JPMorgan Seeks Patent for AI-Generated Stock Ratings
The bank is seeking patent protection for an AI system that makes buy, hold and sell recommendations based on market data, news and sentiment.
JPMorgan is seeking patent protection for an AI system that generates stock-rating predictions, applying AI to one of Wall Street’s most familiar research formats: buy, hold and sell calls.
The AI rater draws on company fundamentals, market data, financial news and sentiment to produce analyst-style stock recommendations that are tested against future returns, according to the patent application, which was published in April and initially filed in February 2025. The system generates one of five outputs: Strong Buy, Moderate Buy, Hold, Moderate Sell, or Strong Sell.
A JPMorgan spokesperson said the application came from the bank’s AI research group and was filed to protect the underlying research rather than to commercialize AI-generated stock ratings.
This is one of the first examples I’ve come across of a major Wall Street firm describing an AI system that could generate analyst-style stock-rating predictions. It is a structured pipeline that compresses news, scores sentiment, packages fundamentals and return data, prompts an LLM to reason through future rating horizons, then grades the output against realized forward returns. It’s not asking ChatGPT what stocks to buy.
I don’t read this as evidence that JPMorgan is about to automate sell-side research. It’s more a framework for overworked analysts dealing with information overload.
From the application:
Traditional stock rating methods rely heavily on the expertise of financial analysts and face several challenges such as data overload, inconsistencies in filings, and delayed reactions to market events. The rapid integration of advanced machine learning techniques, particularly Large Language Models (LLMs), presents opportunities to enhance the equity stock rating process.
JPM also notes that the models face limits around context windows, numerical and tabular data, training bottlenecks and the risk of inaccurate responses.
The bank’s CEO, Jamie Dimon, has long been bullish on AI. In his 2023 shareholder letter, Dimon compared the technology’s potential impact to the printing press, steam engine, electricity, computing and the internet.
PATENT DATA
AI Street Patent Review Tracker
I’ve reviewed the relevant finance and AI patent applications published so far in 2026, and I’m continuing to review new publications as they appear, with the help of AI, of course.
This database is a research aid for paid subscribers. It collects finance, trading, market-structure, banking, AI, and infrastructure-related patent applications that appear potentially relevant to Wall Street, fintech, exchanges, clearing, settlement, fraud detection, and institutional data systems.
So far, the filings include applications from:
JPMorgan around stock-rating predictions and financial time-series analysis
CME and ICE/NYSE around exchange resiliency, matching engines, risk controls, and clearing mechanics
Bank of America around AI-driven data plumbing and application connectivity
BlackRock, Schwab, Fidelity, Morgan Stanley, and others around portfolio analytics, wealth infrastructure, model governance, and institutional data systems
Paid subscribers can scroll down to the end of this post to download the tracker.
More Specifics on JPM’s AI Stock Rater
Core task: Generate stock-rating predictions with an LLM.
Rating scale: Strong sell, moderate sell, hold, moderate buy or strong buy.
Prediction horizons: Over 1, 3, 6, 12 and 18 months.
Basic idea: Build a structured dataset around a company, date and future horizon, then ask the LLM to reason through the information and produce an analyst-style rating.
What data goes into the system
Company identifiers: Company name, ticker and relevant date.
Market data: Historical returns, price data, volatility and other technical indicators.
Fundamentals: Financial metrics such as earnings, revenue, return on assets and other company-level data.
News: Company and sector news, including raw articles and summarized versions.
Sentiment: Scores derived from news summaries, with negative, neutral or positive readings.
Forward-return labels: Future stock-return data used later to train or evaluate the rating predictions.
How the news pipeline works
Filtering: A pre-processing LLM removes articles that are not relevant to the company.
Summarization: The same preprocessing step condenses the remaining articles into short company-specific summaries.
Key-event extraction: The summaries are designed to preserve the important developments without overwhelming the prediction model.
Sentiment scoring: The summarized news is converted into a sentiment score from -5 to +5.
Purpose: The news pipeline turns a large, noisy set of articles into a compact signal the rating model can use.
ICYMI Interview
Back in December 2024, I spoke with one of JPM’s patent co-authors, Tucker Balch, who’s now back in academia at Emory, about where he sees the best AI and investing use cases.
One example that I still remember is using AI to expand data sources in other languages:
For instance, if you can listen to the news in Vietnam, translate it in real time, and identify relevant information for specific stocks, you greatly expand your data sources.
How prediction works
Prompt construction: The LLM receives a prompt telling it to act as a financial analyst and predict stock ratings.
Example answer: The prompt can include an input-output example so the model knows the expected format.
Future dates: The model is asked to identify which future months correspond to each prediction horizon.
Explanation: The model is asked to explain the reasoning behind its ratings.
Final output: The model then produces ratings for the specified future horizons.
How hallucination checks work
Date check: The system asks the LLM to calculate the future dates tied to each horizon.
Verification logic: If the model gets the dates wrong, that is treated as a warning sign about the reliability of the rating.
Explanation check: The model’s explanation is used to see whether the rating is actually supported by the input data.
Chain-of-verification: The system uses these intermediate steps to catch cases where the model may be producing unsupported answers.
How training and fine-tuning work
Prompt-label pairs: Training examples pair a prompt with a correct rating label.
Ground-truth labels: The labels are based on future stock performance, not just analyst opinions.
Forward-return quintiles: Future returns are divided into quintiles and mapped to rating categories.
Loss function: The system computes cross-entropy loss between the model’s predicted rating and the ground-truth rating.
LoRA fine-tuning: The model can be fine-tuned using low-rank adaptation, which updates smaller added matrices rather than retraining the full LLM.
Validation: The system can split the data into training and validation sets to test whether fine-tuning improves performance.
How the system checks whether the prediction was right
Forward returns: The system looks at how the stock actually performed after the prediction date.
Peer comparison: The stock’s return is compared with other companies over the same period.
Sector-relative return: The company’s return can be adjusted against sector performance.
Rating match: If the stock’s future-return quintile matches the model’s rating category, the prediction is treated as correct.
Error measurement: Mean absolute error is used to measure how far the predicted rating was from the ground-truth rating.
What the results show
LLMs did better in shorter-term tests: The application says the LLM may perform better on short-term predictions, while analyst errors declined over longer horizons and were slightly better in the 18-month period.
Fundamentals mattered most: The best-performing setups were the ones using fundamentals, especially fundamentals plus sentiment.
News alone helped less: News summaries and sentiment by themselves did not outperform the fundamentals-based setup.
Sentiment added only modestly: Fundamentals plus sentiment performed slightly better than fundamentals alone.
News may skew positive: The results suggest that news-derived inputs may push the model toward more positive ratings.
Short-term versus longer-term signals: News appears more useful for short-term predictions, while fundamentals appear more useful across the main 3-, 6- and 12-month horizons.
AI Street Patent Review Tracker
Paid subscribers can download the Excel file below, which uses AI-assisted review to identify 300 patent applications published this year that appear tied to AI in trading and investing.
