It is daunting to write an article like this and hope that, after publishing, it could earn some nods. Whether acquiring the public’s acknowledgement or simply holding more than five seconds of attention in social media traffic is our goal, I cannot tell. We are living in a strange world, and sometimes I wonder: if Mother Teresa were still alive, what would she think about the masses, and the new forms of poverty and equality?
The other day, I read Pope Leo’s Magnifica Humanitas. For someone used to killing time sitting at the back of a church flooded with Hindu-looking people, I read the Bible in a copy that had been handed to me. I read the Old Testament because my university courses mostly ignored it, or simply said, this is Jewish, let’s start with the New Testament. The idea of the new really did the trick.
The Old Testament reads like a Chinese Fengshen Bang to me: Noah with his boat, Abraham living to more than 100 years old, Sarah having a child at 90. The serpent gave wisdom to the woman, and both ended up cursed by the Lord. It sounds like a camp with internet censorship, or a prison in Siberia, where the purpose is torture and permanent exclusion from civil life.
Calling it fascinating would be an exaggeration, but I agree that it is largely a mirror of human history. Nothing can be taken at face value, nor can we fully doubt its originality or truth. Did Abraham really have his child with the Egyptian slave girl of his wife Sarah? Was that the same figure later called Ishmael, and later connected to Islam? The spirit of it is largely a narrative of mankind: we live, we die, we pass from generation to generation, we move from one state to another. There is good and bad, and the criteria keep changing. Then a new story comes, with brand new phrases after mankind was erased by the Lord after the great flood. A cycle.
I have to accept that I do have good memories. I read, and I do not forget. I often recite details that even surprise myself. Yet in the era of AI, this skillset is being eclipsed, since AI does not get tired and can master multiple languages simultaneously while I only work on a single core, or occasionally multicore processes.
Deep quant, or simply quant finance, is now facing AI. AI is not new, and cloud computing is not new either. The search for high-performance computing and efficient modeling using mathematical frameworks and technology tools has always been at the heart of quant finance. But we have more data now, and computers are more powerful through distributed systems. This simplifies the end-user experience and makes systems easier to adopt and scale. Yes, that is the word: scale. We are building an ecosystem, a chain of values. Cloud-based data centers, data itself, powerful GPU machines, and even quantum computing, whether annealing- or gate-based, all form part of it. It is still not fully proven to work well under all constraints.
So what are we now? Are we all going to be replaced as low-value human capital, reduced to survive on a universal living credit, counting the birds if there are any left? The real question is: what is the purpose of human beings?
Revolutions. Existential authenticity. Resisting the fallacy of invulnerability. This is a lot, even by my standards.
I will shout that I am low-value human capital, but I have a very strong intrinsic sense of being. AI plays piano better than I do, but I feel immense pleasure and soul-touching emotion while playing it. Maybe I should connect my neurons to a machine so that it can sense this feeling and learn from it.
Anyway, enough brain tickling for now. Below are the use cases, mainly from our own experiences in the field, a summary of the last 20 years across JP Morgan, Barclays, Morgan Stanley, Nordea, Deutsche Bank, Citi, and more. The list is long, and I am happy to share it for free as a contribution to the industry and for more people to benefit. I am lucky to have experienced the 2008 Financial Crisis in JPMorgan’s Credit Derivatives business, pricing CDOs, with little knowledge that Lehman Brothers did not have an equivalent sister.
The AI industry is massive in wealth and power, and it must be distributed to all humankind for the benefit of all. Distribution is not only about profits or wealth itself. AI harvests from human intelligence; in return, it should serve all humans’ life purpose and wellbeing. Using London pub words: be my bitch.
The new digital slavery is invisible, often dangerous labor in mining rare metals, running data centers, content moderation, and click work that sustains AI infrastructure while concentrating profits and power elsewhere.
Should we build a bigger cake and share it for common prosperity, or enslave the masses to build an ever-larger cake?
Remember that artificial intelligence is not new, and there is no intelligence at all. It is basically an aggregation of human data, harvested and trained by the machine. It is the average of humankind. And remember: you do not want to be labeled average, because that hints at mediocrity.
The first use case is the simplification of analytics library maintenance and deployment. Most investment banks and trading firms, including software providers like Bloomberg or Murex, have in-house analytics libraries for derivatives pricing and risk management: market risk, VaR, CVaR, counterparty credit risk, XVA, and even credit risk. A typical example is volatility modeling, such as Heston and SABR. It was built in the 1980s, along with Black-Scholes, as a market model where mathematics sits at the heart of the engineering. In the absence of enough data points to calibrate or build the vol surface, we rely on the modeling framework to fill the data holes and build an implied volatility surface for option markets. That is the nature of the market: there are not enough option prices or strike points to infer volatility across strikes and tenors. The implied vol surface is incomplete, and so the market model became popular for a period. Now, with data and the technology of market microstructure, we explore latency and compress the space of data to infer from it. This has largely reduced the reliance on market models and heavy mathematical assumptions. For a mathematician, that is bad news; for data scientists, it is good news.
The second case is stress testing, or macroeconomic scenario generation with risk-factor generation. This requires massive datasets. When I was at Citi, the challenge was how to infer the prescribed limited scenarios from the CCAR exercise into the infrastructure and risk framework. In order for the risk-management analytics to rerun the scenarios, each risk factor needs proper data input: gas 2Y, Australian housing index, and so on. Some are easier to fill, like those prescribed by the Fed, such as Nasdaq shocks or volatility. Most of them have to be constructed. It is called scenario risk-factor regeneration, and it ends up with around 14,000 risk factors and associated shocks to be fed into different risk models to compute capital or stress loss. This is a scenario-driven exercise, and it is very much a heavy-engineering task that takes months to complete, with many issues.
The third case is speed and HPC, which is useful in Monte Carlo, portfolio optimization, index rebalancing, and algo-trading strategy calibration and review. One example is a Monte Carlo simulation project for CCR and XVA. Without AAD, calculating or computing expected exposure requires multiple loops, repeatedly computing the payoff on the nodes across simulation points and time dimensions. It is at least two matrices, or even more calculation. If one wants to calculate Greeks along these multidimensional vectors, the whole exercise becomes time-consuming and computationally intensive. GPU is a good solution, on top of multithreading or distributed computation, because CPU has its limitations even if multiple process cores are switched on. My view is that it speeds up computation and frees up resources. The efficiency gained could allow more frequent computation, daily or even intraday, whereas some companies today run weekly PFE. For portfolio optimization, the same efficiency gain could reduce the time needed for portfolio rebalancing and calibration, which leads to optimization because the procedure can run within a much shorter timeframe.
Partners include NVIDIA HPC GPU, CUDA, Claude Code, and Gemini ICL. Cloud data centers include AWS, Azure, and Google. Software and data providers like Databricks and Snowflake are also part of the stack. Sometimes it is better to use vendor data, because it is already cleaned and carries intelligence built in, rather than purchasing raw data and hoping intelligence is there. Imperfection creates room and possibility for evolution toward perfection. Being.