Finance and LLMs
What they can learn from each other
As a finance practitioner, and also an active researcher in LLMs, I’ve found many connections between the two fields, particularly in how they simulate possible future paths and assign value to outcomes.
Key Similarity
| Finance | LLMs | |
|---|---|---|
| Simulator | Options: Heston, local vol; HFT/ market making: order book/ flow, queue dynamics. | A pretrained language model generates possible future token sequences. |
| Sampling | Monte Carlo price paths; order, cancellation, and execution events. | Inference / rollout samples many possible reasoning trajectories. |
| Reward | Option payoff; trading PnL adjusted for risk, costs, and market impact. | A reward or verifier assigns value to each generated trajectory. |
| Scale & Infra | Options: pricing speed, Greeks, etc; HFT: latency, networking, co-location. | Inference: throughput, bandwidth, parallelism, serving. |
| Resource Allocation | Expected return, risk, alpha, and correlations guide capital allocation. | Scaling laws guide compute, data, and model-size allocation. |
| Safety | Control tail risk through stress tests, risk limits, and drawdown constraints. | Control harmful behavior through evaluations, guardrails, and monitoring. |
| Data | Massive market data, but rare regimes and high-value signals remain scarce. | Massive training data, but rare, specialized, and high-quality data remain scarce. |
Key Difference
| Finance | LLMs | |
|---|---|---|
| Dimensionality | Traditionally low- to medium-dimensional. | Extremely high-dimensional |
| State space | Continuous in option; discrete / event-driven in HFT. | Discrete token space |
| Dynamics | Stochastic price dynamics or discrete market events. | Discrete autoregressive |
| Applications | Pricing, trading, hedging, and risk management | Reasoning, coding, and personal AI |
What They Can Learn from Each Other
Finance can learn from LLMs by moving beyond individually calibrated, low-dimensional models toward high-dimensional, multimodal market simulators learned jointly across assets, signals, and market states.
LLMs, in turn, can learn from finance by placing greater emphasis on safety, tail risk, uncertainty, stress testing, hard constraints, risk controls, and robustness when models fail.
In short, LLMs can help finance model a larger world, while finance can help LLMs handle failure.
Acknowledgment
Thanks to ChatGPT for helping refine the ideas and wording in this post.