Finance is a Latent LLM
How market simulators connect to LLMs
As a practitioner in finance, who is also an active researcher in large language models (LLMs), I’ve found that finance and LLMs are closely intertwined. In a sense, the finance system can be seen as a latent large language models in the following aspects:
Environment: Market Simulators vs. Token Simulators
In finance, a market simulator is a synthetic finance world model that generates prices, order flows, volatility, and liquidity and indices to test your strategies and models.
LLMs simulate discrete token sequences to model language and reasoning trajectories.
Reinforcement Learning
- Finance: option pricing, hedging, and insurance → minimize tail risk
- LLMs: RLHF / RLVR → align behavior with user preferences
RL functions as risk control in finance versus behavior alignment in LLMs.
Optimization & Allocation
- Finance: portfolio optimization hedges risk across dissimilar (anti-correlated) assets
- LLMs / Tech: recommendation systems allocate similar items to similar users
Both are constrained allocation problems under uncertainty.
Scale & Infrastructure
- HFT: nanoseconds, latency, hardware dominance
- LLMs: tokens/sec, pipeline parallelism, bandwidth limits
At scale, systems engineering dominates algorithmic details.
Alpha–Beta vs. Scaling Laws
Finance uses alpha for excess returns and beta to model market exposure.
LLMs rely on scaling laws to predict final loss and determine when to stop training.
Both guide capital and compute allocation.
State & Control
- Finance: hidden latent state inferred from noisy prices
- LLMs: hidden activations inside deep neural networks
Finance infers the state; LLMs are the state.
Prompt Engineering vs. Technical Analysis
- Finance (technical analysis): conditioning trades via patterns in past prices
- LLMs (prompting): conditioning behavior via input structure
Control without retraining.
Safety
- Finance: minimize catastrophic loss via VaR, stress tests, and drawdown limits
- LLMs: minimize harmful outputs by filtering toxicity, hallucination, and misuse
In both systems, tail risk matters more than average performance.
Fundamental analysis v.s. pretraining
model training
pretraining + fine-tuning.