Modular NLP pipeline to download, preprocess, and classify SEC 10-K filings. Combines TF-IDF + logistic regression and fine-tuned BERT with GPT-3.5-assisted labeling to extract financial highlights, risks, products, and AI mentions. Open-source, scalable, and designed for transparent financial text analysis.
This README document explains the workflow of the SEC 10-K NLP pipeline, including how the program works, required edits, and commands to run each step.
-
Download 10-K Filings What it does: Downloads raw 10-K HTML filings directly from the SEC EDGAR database and saves them locally in the data/ folder. File: notebooks/download_10k.py What to edit: Edit the companies list near the top of the file. For each company, specify: { "name": "Company Name", "url": "https://www.sec.gov/Archives/edgar/data/XXX/filing.htm", "save_path": "data/company_10k.html" } Run command: python notebooks/download_10k.py
-
Label Paragraphs with GPT-3.5 What it does: Automatically labels paragraphs in downloaded 10-Ks using OpenAI’s GPT-3.5, categorizing them into financial highlights, risks, products, AI mentions, or other. File: notebooks/ai_label.py What to edit: Update the files list at the top of the file to include paths to downloaded 10-K HTMLs, e.g. "data/company_10k.html".
Run command: python notebooks/ai_label.py Output: Labeled JSON files saved in ai_labeled/.
-
Train Logistic Regression Classifier What it does: Trains a TF-IDF vectorizer and logistic regression model on labeled data. File: notebooks/train_logistics.py What to edit: Ensure your combined manual and GPT-labeled data is saved as data/labeled_data.json.
Run command: python notebooks/train_logistics.py Output: Model and vectorizer saved in models/logistic_model.pkl and models/vectorizer.pkl.
-
Batch Inference (Logistic + BERT) What it does: Runs both logistic regression and fine-tuned BERT models on paragraphs from a 10-K file. File: notebooks/batch_predict.py What to edit: Change the filepath variable to point to the desired 10-K HTML in data/.
Run command: python notebooks/batch_predict.py Output: Predictions saved as JSON in predictions/.
-
Logistic-Only Prediction & Evaluation What it does: Runs logistic regression predictions on a 10-K and evaluates against ground truth if available. File: notebooks/predict_logistics.py What to edit: Change new_10k_path to your target 10-K HTML file.
Run command: python notebooks/predict_logistics.py Output: Predictions saved in predicted_labels/, evaluation metrics printed if ground truth exists.
-
Evaluation of Predictions What it does: Compares predicted labels to ground truth labels and prints classification metrics. File: notebooks/calc_scores.py
Run command: python notebooks/calc_scores.py
-
Interactive Paragraph Classification What it does: Allows manual input of paragraphs to classify in real-time using both logistic and BERT models. File: notebooks/predict.py
Run command: python notebooks/predict.py **Additional Notes
- Make sure to have your OpenAI API key set in your environment for GPT labeling (ai_label.py).
- Ensure dependencies (transformers, scikit-learn, torch, requests) are installed.
- For BERT inference, GPU use is optional but recommended for speed.
- Always keep your labeled data and models synchronized when retraining or evaluating.