Renaissance Technologies built the most successful hedge fund in history using machine learning before most people had heard the term. The Medallion Fund has returned 66% annually (gross) since 1988. They employ 300+ people, mostly scientists and engineers, working on models that process petabytes of data.
So when people ask "what's new about LLMs for finance?", it's a fair question. Quant funds have used ML for decades. Two Sigma, DE Shaw, Citadel, and hundreds of others have sophisticated machine learning pipelines. What exactly would change?
The answer: almost everything. The models Renaissance uses are fundamentally different from LLMs in ways that matter for training, deployment, and capability. Understanding this difference is essential for anyone trying to build financial AI today.
What Traditional Quant ML Does
Traditional quantitative machine learning in finance typically does one or more of the following:
Signal Prediction
Given features X (price patterns, volume, fundamentals, etc.), predict outcome Y (price movement, volatility, etc.). The model learns statistical relationships between inputs and outputs.
Pattern Recognition
Identify recurring patterns in market data. Technical analysis, but automated and at scale. If pattern A historically precedes price movement B, trade accordingly.
Optimization
Given a set of possible trades and constraints, find the optimal portfolio allocation. Mean-variance optimization, risk parity, and similar approaches.
Classification
This trade/customer/transaction belongs to category A or B. Used for fraud detection, credit scoring, and transaction classification.
These tasks share common characteristics:
- Well-defined inputs and outputs: The feature space is known in advance
- Numerical data: Prices, volumes, ratios, and other quantitative measures
- Narrow scope: Each model does one specific thing well
- Static deployment: Train once, deploy, retrain periodically
Traditional quant ML is narrow intelligence: powerful within specific domains, but unable to generalize or adapt to tasks outside its training distribution.
What LLMs Do Differently
Large language models are fundamentally different beasts:
General Capability
LLMs aren't trained for a single task. They learn a compressed representation of human knowledge and reasoning that can be applied to many tasks, including tasks never explicitly seen during training.
Language Interface
LLMs interact through natural language. This means they can be instructed, can explain their reasoning, can ask clarifying questions, and can integrate unstructured information like news articles or analyst reports.
Reasoning Chains
LLMs can perform multi-step reasoning. Rather than mapping input directly to output, they can think through problems, consider alternatives, and arrive at conclusions through logical steps.
Tool Use
Modern LLMs can use external tools: call APIs, execute code, query databases. This extends their capabilities beyond what's encoded in their weights.
Adaptation
LLMs can be prompted to behave differently for different contexts. The same model can be a conservative risk analyst in one prompt and an aggressive trader in another.
Why This Changes Training Requirements
The training data for traditional quant ML is straightforward: historical prices, volumes, fundamentals, and labels for what you're trying to predict.
LLM training for financial applications requires fundamentally different data:
Reasoning, Not Just Outcomes
A quant model needs to know: "When RSI crosses above 70 and price breaks resistance, go long." It doesn't need to know why.
An LLM needs the reasoning: "RSI crossing 70 suggests momentum is strong. Price breaking resistance confirms buying pressure. The combination suggests continuation is likely. However, check volume for confirmation and consider profit-taking resistance at the next level."
The reasoning trace is the training data, not just the conclusion.
Language Context, Not Just Numbers
A quant model consumes feature vectors: [price: 65000, volume: 1.2M, RSI: 72, ...]
An LLM needs to understand that "the Fed signaled hawkish intent at Jackson Hole" connects to interest rate expectations which affects risk asset prices which influences the decision at hand.
Training data must include the linguistic and conceptual context, not just numerical features.
Decision Episodes, Not Just Signals
A quant model needs (features, outcome) pairs. See these inputs, predict this output.
An LLM needs complete decision episodes: what was the situation, what was the intent, what reasoning was applied, what tools were used, what action was taken, what happened, and what should be learned.
A quant fund with decades of price data and extensive feature engineering has almost nothing usable for LLM training. The data formats, structures, and information content are fundamentally incompatible.
The Integration Opportunity
This isn't either/or. The future likely involves LLMs and traditional ML working together:
LLMs for Interpretation
LLMs can process unstructured information that quant models can't handle: news, social media, research reports, regulatory filings. They can extract features that feed into traditional ML pipelines.
LLMs for Decision-Making
Given signals from quant models, LLMs can make decisions that integrate multiple factors, consider constraints, and adapt to context in ways that pure numerical optimization can't.
LLMs for Explanation
Quant models are black boxes. LLMs can provide natural language explanations of what models are doing and why, enabling human oversight and debugging.
LLMs for Strategy Generation
Rather than manually specifying strategies, LLMs could propose and iterate on trading approaches, which are then validated by rigorous backtesting with traditional tools.
What Doesn't Transfer
Financial firms hoping to leverage their existing ML capabilities for LLM development will find some things don't transfer:
Data Assets
Historical price data isn't useful for LLM training. Decision data, reasoning traces, and human feedback are what matters.
Training Infrastructure
Infrastructure for training gradient boosted trees or simple neural networks doesn't scale to LLM training, which requires massive compute clusters and different optimization techniques.
Evaluation Methods
Backtesting a quant strategy is well-understood. Evaluating whether an LLM can make good financial decisions is harder and requires different methodologies.
Development Workflows
Quant development: research signal → backtest → refine → deploy. LLM development: pre-train → post-train → prompt engineer → evaluate → iterate. Different skills, different tools.
What Does Transfer
Some things are valuable regardless of the technical approach:
Domain Expertise
Understanding markets, risk, and trading remains essential. LLMs need to be trained and evaluated by people who understand finance.
Risk Management Culture
The discipline of testing, validating, and monitoring models applies equally to LLMs. If anything, LLMs require more rigorous oversight given their complexity.
Market Access
Execution infrastructure, data feeds, and market relationships are valuable regardless of what's making decisions.
Regulatory Understanding
Knowledge of compliance requirements, reporting obligations, and regulatory constraints applies to LLM-based systems as much as traditional quant.
Implications
If you're at a financial firm thinking about LLMs:
- Don't assume your ML team can just pivot. LLM development requires different skills. You'll likely need new expertise.
- Your data advantage may not be what you think. Price data doesn't help. Decision data does. Most firms have the former, not the latter.
- Integration is more realistic than replacement. Use LLMs to enhance existing workflows rather than rebuild from scratch.
- Start building decision data infrastructure now. The firms that capture high-quality decision data today will have LLM advantages tomorrow.
Traditional quant ML and LLM training are complementary but distinct disciplines. They solve different problems, use different data, and require different expertise. Financial firms need both, but shouldn't assume capability in one implies capability in the other.