Strait of Hormuz: Causal Models for Rare Events | Alexander Denev S2E11 | CausalBanditsPodcast.com

*How do you forecast an event that has never happened before?*

How do you forecast an event that has never happened before?

The recent closure and reopening of the Strait of Hormuz are unique events. For events like these, traditional risk models lose their statistical basis: repetition. Alexander Denev returns to the podcast to show how causal models (Bayesian networks) let us reason about rare events despite this limitation.

In this episode, we cover:

- Why value-at-risk and other correlation-based models break exactly when you need them most
- How a causal structure can "hold in time"
- Building scenarios with LLMs - benefits, drawbacks, and lessons learned
- Historical analogy as a modeling tool: Bosphorus, Hormuz, and more
- A three-way robustness test for any Bayesian network
- How the model's call held up: a ceasefire, a still-closed strait, and lasting infrastructure damage keeping oil elevated

"History doesn't repeat itself, but it rhymes."

------------------------------------------------------------------------------------------------------

Video version available on the Youtube: https://youtu.be/FzKy2ws-7qs
Recorded on May 29, 2026 in London, UK.

------------------------------------------------------------------------------------------------------

*About The Guest*
Alexander Denev works at the intersection of quantitative finance, causality, and AI. He's the CEO of Turnleaf Analytics and the author of two books on applying Bayesian networks and probabilistic graphical models to finance and scenario analysis.

Connect with Alexander:
- Alexander on LinkedIn: https://www.linkedin.com/in/alexander-denev-66a25824/
- Alexander's web page: https://turnleafanalytics.com/

*About The Host*
Aleksander (Alex) Molak is an independent machine learning researcher, educator, entrepreneur and a best-selling author in the area of causality (https://amzn.to/3QhsRz4 ).

Connect with Alex:
- Alex on the Internet: https://bit.ly/aleksander-molak

*Links*
Web

- Alexander's LinkedIn post, Bayesian-network scenario for the Strait of Hormuz / Israel-Iran-US conflict: https://www.linkedin.com/posts/alexander-denev-66a25824_when-modelling-the-impact-of-events-that-share-7442892381668048896-JDs5/
- Risk.net article, "Iran confusion makes the case for causal modelling": https://www.risk.net/our-take/7963361/iran-confusion-makes-the-case-for-causal-modelling

Books

- Rebonato, R. & Denev, A. - Portfolio Management under Stress: A Bayesian-Net Approach to Coherent Asset Allocation (https://amzn.to/3vE6Jc1)
- López de Prado, M. - Advances in Financial Machine Learning (https://amzn.to/3PXD8kH)
- Molak, A. - Causal Inference and Discovery in Python (https://amzn.to/3VVK4m3)
- Denev, A. - Probabilistic Graphical Models: A New Way of Thinking in Financial Modelling (https://amzn.to/3VQeLJm)
- Pearl, J. & Mackenzie, D. - The Book of Why (recommended entry point) (https://amzn.to/4e0ATrZ)
- Pearl, J. - Causality: Models, Reasoning and Inference (for advanced readers) (https://amzn.to/49zBKf5)
- Rebonato, R. - Coherent Stress Testing: A Bayesian Approach to the Analysis of Financial Stress (https://amzn.to/3RC411e)

*Perks & resources*
🚀 Join FREE Causal Python Weekly Newsletter: https://causalpython.io
📽️ FREE Online Course on Causality: https://causalsecrets.com/
📕 My Book on Causality: https://amzn.to/3SKRXIw
🔥 Causal Bandits Community Beta Wait List: https://causalbandits.com/
🎙️ Get notifications about new episodes: https://causalbanditspodcast.com

*Let's connect!*
👉🏼 Linkedin: https://www.linkedin.com/in/aleksandermolak/
👉🏼 Bluesky: https://alxndrmlk.bsky.social
👉🏼 Tiktok: https://www.tiktok.com/@alex.molak

*Business*
👉🏼 Consulting and Causal AI Training For Your Team: hello@causalpython.io

*Podcast Playlist*
https://www.youtube.com/playlist?list=PLhKKv6iMja4p5FbJIgzTOE67E1M6c8lnB

*Causal Bandits Team*
Project Coordinator: Taiba Malik (https://www.instagram.com/taibasplay/)
Video and Audio Editing: Navneet Sharma

#machinelearning #causalai #causalinference #causality

Transcript

CAUSAL BANDITS PODCAST
Episode S2E12 (CB032)

Title:    Strait of Hormuz: Causal Models for Rare Events
Guest:    Alexander Denev (CEO, Turnleaf Analytics)
Host:     Alex Molak
Date:     2026-05-29
Location: London, UK
Links:    https://causalbanditspodcast.com  |  https://turnleafanalytics.com

----------------------------------------------------------------------
TRANSCRIPT
Speaker labels are inferred from context. A few uncertain turns are
marked [speaker?]. Lightly edited for readability (filler words removed,
obvious transcription errors corrected); wording is otherwise faithful.
----------------------------------------------------------------------

[00:35] Alex Molak: Welcome back to the podcast, Alexander.

[00:37] Alexander Denev: Thank you for having me back. It's a great pleasure.

[00:40] Alex Molak: A few weeks back, you published a LinkedIn post about the closure of the Strait of Hormuz, related to the conflict between the US, Israel, and Iran. In that post, you presented an idea: in cases like this, rare events that we might observe only once in an extended period of time, it's very difficult to use any traditional statistical modeling technique to reduce our uncertainty about the future trajectories of events. Why is that, and how can we deal with it?

[01:27] Alexander Denev: That's a very good question. I already talked on your podcast, giving a general framework which I believe is very suitable for situations like that. Financial markets, which are my focus, are not a closed system. They're continuously hit by exogenous shocks, and most of them come all of a sudden. They're unpredictable: the so-called black swans. Some of them are predictable, like gray swans: elections, referendums that are announced in advance. But what is common across all these shocks is that they're unique. And if they're unique, it means we don't have a statistical repetition of the event. So what is happening now in the world with the conflict in the Middle East is something unique. We don't have any reliable episodes in the past from which we can do statistical inference and really understand how this conflict is going to unfold and end. So we need a different type of modeling.

[02:47] Alexander Denev: Where would you look to understand what is going to happen, even if it's an event happening for the first time? Are there any stable transmission mechanisms and patterns that can be used to make a prediction of how this conflict is going to unfold? It's a difficult question, and I believe a Bayesian network framework, which is firmly grounded in causality, can help. This is essentially my post that you noticed: a Bayesian network built on how this scenario was going to unfold back when I published it, around eight weeks ago. What are the possible transmission mechanisms, and what would be the impact?

[03:37] Alexander Denev: It's important to say that I worked on this scenario many years ago. The first time was in 2012, so about 14 years ago, and I published it, so it's in the public space: what a potential conflict between Iran, Israel, and the US would look like. The curious thing is that it hasn't changed much. I sat down with an economist back in those days, and we used the Bayesian network methodology to come up with these scenarios, with transmission mechanisms and economic impacts, and things have stayed the same. Even though this event never happened in the past, the certain background structure that would be triggered if it happened has stayed the same for this period. So you might have shocks, but the way shocks propagate depends very much on the underlying transmission mechanisms, on the underlying causal structure that has been stable over time.

[04:53] Alex Molak: The model you build returns certain results, and I'd like to discuss those with you as well. But before we move to the results, I wanted to ask about the way you built this model this time. You mentioned that last time you built a similar model, I think 14 years ago, and that required a significant amount of work from you: talking to experts, to economists, to understand the structural connections between the building blocks, if I understand you correctly. So that gives you a background to understand, today, how to rebuild this model. You already have the building blocks ready, mentally speaking. Right?

[05:43] Alexander Denev: That's correct. I'm not an economist, and even if you're an economist, you specialize in a certain area. Maybe you monitor the United States, maybe Latin America. There is no single expert on everything. At the time, when I used Bayesian networks to create these scenarios, of course I sat down with an expert in the area. We designed together what the potential variables were that could be triggered, what the transmission channels were, and the different probabilities on the links between events that could materialize. I also worked, for example, on a scenario for the Scottish referendum. In that case I sat down with a UK economist. We spent around one month to come up with a scenario. Of course, the Scottish referendum didn't happen, so we can't back test that scenario, but it could happen again in the future, because it's still in the minds of people in Scotland. So that could be a repetition.

[06:47] Alexander Denev: But each scenario requires deep expertise to understand the stable patterns and mechanisms that could come back into play if the scenario materializes, and it is difficult for us to find the right experts. It's difficult to involve more than one expert, especially in global scenarios where you require a multitude of experts. In this case you need an expert on Iran, on the Middle East, on the US. You also need an expert on oil markets and on shipping. So it was very difficult to create the scenarios, and each one would require a room full of experts, which is very difficult to organize logistically. And markets move so quickly nowadays that by the time you finish designing your scenarios, the markets have already moved. So it was very difficult, at the time, to devise it to be quick.

[07:55] Alexander Denev: Now things have changed. As you know, I've written two books on the subject, but they remained confined to the theoretical toolset. They never saw the light of day, because the methodology was impractical given the constraints I just mentioned. But things have changed recently. Since ChatGPT was released a few years ago, we've seen a lot of advances on that front. We've seen different versions of ChatGPT; we have Claude; we have Grok; we have a variety of tools that are becoming better and better. We practically have access to a multitude of experts at the same time, instantaneously, not only in the way the LLMs are structured to understand your query, but also in the way they go and retrieve information.

[09:04] Alexander Denev: There is a lot of information in the public space, with different experts speaking about Iran, the US, oil prices, gold, and so on. This information can be collected quite promptly, thanks to these LLMs, with a series of prompts, which was impossible at the time because the technology didn't exist. Now, having a multitude of experts is equivalent to having a series of the right prompts. And it's very quick. You don't need to do whiteboarding sessions with experts for several days in a row, where the experts' biases can resurface and bias the scenarios. Now you have access to information almost instantaneously.

[09:56] Alexander Denev: Moreover, these LLMs do understand Bayesian networks. I fed my books into Claude in these specific examples, and it read and understood the background of the methodology and how it should be applied to economics and finance. It already had all the theoretical constructs of Bayesian networks embedded in its knowledge. So it was very easy for me to prompt Claude and say, "Claude, this is the methodology, please read it." It took very few minutes to read all my life's work, and then: "Given the information we have in the public space today, please construct a Bayesian network around the Iranian scenario." Of course, the first output you receive is not perfect, it's not the final one. You need to perform some sensitivity tests, you have to reprompt sometimes, you have to inspect it and make sense of the scenario. So it takes iterations to build it. But going from sitting a week with experts to having the right series of prompts and doing this in 15 to 30 minutes, it's a huge leap.

[11:19] Alexander Denev: So that methodology, which was essentially non-scalable, not applicable in the financial sector to the required standard of speed and promptness, is now becoming scalable, very quick, and very reactive. We're seeing this quantum leap in the implementability of this methodology based on causality, applied to economics, finance, and geopolitics in general, because finance is mostly driven by geopolitics nowadays. Having all these factors put together is becoming very easy, and this methodology, after years of writing about it only theoretically, is now becoming practicable.

[12:06] Alex Molak: You told us about the bright side of working with systems like Claude for developing the structural models, and that it's much faster than it used to be. What challenges, or let's say not red lights but orange lights, have you encountered in working with these models? You mentioned reprompting. So what were the challenges, and what helps you address them?

[12:40] Alexander Denev: The risk, and everybody knows and speaks about this, is hallucinations. They do hallucinate, and the risk is being complacent about the output of Claude, or Grok, or ChatGPT, and taking everything at face value. You should always cross-check, because the negative side is that it can stimulate human laziness. The LLM returns some very nice Bayesian networks, a very nice scenario; at face value everything looks all right, and you become complacent and say, "This is it." But hallucination is a problem and should be tackled. This is why I said one prompt is not enough. It returned the Bayesian network in one prompt, but I needed to reprompt at least 10 more times and perform sensitivity analysis, challenging some of the assumptions, and that was just for the sake of publishing my LinkedIn post. Of course, if you're an institution, you want something rigorous, so you have to spend a little more time on that and not be complacent. You should always challenge the results.

[14:03] Alexander Denev: In the book we present different methodologies. Even if you build a purely expert-driven scenario, you have to do some checks, because experts can also be biased, they can be tired at the end of the day, they can give very inaccurate numbers that they haven't thought deeply about. So even with experts you have to perform certain checks. In the first book, "Portfolio Management under Stress," we suggest a recipe for how to make sure the results make sense, how to perturb the network of the scenario to inspect whether the results still hold. There's a recipe to follow to make this robust, and that recipe should be followed in the case of LLMs too, not just experts. The key here is: check, check, and triple-check. This takes time, but again, if you compare one week of experts in a room arguing over pizzas and sandwiches to something much more compressed in time, the advantages are clear.

[15:19] Alex Molak: You mentioned checks and sensitivity analysis. For someone interested in building a Bayesian network for themselves, in finance or outside it, what would be two or three of the most useful, strongest checks for a structure like this?

[15:42] Alexander Denev: I mentioned sensitivity analysis because this is a probabilistic model, not a deterministic one. You start from a root node, which in this case is the current war, and you go through different transmission channels: it can be the Strait of Hormuz, an attack on American bases in the region, an escalation to a broader regional conflict. These are events that depend on each other. They're linked in a cause-and-effect framework through the Bayesian framework, but probabilistically, not deterministically. If the root node materializes, which is the trigger for the scenario, certain things can happen with a certain probability. And when experts assign probabilities, they tend to underestimate or overestimate low or high probabilities. If you're close to zero or one, the human mind gets confused. If you say more than 90%, for the human mind that's 100%. So experts tend to be very confident at the upper or lower bound of the probability spectrum.

[17:12] Alexander Denev: So the first thing is really to perturb these probability assumptions, and now it's easy with Claude. You inspect your results again, and the results are probability distributions. Let's say you want to know the impact on a portfolio of US equities or US treasuries. As a result of this perturbation, you don't get something exactly precise, but something more spread out. This is one of the sensitivity tests: that your results don't vary too much. You have to make sure of the robustness of your network with regard to the changes you make. And when we devise these networks, you have different nodes, different variables that can play out in a conflict scenario, but this is not a closed set. You can always expand it, add additional transmission channels. So you can think of two or three more and see if the results of the network change if you add or trim branches.

[18:28] Alexander Denev: So the most important step is to perturb the variables you've put in your scenario. Second, the structure: how they're linked. Ideally you should be able to remove and add different links. And the third is the probability you assign to each relationship. So it's a three-way robustness test, and that's the most important thing to do. And of course, since all these models have access to different sources, you can direct the LLM to take information from different sources. Let's say it extracted information from a certain news website, or from Polymarket. When Claude built the network, it had access to Polymarket to extract oil prices in the future. But you can swap data sources: "Okay, this is Polymarket's point of view. What if I use information from the Energy Information Administration in the US? How do my results vary?" So you can also perturb with respect to data sources.

[19:47] Alexander Denev: The remarkable thing about these objects is that they're not very sensitive, because if you think about how an event unfolds, the transmission channels robustify each other, they can cancel each other out. So a single perturbation in a single node, in the assumptions you made, can be tamed and could have a very small impact on the final results. In the end, if you think of a causal model with, say, 10 Boolean variables, it creates a joint probability table with 1,024 different combinations, according to the different events happening or not. And when you do sensitivity analysis, you discover that usually there are around 10 probabilities, out of all the ones you've assigned, that really matter. So it's important to understand what the biggest amplifiers or suppressors are in the scenario. There are many tools, but the most important are tools to assess the robustness of the claims you're making.

[21:22] Alex Molak: In marketing, sometimes what we do is try to back test the model. Maybe we have a marketing mix model, and we look at past events and check whether the current model we've built, if we feed it those past shocks, produces outcomes that resemble what really happened back then. Is that something you find useful as well, for validating the structural assumptions, the distributional assumptions?

[21:54] Alexander Denev: When we speak about back testing in finance, back testing a trading strategy, you're still speaking about statistical repetitions.

[22:03] Alex Molak: Yeah.

[22:03] Alexander Denev: Which means you might have a very simple linear regression model, trying to predict equity markets based on a few factors. You can roll your model back into the past and start a walk-forward back test from 10 years ago, and see how your model would have behaved out of sample, with all the caveats around overfitting. Here we don't have that luxury, in the sense that Brexit happened only once. We don't have different instances of Brexit in the past to go back to. So all this statistical repetition, rolling forward your training window, predicting something that will happen one day or one month ahead, is not possible. These are one-off events. But instead of back testing, you can use historical analogy when building these models. You don't use historical episodes to back test your model, but to build your model.

[23:13] Alexander Denev: [speaker?] It's very much like transfer learning in AI: you use similar situations to calibrate your current model, while being conscious that you're building something for an event that cannot be back tested because it never happened. When it comes to the Iran scenario, this very much resembles the situation during the First World War, when Turkey decided to close the Bosphorus. All the wheat exports, an essential commodity coming from Russia, were blocked, so food prices spiked. People started fearing inflation here in London. Great Britain was the superpower at the time, so they were scared about inflation, like we are scared of inflation with what's happening today. Then several plans were discussed, among them boots on the ground, which is also being discussed in the current situation. At that point in time, during the First World War, that attempt failed. But there are similar escalation paths: blockade of a commodity, inflation, intervention from the sea to attack the fortifications blocking the Bosphorus. It didn't succeed, then boots on the ground. So things that happened more than 100 years ago are being discussed today, and this also helps in building these scenarios and calibrating the outcomes: how successful a ground operation could be, the nature of the terrain in Iran compared to Turkey, the level of technology now versus then. These have a historical analogy but not a historical identity. It was a different time, different countries, but it's all a kind of transfer of learning: you extract lessons you can use to build the scenario as of today. History does not repeat itself, but it rhymes, as they said.

[25:35] Alex Molak: Correct, correct. The wise man said it once.

[25:39] Alex Molak: After your post, there was an article published in Risk.net about your model, about your post. Were you contacted by other people interested in this approach?

[25:51] Alexander Denev: Yes. I had a conversation last week with a big sovereign wealth fund who were interested in applying this methodology. Of course, my current focus is on a different type of approach: I run a forecasting company that uses different approaches. So I'm not actively engaged in developing this further, but it's open ground, given all my previous work, the work of Riccardo Rebonato, and the recent work of Marcos López de Prado. So I've opened a possibility for other people to explore these kinds of scenarios more in practice. And that article reminds me of another useful feature of causal graphical models: they're extremely good objects for putting causal narratives into a mathematically rigorous framework, and they also help expand the human mind in managing very complex scenarios. If you think about it, we are causal machines: we observe the world and extract cause-and-effect relationships. But the human mind is very limited. In my mind I can analyze three or four causes, maybe see the first-round effect, rarely the second. The graphical causal framework of a Bayesian network lets you put this on paper in a mathematically rigorous form.

[27:25] Alexander Denev: So you can encode many more possibilities in your scenario. As in the example I mentioned, if you have a scenario involving 10 variables, transmission channels, and impacts on different asset classes, even a simple Boolean network, yes or no, this event happens or doesn't happen, encodes 1,024 different possibilities. It's a very rich framework. I cannot keep 1,024 different scenarios in my mind. When you analyze the output of the network, the different combinations of events and their probabilities, you can come up with something where you say, "Wow, I never thought about that." In that article, I discussed with the journalist who interviewed me what would happen if the ceasefire held, there were no further attacks, and the Strait of Hormuz opened. So you have different scenarios: ceasefire yes or no, Strait of Hormuz open yes or no. According to the model, the oil price would be up for a long time, because an important variable is the damage to the existing infrastructure. Even if things finish abruptly today with a positive outcome, there's still some damage that's going to last.

[29:05] Alexander Denev: Markets have been quite complacent about this conflict. A single tweet, "we open the strait," happened a few weeks ago, and people really forgot about the infrastructure. If you have that in your model, you can really understand: it's there, I have to think about it. Even if things go positively and the strait reopens to commercial traffic, there's still some lasting damage. As I mentioned, there are 1,024 different combinations out of this simple model, and some of them are very peculiar, things you wouldn't think about. I discussed this scenario with the journalist, and now it also seems the majority of analysts are saying: there's some damage, there's so much traffic that has to be cleared, so many delays, freight rates are up. This adds additional cost to oil prices, and the infrastructure cannot immediately accommodate a return to normality, so prices are going to be up for a long time. There was a forecast yesterday from Goldman Sachs mentioning exactly this type of scenario. So yes, Bayesian networks are extremely good causal mental maps that encode many things in a mathematically rigorous way and extend your mind, because you branch different possibilities. You always reason locally when you build the network, this event causes that one, and when you look at the global picture, once it's finished, you find a lot of interesting things.

[31:05] Alex Molak: Your article was published five or six weeks ago, around there. You also shared some preliminary results. As you said, it was a quick exercise, but the results were there. Did the results hold after these six weeks? Some things have changed in the meantime; the situation was very dynamic. I was very curious how you look at the results you got on that day, from today's perspective.

[31:48] Alexander Denev: At the time, the question was about a ceasefire: yes or no. It was discussed but never implemented, so it was still a probabilistic event. Now we know it happened, and there are no more war activities. But the other variables, like whether the Strait of Hormuz is open or closed: the answer is closed. A lot of the other variables have not yet unfolded. So one event is certain, at least for now. When you build these networks, you have to think about your horizon: is it one day, one month, one year? In that case it was one month. So in this month, the ceasefire node materialized, but the "Strait of Hormuz open" node has not. Other nodes that imply a further extension of the war have not unfolded, but they have become less probable, given that a ceasefire is happening. Once you instantiate one of the variables, once you have certainty that it has happened, like the ceasefire, it automatically suppresses different probabilistic paths leading to the larger nodes.

[33:09] Alexander Denev: When it comes to the impact on asset classes, I was putting myself in the shoes of a US asset manager. The bond markets are nervous, and that materialized. The move of the dollar also materialized in the right direction. But equity markets have been unusually complacent. Again, this is a probabilistic model. It's not saying equity will crash with 100% certainty; nobody can have that certainty, because not all of the events have materialized yet. Given what has happened, the model assigned probabilities, and now that the ceasefire is in place, it makes an equity crash less probable than in the case of no ceasefire and continued war. So at least that headwind disappeared, but there are still other possibilities, and not all of the network has materialized. I'll have to update the network now to see how the information has moved and how the network has changed. The beauty of this thing is that it gives you probability distributions with a certain variance, not exact results.

[34:42] Alex Molak: What would be your advice for someone interested in building a similar model on their own?

[34:51] Alexander Denev: Now it's becoming easier. The first step is to understand that this can be really useful, and that in these circumstances all traditional risk models, like value-at-risk or expected shortfall, fail. They fail because they're calibrated on historical data with a certain volatility, and now you have much bigger volatility. Even if you include periods of stress in these value-at-risk models, which is a regulatory requirement, you have to have a stress VaR, people usually calibrate the stress models on past periods like the great financial crisis or COVID, which are totally different. So the first step is to understand that what's needed is a more forward-looking view, and that causal structures are more stable than simulations based on a historical correlation matrix. Exactly in times of crisis, all the historical correlations tend to break.

[36:01] Alexander Denev: So the first step is to understand that, although the standard practice in markets is to have these value-at-risk measures, they're not useful. A change of mentality is needed. Regulators also call for forward-looking scenarios, and this is a framework where those scenarios can be done. So that's my first piece of advice: to understand that this type of forward-looking methodology should supersede pure historical-correlation-based modeling. The second step, and by no means am I saying this is the only methodology out there: sometimes a simple decision tree could be more useful than historical-correlation simulations. To understand the methodology more, as you cover different domains on your podcast, it's becoming more and more widespread. It can be in healthcare, you mentioned marketing as well, it has many applications in industrial production. So there's more and more awareness that this is a useful technology.

[37:14] Alexander Denev: Get acquainted with this, because in one shape or another, if not in finance, causality is going to be the future. Even when building artificial general intelligence, in my opinion, it must fully understand cause and effect to really be called AGI. There are a lot of free tools out there, a lot of code libraries, and you've written about this abundantly and discussed it on your podcast. And now we have Claude. The first step is just to experiment. If you want to know what's happening, you have your traditional tools, expert analysis or a simple decision tree, but now you have all the means to explore something much more powerful, in a very simple way. So experimentation is key. As I said, I threw this into the public space through the article and through my post, but the next phase is experimentation: to find if there are any limitations to the approach. It's an important step to start experimenting now that we have all the tools necessary.

[38:33] Alex Molak: To what extent do you think feeding the model, feeding Claude with your books, was helpful for getting meaningful results?

[38:48] Alexander Denev: This is a bit of a counterfactual query. Maybe I could try with a different account where I didn't feed the books. But there's a lot of information in the public space about this methodology now. With regard to the two books, I have some free articles that describe the methodology, and since they're public, Claude has already ingested them. I'm not saying, "Guys, buy my book as a PDF and upload it to Claude." I think there's enough out there to start with. But it's a good point; I'll try to experiment. Given that Claude already ingested this, I think it's already embedded in the LLMs. Anyway, there's plenty, so experimentation can start without feeding my books.

[39:51] Alex Molak: Many experts would say it was a very good idea, providing the model with additional context that's available to it right now, not something encoded in the weights. That typically improves performance. So that might be a very good idea: adding context to help it ground itself in knowledge in this space.

[40:24] Alexander Denev: Yes. We still don't fully know how these models work, but context is important. Of course, giving more context around the methodology or the scenario you want definitely helps. Contextualization helps.

[40:39] Alex Molak: I agree. Great. Alexander, what would be your message to the community before we close?

[40:44] Alexander Denev: As I mentioned, it's an old technology, and we're seeing breakthroughs in causality. Causal AI is being mentioned more and more, so please go and experiment. As I said, I'm focused on other types of research at the moment, but I'd be curious to see other people experimenting with this and giving me feedback, positive or negative. I hope other people will be able to advance this field and make it more and more practicable and widely usable, because being forward-looking, having a rigorous mathematical framework to behave in situations like this, could save a lot of problems and a lot of money for asset managers. The ultimate goal is always your outcome: to be better at what you do. So my suggestion is: please do experiment, and give me feedback. I'm happy to discuss and work with you on that.

[42:05] Alex Molak: And for someone who is just starting with causality, what is one resource you'd recommend to familiarize themselves?

[42:14] Alexander Denev: Your book, of course. It's very well written, and I'd highly recommend it as a very good entry level. And "The Book of Why" by Judea Pearl, which is the simpler version of his "Causality" book; that one is more for advanced users. For specific applications to finance, I'd recommend Riccardo Rebonato's book, "Coherent Stress Testing." Also Marcos López de Prado's book. And if you want to build scenarios in the way I mentioned, of course my two books, but the content of those is already available in many free papers out there. So that's a good resource if you want to apply the causality methodology to the US use case I described.

[43:17] Alex Molak: Great. Alexander, thank you so much. It was a pleasure to have you again, and I hope that's not the last time we speak on this podcast.

[43:25] Alexander Denev: I hope so. Thank you, Alex, for having me. Thank you.