We trained a compact Qwen3-32B forecaster to estimate whether disclosed SEC filing risks would materialize within a specified horizon. The paper reports that the trained model outperforms GPT-5 on probabilistic accuracy and calibration.
What we did
- Converted SEC risk disclosures into firm-specific, time-bounded forecasting questions.
- Resolved each risk query against later public filings and outcomes.
- Fine-tuned Qwen3-32B and evaluated it against GPT-5 and baselines.
Read more
Primary artifacts for this case study.
Blog post blog.lightningrod.ai
AI-Driven SEC Risk Prediction: Separating Signal from Noise
How to train a model to predict SEC enforcement from public filings
Paper arxiv.org
Foresight Learning for SEC Risk Prediction
Paper describing the SEC risk-query generation, training, and evaluation pipeline
Dataset huggingface.co
SEC risk questions test set
Evaluation set of SEC filing questions with resolved outcomes