NVIDIA NeMo generates synthetic financial data for AI

NVIDIA announced an approach to synthetic generation of financial data to improve fine-tuning of LLMs in the financial sector. The development addresses a fundamental problem: real financial data is systematically unbalanced by event types.

Why Financial Data Ends Up Unbalanced

When companies and researchers train language models on financial texts, they face a skewed distribution of events. Financial news is flooded with information about quarterly earnings and stock price movements, while rarer and more specific events occur in insufficient volume.

• Overabundance: quarterly reports, stock price movements
• Scarcity: credit rating changes, product approvals, labor conflicts
• Result: models become overtrained to predict frequent events but respond poorly to rare ones

This creates a serious problem for algorithmic trading and risk management, where missing a rare but significant signal can lead to massive losses.

How Synthetic Data Closes the Gap

NVIDIA proposes using synthetic generation to create missing training examples. The approach allows targeted augmentation of the dataset with events that rarely occur in the real stream of financial news.

This does not mean completely fictional training — it is about generating controlled examples for underrepresented categories. Synthetic data is created with quality checks and alignment with real financial scenarios.

Where It's Applied

Financial LLMs are rapidly becoming standard in the financial sector. NVIDIA highlights three key directions for applying synthetic financial data: trading research (models for news analysis and market movement prediction), risk modeling (assessing scenario impact on portfolio or balance sheet) and surveillance (detecting market anomalies and potential manipulation).

Banks and hedge funds are investing significant resources in AI systems for real-time analysis of market data. However, model quality depends entirely on training data. Imbalance in the dataset leads to biased forecasts and missed signals. NVIDIA's approach eliminates the need to collect multi-year archives of rare events.

What It Means

Financial AI is gradually transitioning to synthetic data as a tool for overcoming practical limitations of real datasets. This opens the path to more reliable and balanced models in trading and risk management — areas where errors are very expensive.