原则上在1946年，麦卡洛克和皮茨如果拥有我们现在的全部数据和技术，可能七十年前就完成类似成果了。

In general, agent based modeling crowding, it's a vastly underexplored field. And by the way, talking about technologies that are old, neural networks are older than factor models. They have gone through two cycles of hype, but the reality is it's not even 2006 to present. I mean, it's really the past ten ish year for a very old techno I mean in principle in 1946, McCullough and Pitts had all our data and all our technologies. Maybe they would have done something like this already seventy years ago.

所以老东西并不差。

So old stuff is not bad.

本着越来越有主见甚至可能有点尖刻的态度，我给读者一个启发式建议：如果你看到书名是《X技术用Y语言应用于Z领域》，不要买那本书。从技术出发试图映射到实际问题的书，几乎从概率上就注定失败。你不会说'哦我掌握了机器学习，现在要把它应用到金融领域'。你不会写什么《Python在金融中的应用》之类的书。这种内容很快就会过时。

In the spirit of I'm getting more and more opinionated and maybe a little snarky, but as a heuristic for the readers out there, If you see a book titled technology x in language y applied to field z, don't buy that book. A book that starts from a technology and tries to map it to a real problem is almost by definition probabilistically doomed to failure. You don't say, oh I have machine learning and now I'm applying it to finance. You don't write a book about Python in finance or whatever. It will get old very quickly.

任何人都可以读一本强化学习的书，然后肤浅地将其改造应用于金融领域。但这并不是推动一个领域进步的方式。理想情况下，我希望写一本只使用50年以上数学知识的书。在某种程度上确实如此，但也不完全是，因为我使用了Rademacher复杂度——这是一个机器学习概念，在机器学习领域有二十年历史，但在函数分析领域其实已有四十年。所以我确实使用了一些高级工具。

Everybody can read a book on reinforcement learning and retool it superficially for finance. That's not how you create progress in a field. Ideally, I wanted to write a book that used only 50 year old plus math. To some extent, yes, but I'm also not because I'm using Rademacher Complexity, which is a machine learning concept, twenty year old in machine learning, but it's really 40 years old in function analysis. So I do actually use a little bit of advanced tools.

我会深感荣幸。

I would be delighted.

也许这个名字不够抓人眼球。

Maybe it's not catchy enough.

谢谢你。

Thank you.