IPO Underpricing Prediction — ML Capstone - Live Demo

CS 0451 Machine Learning · Middlebury College · Spring 2026

A machine learning pipeline that predicts whether an IPO will be underpriced, fairly priced, or overpriced before opening day — combining financial data, macroeconomic indicators, and NLP-extracted signals from SEC S-1 filings.

---

The Problem

IPO underpricing costs companies billions annually. When Airbnb went public, its stock opened at nearly double its offer price — leaving over $3 billion on the table. On average, IPOs are underpriced by 10–20%. This project asks: can ML detect the patterns behind mispricing before trading begins?

---

Data Sources

Source	Features
Bloomberg Terminal	Offer price, deal size, shares outstanding, industry, lead underwriter
FRED API	Interest rates, CPI, unemployment — matched to each IPO date
SEC EDGAR S-1 Filings	NLP-extracted risk scores, growth signals, business complexity via GPT-4o-mini

Integrating structured financial data, macroeconomic time series, and unstructured regulatory text into a single feature pipeline was the core engineering challenge of this project.

---

Results

Binary Classification (Full Dataset, n=6,110)

Model	Accuracy	Macro F1
Logistic Regression	58%	0.58
Neural Network	69%	0.64
Random Forest	70%	0.65
XGBoost	71%	0.70

XGBoost beat the naive baseline (always predicting "underpriced" = 60% accuracy) by 11 points and was the most balanced across classes. A naive baseline that always predicts "underpriced" would achieve 60% accuracy without learning anything useful — Macro F1 tells the real story.

3-Class Classification (Full Dataset)

Model	Accuracy	Macro F1
Logistic Regression	50%	0.45
Neural Network	53%	0.51
Random Forest	63%	0.49
XGBoost	60%	0.55

Top Features (XGBoost — Binary)

Offer Price dominated with an importance score of ~0.18, nearly double the next feature. Market Cap at Offer and GDP followed at ~0.09 each, with CPI and bulge-bracket underwriter status rounding out the top 5 — confirming that pricing decisions and the macro environment matter more than firm-specific characteristics.

Risk Feature Impact (XGBoost on SEC S-1 Subset, n=862)

Adding GPT-4o-mini risk scores from S-1 filings produced small, consistent gains across precision, recall, and F1 — but no single improvement exceeded 3 percentage points. A larger dataset may reveal stronger effects.

📓 Full confusion matrices, feature importance plots, and per-class breakdowns in report.ipynb

---

Key Technical Highlights

• End-to-end feature pipeline combining structured financial data, macroeconomic time series, and unstructured text
• NLP feature extraction from SEC S-1 filings using GPT-4o-mini to score regulatory risk, competitive risk, financial risk, and overall risk on a 1–10 scale
• Four model types — Logistic Regression (custom PyTorch), Neural Network (custom PyTorch MLP), Random Forest (custom from scratch), and XGBoost — all trained and evaluated on identical 80/20 stratified splits
• Class imbalance handling via class weights across all models (60% of IPOs underpriced in full dataset)
• Feature importance analysis identifying offer price, market cap, and macro conditions as the strongest predictors

---

Project Structure

ipologists/
├── data/               # Raw and processed datasets
├── notebooks/          # Per-model Jupyter notebooks
├── scripts/            # Data collection (Bloomberg, FRED, SEC EDGAR)
├── src/                # Core pipeline modules and saved models
├── report.ipynb        # Full analysis, results, and visualizations
└── requirements.txt

---

Setup

pip install -r requirements.txt

Mac only (required for XGBoost parallel processing):

brew install libomp

Create a .env file with your API keys:

FRED_API_KEY=your_key
OPENAI_API_KEY=your_key

---

Quick Demo

import pickle, pandas as pd
from src.pipeline import features, prepare_xgboost

full = pd.read_csv("data/final/dataset_full.csv")
with open("src/models/saved_models/xgBoost/xgb_binary_full.pkl", "rb") as f:
    model = pickle.load(f)

_, X_test, _, y_test = prepare_xgboost(full, "underpriced", features)
preds = model.predict(X_test)  # 71% accuracy on held-out test set

---

Authors

William Zambito · Anna Kester
Middlebury College — Computer Science & Economics

---

For educational and research purposes only. Not financial advice.