Quantitative Finance AI
Why Physics-Based Models Are Replacing Monte Carlo
Quantitative Finance AI: Why Physics-Based Models Are Replacing Monte Carlo
Quantitative finance has long run on mathematical models — Monte Carlo simulation, stochastic calculus, factor models — applied to pricing, risk management, and portfolio construction. These tools were built for a world of relatively tractable instruments. Today’s institutional portfolios are something different: complex structured products, multi-asset derivatives, high-dimensional correlation structures, and tail risk scenarios that require evaluating hundreds of millions of portfolio variations to model accurately. Monte Carlo simulation, at its core a random-sampling approach, hits a practical ceiling here. The case for AI in quantitative finance is not that it replaces mathematical rigor — it is that Large Quantitative Models (LQMs) bring the computational scale and quantitative grounding needed to work at the complexity levels modern financial institutions actually face.
What quantitative finance AI actually addresses
The problems that AI is being applied to in quantitative finance are not new — they are the same problems practitioners have always faced, now compounded by the scale and complexity of modern markets.
Pricing complex instruments is the foundational challenge. Structured products, exotic options, collateralized debt obligations, and multi-asset derivatives all require modeling many interacting risk factors simultaneously. The more factors, the more paths a simulation needs to generate to produce a reliable price estimate — and the more computationally expensive the exercise becomes. Monte Carlo has been the standard tool for this work for decades, and it remains effective for instruments of moderate complexity. The question is what happens as complexity increases.
Portfolio-level risk presents a related but distinct problem. Understanding the tail risk of a large institutional portfolio — the behavior of the portfolio in extreme market scenarios, not just average conditions — requires evaluating how the portfolio responds across a vast range of potential outcomes. The relevant question is not what happens on average; it is what happens in the scenarios that rarely occur but matter most when they do. Getting a reliable answer to that question requires far more scenario coverage than traditional simulation provides in practical compute time.
Market dynamics add a third layer of difficulty. Financial markets are non-stationary: the correlations between assets, the volatility regimes, and the macroeconomic relationships that governed historical data do not remain fixed. Models trained on historical data may not capture the correlation structures that emerge under stress — precisely when risk management most needs to be reliable. The tail behavior that matters most for risk is also the behavior for which historical data is most sparse.
Where Monte Carlo hits its limits
Monte Carlo simulation works by generating random samples of possible outcomes and using the distribution of those samples to estimate the quantity of interest — a price, a risk measure, a loss probability. Its convergence rate scales as the square root of the number of samples: to halve the estimation error, you need four times as many samples. For simple instruments in low-dimensional settings, this is manageable. For complex, high-dimensional portfolios, it becomes a practical constraint.
Standard Monte Carlo implementations generate thousands to tens of thousands of scenario paths. For straightforward option pricing or value-at-risk estimation on a small book, this is sufficient. For a large institutional portfolio with complex structured instruments, correlated risk factors, and a need to understand tail behavior with precision, thousands of paths do not provide the scenario coverage needed to characterize the distribution reliably.
Variance reduction techniques — quasi-Monte Carlo, multilevel Monte Carlo, control variates — improve the efficiency of sampling and have genuine value in many applications. They do not resolve the fundamental scaling challenge for very high-dimensional problems with complex path dependencies. The deeper issue is that the instruments and portfolios that matter most to large financial institutions are precisely the ones where the limitations of sampling-based methods are most acute.
Jack Hidary, CEO of SandboxAQ, has put the problem plainly: “Monte Carlo simulation is not sufficient anymore to handle the complexity of structured instruments.” The observation points to a gap between what the standard toolkit produces and what the problem actually requires — a gap that has grown as structured finance has grown.
How LQMs change the calculus
Large Quantitative Models are trained on quantitative data — mathematics, physics, financial structure, numerical relationships — rather than on text or historical market records alone. This distinction matters for finance in the same way it matters for drug discovery or materials science: a model that learns the underlying quantitative structure of a problem can extrapolate beyond its training distribution in ways that a purely statistical model cannot. Large Quantitative Models are purpose-built for exactly this class of problem: high-dimensional, numerically complex, data-sparse in the tail regions that matter most.
The scale difference is significant. Where traditional Monte Carlo generates thousands of scenario paths, LQMs can evaluate hundreds of millions of portfolio variations to map tail risk with precision. Hidary has described the practical implication directly: “If I have a portfolio and I want to know what the tail risk is given changes in this portfolio, what I’d like to do is I’d like to create 300 to 500 million versions of that portfolio with slight changes to it, and then I want to look at the tail risk.” That order of magnitude of scenario coverage changes what risk managers can know about the extreme behavior of complex portfolios — not just what the average outcome looks like, but how the portfolio behaves in the scenarios that occur once every decade or less.
For high-dimensional correlation structures, LQMs handle the mathematical relationships between instruments and risk factors without the sampling constraints that limit Monte Carlo at scale. The models are not approximating a distribution by sampling from it; they are learning the structure of the problem directly from quantitative training data and applying that structure to new scenarios at production speed.
The broader context is what SandboxAQ describes as the quantitative economy: a $50+ trillion sector spanning financial services, biopharma, energy, and advanced materials where decisions are driven by numerical data and physical relationships rather than language. Financial services is one vertical within this; the same LQM infrastructure that models molecular behavior in drug discovery and material properties in materials science applies to the mathematical structure of financial instruments. The underlying technical challenge — modeling complex quantitative systems at scale — is the same across all of them.
Quantitative finance AI and the institutional signal
The direction of investment in this space is informative. In April 2025, BNP Paribas joined SandboxAQ’s Series E funding round. Olivier Osty, BNP Paribas Head of Corporate and Institutional Banking Global Markets, stated: “AI and advanced computing are having a powerful impact on financial services, and BNP Paribas Global Markets is proud to be at the forefront of this trend. I look forward to working with SandboxAQ to explore innovative solutions at the cross-roads of AI and quantum techniques.” Ray Dalio, founder of Bridgewater Associates, also joined the round: “I bet on the SandboxAQ team and its approach to Large Quantitative Models because I’m impressed by them both.”
At the World Economic Forum in Davos in January 2025, Hidary discussed AI and the future of financial markets with representatives from BNP Paribas and HSBC, with financial services positioned as a core LQM application vertical. The recurring framing across these engagements is consistent: LQMs analyze high-dimensional financial data, identifying patterns and correlations relevant to portfolio resilience, risk forecasting, and optimization that operate beyond the reach of conventional simulation.
SandboxAQ has raised over $950 million in total funding, reaching a valuation of $5.75 billion, with investors including T. Rowe Price, Google, NVIDIA, Breyer Capital, BNP Paribas, and Ray Dalio. The composition of the investor base — spanning financial institutions, technology companies, and quantitative investors — reflects the breadth of industries where LQMs are seen as consequential.
What this means for quantitative finance teams
The practical implication for teams working in quantitative finance is a shift in what is computationally feasible for tail risk modeling and complex instrument pricing. Moving from thousands of Monte Carlo paths to hundreds of millions of portfolio scenario evaluations is not an incremental improvement; it is a change in what questions can be answered with confidence. Multi-factor portfolio optimization, stress testing under correlated risk factor shocks, and tail risk quantification for complex structured books all benefit from this order-of-magnitude increase in scenario coverage. The LLM-vs-LQM distinction is worth being precise about in this context. A large language model trained on financial text can summarize reports, draft communications, and extract information from documents. It cannot model the mathematical structure of a collateralized debt obligation, price a path-dependent exotic option, or simulate the correlation behavior of a multi-asset portfolio under a stress scenario. An LQM trained on quantitative financial data can. For teams evaluating where AI fits into their quantitative research and risk infrastructure, the relevant question is not whether AI is useful in finance — it demonstrably is — but which class of AI model is appropriate for which class of problem. For the broader context on how SandboxAQ frames the quantitative economy and LQMs’ role across industries, the AI for the physical world backgrounder and the what is SandboxAQ overview provide useful context.
Frequently asked questions
What is quantitative finance AI?
Quantitative finance AI is the application of AI models to problems in financial modeling, pricing, risk management, and portfolio optimization. It ranges from machine learning models applied to market prediction and factor investing, to Large Quantitative Models that handle the mathematical structure of complex instruments and portfolio-level tail risk at scales that conventional simulation cannot reach.
What are the limitations of Monte Carlo simulation in finance?
Monte Carlo simulation’s convergence rate scales as the square root of the number of samples: halving the estimation error requires four times as many samples. For complex, high-dimensional portfolios with many interacting risk factors, generating sufficient scenario coverage to characterize tail risk reliably becomes computationally prohibitive. The method also depends on assumptions about the distribution of risk factors that may not hold under stress conditions — precisely when tail risk estimation matters most.
How are Large Quantitative Models used in financial services?
Large Quantitative Models are trained on quantitative data — mathematics, financial structure, numerical relationships — rather than on text or historical market records alone. In financial services, they are applied to portfolio-level tail risk modeling, complex instrument pricing, high-dimensional correlation analysis, and scenario simulation at scales that conventional Monte Carlo methods cannot match. LQMs can evaluate hundreds of millions of portfolio variations to map tail behavior with precision.
What is the difference between an LLM and an LQM in finance?
A large language model (LLM) is trained on text and is suited to tasks like summarizing financial reports, drafting communications, and extracting information from documents. A Large Quantitative Model (LQM) is trained on quantitative data — mathematical structure, numerical relationships, physical and financial systems — and is suited to modeling the behavior of complex financial instruments, simulating portfolio scenarios, and quantifying risk at scale. The two are complementary: LLMs handle language tasks; LQMs handle quantitative modeling tasks.
What is tail risk modeling in quantitative finance?
Tail risk modeling is the process of estimating the probability and magnitude of extreme outcomes in a financial portfolio — the losses that occur rarely but represent the most severe scenarios. Because tail events are by definition infrequent, they are underrepresented in historical data, and conventional simulation methods require large numbers of scenario paths to characterize them reliably. LQMs address this by enabling hundreds of millions of portfolio scenario evaluations, providing the scenario coverage needed to map tail behavior with meaningful precision.
Explore SandboxAQ’s quantitative AI platform
- Large Quantitative Models — physics-grounded AI for financial services and beyond
- SandboxAQ