Better Financial AI Starts with Better Writing Patterns

When organizations evaluate AI performance, most attention goes to model selection, data quality, or retrieval.

Far less attention is given to something much simpler:

How should the model write?

Most production systems rely on hand-written prompt instructions that tell an LLM to be concise, professional, or analytical. Those instructions are often based on intuition rather than evidence, and few organizations measure whether they actually improve answer quality.

Our latest research asks a different question:

Can the writing patterns found in professional financial research be automatically extracted and used to improve AI-generated answers?

The results suggest they can.

Turning Financial Research Into a Style Guide

Rather than writing prompt instructions by hand, we analyzed approximately 12,800 professional financial research documents spanning equities, credit, commodities, macroeconomics, and foreign exchange research.

From each document, we extracted characteristics such as:

• Writing style and sentence structure
• Tone and use of analytical language
• How numerical data is integrated into prose
• Technical vocabulary
• Common analytical relationships and financial concepts

Those individual analyses were then synthesized into a single corpus-derived style template that could be added to a model's system prompt.

Importantly, this process changed how the model answered, not what information it received.

Both the baseline model and the templated model were given identical source material. The only difference was the guidance used to generate the final response.

That allowed us to isolate the impact of writing style alone.

Better Answers Without Changing the Data

The improvement was substantial.

Across blinded A/B evaluations, independent AI judges consistently preferred answers generated using the corpus-derived template.

Using Claude Opus 4.8 as the answering model, the templated responses were preferred 90.5% of the time.

The results were not limited to a single judge or provider.

Three independent non-Anthropic judges, including Mistral, Amazon Nova, and DeepSeek, also preferred the templated answers, with win rates ranging from 79% to 92%.

Perhaps more importantly, the gains were not simply cosmetic.

The study also measured key fact recall and found that templated answers consistently included more of the important information from the provided research.

In other words, the models were not just writing more like analysts.

They were producing more complete answers.

Bigger Models Are Not the Only Answer

One interesting finding is that prompt quality can materially improve outputs without changing the underlying model.

Organizations often assume that better AI requires upgrading to the newest frontier model.

This research suggests another opportunity.

Well-designed guidance, derived from expert human writing, can significantly improve answer quality using the same model and the same source material.

Prompt engineering has long been treated as something of an art.

This work suggests it can become a measurable, data-driven process.

One Template Was Better Than Many

The research also tested whether highly specialized prompts would outperform a single global template.

Researchers created 89 topic-specific templates covering areas such as sectors, asset classes, and research domains.

Surprisingly, they did not outperform the global template.

Even when the correct topic template was selected perfectly, the single global template performed just as well or better.

For organizations deploying AI at scale, this is encouraging.

Rather than maintaining dozens of specialized prompts, a single well-constructed template may deliver better consistency while dramatically simplifying implementation.

What This Means for Financial AI

As organizations build production AI systems, three factors increasingly determine answer quality:

• Access to high-quality financial data
• The model's ability to retrieve and synthesize that information
• The guidance that shapes how answers are constructed

This paper focuses on the third element.

It demonstrates that professional financial writing contains consistent patterns that can be extracted automatically and used to improve AI-generated research.

Rather than relying on manually written prompt instructions, organizations can derive those instructions directly from the expertise embedded in thousands of analyst reports.

Improving AI by Learning From Analysts

Financial analysts have spent decades refining how research is communicated: leading with conclusions, integrating evidence naturally, expressing uncertainty appropriately, and organizing complex information into clear investment narratives.

Those patterns are now teachable.

By learning from the collective writing style of thousands of professional research documents, AI systems can generate answers that are not only more natural, but also more complete and more useful for financial professionals.

The findings suggest that better financial AI is not only about choosing a stronger model or providing better data.

It is also about teaching models to communicate the way experienced analysts already do.

Download the full Corpus-Derived Style Templates Improve LLM Financial Research Q&A paper to explore the complete methodology, evaluation framework, experimental results, and implications for financial AI systems.