Prospects of AI Integration into MBS Analytical Environment & Challenges That Can Be Mitigated Only By a Fantastic "Le Carignan”.

In the picturesque village of Montpeyroux, Domaine d’Aupilhac stands as a beacon of tradition and innovation in the Languedoc region. This estate, which has been nurtured by three generations of the Fadat family, is deeply rooted in history, with vineyards that trace their origins back to Roman times. However, it was in 1989 that Sylvain Fadat officially registered the domaine as a vigneron indépendant, marking a new chapter in its legacy.

This is a fascinating story, and we will explore it right after we address the main topic of this publication: practical AI implementation within the context of fixed income analytics and portfolio management.

Challenges & Rewards of AI Implementation In the MBS Universe

* Before we begin, we must warn our readers that this publication is not directly related to the present relative value landscape of the MBS market, or even Fixed Income Market in general. Yet, the issue we attempt to address in this brief publication, will be critical to the process of relative value discovery going forward.

* In fact, the overwhelming majority of our recent visits have involved a discussion with investors regarding the role that "AI", in all of its manifestations, will be playing in the ancient art of Portfolio Management. These conversations go especially deep during meal sharing rituals with various players in the MBS markets. Thus, (even at a risk of overgeneralization), we can aggregate concerns, voiced by the portfolio managers, analysts, and C-suite members of countless organizations, we have spoken to:

"While we are currently in the beginning stages of an AI implementation, we are concerned that we are entering too late, and our "larger" competitors (aka Big boys) are already light years ahead of us, forcing us to play a never-ending catch-up game with near-zero probability of success. And yet, not entering into the realm of AI is simply not an option. Additionally, we are unsure on how to prioritize our investments in AI, since those funds can be deployed in many directions and areas (such as operations, risk management, client interaction, cybersecurity etc.) to boost our efficiency and enhance our modeling capabilities. Lastly, we are worried about risks that are well-known (privacy violations, cybersecurity, IP etc.) and more importantly those of which we have limited understanding".

* Today, our conversation will indeed be focused on the risks, emphasizing those, that are either more "manageable", or less obvious (and therefore, possibly, more dangerous). And since the MBS analytics and the Investment Lifecycle Optimization (ILCO) are our areas of expertise, we will focus on them at the expense of other possible AI deployment projects (such as operations, marketing, cybersecurity or IP), of which we know far less.

* Before we address these potential risks (and opportunities), allow us to express our deeply-held view that an AI integration is a long-haul game. We often see institutions that are attempting to play a "catching up with Joneses" game with AI adoption, while having no clear prioritization of goals that this (very costly) implementation will help them achieve. It seems to us that a mere displeasure of the Board of Directors regarding the lack of activity on the AI front does not merit an aggressive AI adoption.

* The overriding concerns that many PMs have expressed to us is that someone else will build a superior AI "dragon", putting them at a disadvantage. If you are in this camp, please allow us to put your mind at ease - someone has already built an AI structure that is very likely to be way superior to yours. Yet, looking at the story as a simple race to find the perfect algorithm may produce some counterintuitive results.

* Let's imagine that Goldman or BoA or Chase (or all three of them) build the most amazing algorithm that is capable of profitably identifying Value in the MBS market before we can even turn our machines on. This Ultimate AI gets better every day, as it learns to be even more powerful, thus leaving all the rest of us in the dust. As the rewards become asymmetrically massive, so will the odds of magnificently escalating investments by competitors to reverse-engineer this system.

* And what if this system is so good that if simply can not be replicated? Well, then, it can be misdirected. The system of this quality and magnitude will have to rely on many inputs, both internal, and those residing in the public domain. And if other market participants can not replicate this "magical" model, they can (and will) poison the "data well", trolling the inputs in order to destabilize the outputs. Thus, the fears of someone building an analytical AI engine that is vastly superior in, let's say, asset allocation or security selection game is greatly overblown.

* Yet, we must admit that several institutions are already in possession of very sophisticated AI systems in multiple "planes" including asset management (going way beyond mere Algo trading). Yet, their power rests as much with the input quality, as it does with their AI algorithmic advancements. And here is where smaller players might be at a considerable disadvantage over the large ones.

Buy Historical Data - they don't make it anymore.

Collect New Data - They Make too much of it

* Since the dawn of the Internet, we have been taking the exponential rise in data availability for granted, with firms practically giving away giant sets of proprietary data, not really knowing what to do with it, while individuals traded their social security numbers for pictures of cute cats.

* This era of "data promiscuities" is now officially over. Yet, many investors are still comfortable in relaying on data providers (such as Bloomberg, Yield Book and others), who offer them a neat-looking time series to be fed into their in-house regression models and other analytical constructs.

* Few investors pause to ask themselves if this data will be available to them tomorrow, or next year. Yet the noose is tightening, as the analytical providers are charging increasingly skyrocketing fees for essentially the same functionality. Since an increased number of data inputs can be obtained through only one or two providers, investors will have no choice but to continue paying those high subscription fees in order to have access to the data.

* Without the data, no AI build-up will be possible, thus the first step in laying the groundwork for AI integration is building up a dataset (ideally in-house) so the future success or failure of one's analytical constructs is not in the hands of the data gatekeepers.

* The very power of AI (such as neural networks, which are often used in the analytical settings) is an ability to detect unusual relationships between seemingly unrelated data patterns - something that even a powerful regression is simply unable to do. Yet, in order to establish these relationships between a great many data inputs, one must first be in possession of these treasures, and make sure that they are real (as opposed to being fake, or simply incorrect).

Quick Clarification: What AI are We Talking About?

Let's do it really fast: we like to think of the AI plane as consisting of two broad strands:

(1) data-driven machine learning systems;

(2) rule-based approaches such as deterministic chatbots.

* Machine learning contains traditional statistical models and artificial neural networks, which aim to replicate the learning process of the human brain. These models can capture non-linear properties of data, and apply previously gained knowledge to new problems.

* Recently, the capabilities of artificial neural networks have been significantly boosted by increasing their complexity by training them on a vast amount of data. The rise of this new class of models, generally called foundation models, was mainly enabled by the decreasing cost and increasing efficiency of the computational power.

* These models are “trained” in a self-supervised manner on a vast amount of both structured (e.g. tables) and unstructured (images, sound, text) raw data, with only minimal human intervention. In the pre-training phase, the model learns the fundamental structure (“ground truth”) of the data in a generic way, covering aspects like the use of human language, recognition of objects and images, and numerical input. In fact, Generative AI models can make good use of the knowledge of foundation models.

* A key feature of generative AI is its ability to produce unique output, which shares some properties of the input data, but differs in others (generative capabilities). Most current generative AI models are based on text (large language models, or LLMs), thus eliminating the need for proficient coding skills to modify or use them.

* The performance of foundation models can be enhanced by providing additional training on task-related data (fine-tuning), or by embedding additional tools like search engines. Foundation models and the generative AI based on such models add new aspects to consider, when assessing implications for the financial system. Therefore, this discussion focuses explicitly on these models.

Back to the Risk of AI Implementation: Risk of External Interference

* Aside from obvious human interference of "bad actors" on the inside of a given financial institution, here, we are talking about a direct threat of external interference. Clearly, a cyber attack immediately comes to mind. Yet, this is too obvious of a risk for the purposes of our discussion. One of the main reasons financial organization are investing billions into analytical AI systems is to establish relationships that go far beyond the well-understood groups such as "rates down, prepayments up, OAS on ITM MBS widens".

* An advanced regression model can easily capture these. The very reason to seek the assistance of neural networks is to bring data from the periphery of the traditional analytical data sandbox, utilizing various measures of sentiment (political, economic, social, etc.), while incorporating data that does not "relate" to our dependent variable set in a traditional linear format, yet exerts an undoubtable influence on it.

* Note that the overwhelming majority of this new input data is likely to come from the "outside" of the organization, either from the public domain or from a third party. Even if a historical set of data, on which a NN model is being trained is "clean", the ever-growing "voice" of bots in the public domain discourse (including in malicious data manufacturing), will pose an increasing risk.

* This risk will intensify exponentially with the size and the level of success of a specific AI system, which will make those data fakes purposefully targeted to be incorporated in the input of this "more advanced" AI model in order to derail its productivity.

* Consider the fact that during the election campaign of 2024, the estimates of non-human participation in the on-line discourse are as high as 48%, which means no matter how ones attempts to incorporate political sentiment of various segments of the population as a model input, the voice of non-human players will pollute this data set. And, needless to say, these non-human participants are here to fulfill someone's agenda.

Risk of Systemic Amplifiers

* There are two systemic amplifiers, through which the implications of AI for a single firm could become systemic.

The first amplifier is a technological penetration. If an AI is widely adopted across different financial entities for an increasing number of processes and applications, more areas of the financial system will be affected by the challenges and opportunities associated with this AI structure. The second is the risk of fathomability.

Fathomability Risk

* AI models and, more importantly interaction between them, are getting increasingly more complex, so that even their creators can not account fully for the logic behind their results and predictions. Given this unfathomability, it is more difficult to detect model hallucinations (presenting false or misleading information as facts). This problem is not likely to be resolved in the near future, requiring financial institutions to set up a rather complex result verification process, which, in turn, is likely to depend on use of an equally complex AI system.

* Conceptually, AI can be thought of as a filter, through which information is gathered, analyzed and assessed. The interpretation of information may become more uniform, if increasingly similar models with the same embedded challenges and biases are widely used to understand financial market dynamics. As a result, AI may make market participants’ conclusions systematically biased, leading to distorted asset prices, increased correlation, herding behavior, or asset bubbles.

* Should many institutions use AI for asset allocation decisions and rely only on a few AI providers, for example, supply and demand for financial assets may be distorted systematically, triggering costly adjustments in markets that harm their resilience. Similarly, extensive use of AI by retail investors may result in large and similar shifts in retail trading patterns, which would increase volatility in market sentiment, trading volumes and prices.

In summary, we do not buy the argument that slower (and more measured) adoption of AI is any more risky than a hasty one (which represents a large majority of the cases that we are encountering on a daily basis). Like any information technology, AI will increase concentration risk, thus contributing to a further shift in market power. Yet, AI implementation is a game that has to be steady and purposeful, with a focus on the entire market architecture, including privacy, IP, security etc. Any attempt to trade off these foundational values for speed are likely to be very detrimental.

2021 Vin de France “Le Carignan”

Domaine d’Aupilhac

?

The Historic Aupilhac Terroir?

The heart of Domaine d’Aupilhac is the renowned eighteen-hectare lieu-dit known as Aupilhac. Nestled beneath the ruins of Montpeyroux’s chateau, these terraced vineyards are a mix of limestone scree and marl, perfectly positioned to soak up the southwest sun. This unique combination of soil and exposure results in wines with dense, chewy textures and unmistakable Mediterranean character.

But Aupilhac’s story doesn’t end there. Sylvain ventured into the challenging terrain of Cocalières, an ancient volcanic amphitheater filled with marine fossils and limestone boulders. Over years of grueling labor, he cleared the land to plant vineyards that now sit at 350 meters above sea level. These high-elevation vines, with their cooler northern exposure, yield wines of incredible purity, finesse, and striking minerality.

A Commitment to Sustainable Practices

Sylvain’s work in the vineyard reflects his deep respect for nature and commitment to sustainability. In 2014, after years of organic certification, Domaine d’Aupilhac achieved biodynamic certification—a testament to Sylvain’s belief in creating a harmonious ecosystem. His approach to viticulture involves regular plowing to force vine roots deep into the soil, seeking cooler, more humid layers that protect against drought and intense summer heat.?

This philosophy stems from Sylvain’s conviction: “We believe that work in the vineyards has far more influence on a wine's quality than what we do in the cellar.” His dedication to organic and biodynamic practices not only ensures the health of the vines but also enhances the authenticity and expression of the terroir.

?Masterful Winemaking: From Vineyard to Bottle

?In the cellar, Sylvain strikes a delicate balance between tradition and craftsmanship. His wines are known for their remarkable equilibrium—combining ripe fruit with silky tannins, and power with elegance. These qualities make Domaine d’Aupilhac’s offerings equally captivating whether enjoyed young or aged. Over time, the wines develop a complexity that rivals bottles sold for many times their price.

?One of Sylvain’s standout creations is his Carignan, which he calls “a symbol of the essence of the region and its history.” Crafted from old vines with low yields in poor soils, this wine is a smoky, deep powerhouse capable of aging for decades. It challenges perceptions of the Languedoc, showcasing the incredible potential of this undervalued region.

?A Family Legacy

The Fadat family’s connection to Aupilhac runs deep, spanning generations of dedication to the land. Sylvain’s work continues this legacy, honoring the efforts of those who came before him while pushing boundaries to elevate the Languedoc. His vision for the domaine has transformed Aupilhac into a shining example of what can be achieved through hard work, sustainability, and respect for tradition.

?Domaine d’Aupilhac Today

Every wine from Domaine d’Aupilhac tells a story of its unique terroir and the meticulous care with which it was crafted. Whether it’s the bold, structured reds from Aupilhac’s limestone terraces or the refined, mineral-driven wines from Cocalières, each bottle reflects Sylvain’s mission to valorize the Languedoc’s great terroirs and traditional grape varieties.?

At Domaine d’Aupilhac, history and innovation come together in a celebration of the Languedoc’s rich heritage and boundless potential. With wines that captivate both heart and palate, Sylvain Fadat continues to shape the region’s future while honoring its storied past.

Thank you for thinking & drinking with me; Slava Ukraine!

Sincerely Yours,

Kirill A Krylov, CFA, PhD