Live demo. Data is synthetic and labelled as such in-app.
An XGBoost pipeline that predicts IPO listing-day returns — feature engineering, time-series CV, calibrated confidence bands. No API key; seeds and trains itself on first boot.
▶ Live demo: https://siddharthgaur1-siddharthipo-gmp-srcapp-htnhfl.streamlit.app/ — clickable with no API key (Dashboard, Predictor, Backtest, Model Info tabs on synthetic data). (First load may take ~30s if the app is asleep.)
⚠️ The dataset is synthetic. Every metric here is a real cross-validation result on synthetic data — it demonstrates the ML pipeline, not real-market predictive power. Details in Data.
git clone https://github.com/siddharthgaur1/ipo-gmp
cd ipo-gmp
pip install -r requirements.txt
streamlit run src/app.py # seeds data + trains models on first boot (~2s)No API key, no external data, no cloud. Measured CV results (synthetic data): MAE ≈ 12.7%, R² ≈ 0.87, directional accuracy ≈ 91%, confidence band ±4.7%. Security notes: SECURITY.md.
Predicts Indian IPO listing-day returns from pre-listing Grey Market Premium (GMP) and subscription data. GMP alone is the strongest available pre-listing signal but noisy on its own; this combines it with subscription mix (QIB/NII/ retail), GMP momentum, deal size, and sector to output a point estimate, confidence band, and a STRONG BUY/BUY/NEUTRAL/AVOID signal, plus a dashboard for exploring historical patterns and backtesting the signal.
The dataset is entirely synthetic, generated by scripts/seed_data.py.
listing_gain_pct (the model's prediction target) is constructed directly as
gmp_pct * random_factor + noise — the generator hand-encodes the exact
relationship the model then learns to recover. The accuracy numbers in this
README are real outputs of this codebase, and the feature-engineering and
time-series-CV methodology are both sound, but none of it has been
validated against real historical IPO outcomes. Treat this as a
demonstration of the modeling pipeline, not evidence the strategy works on
real markets. See "What I'd improve" for the actual next step.
scripts/seed_data.py (synthetic IPO generator, seed=42)
│
▼
data/ipo_gmp.db (SQLite, 800 IPOs)
│
▼
src/model.py
├── engineer_features() GMP momentum, QIB/NII ratio, log-scaled size, sector encoding
├── train() 5-fold TimeSeriesSplit CV → final fit on all data
│ ├── XGBRegressor predicts listing_gain_pct
│ └── XGBClassifier predicts P(listing_gain_pct > 0)
├── _backtest() rolling monthly signal vs. all-IPO baseline
└── predict_one() single-IPO inference + confidence band
│
▼
models/{regressor,classifier}.pkl, models/meta.json (CV metrics, importances, backtest)
│
▼
src/app.py (Streamlit, 5 tabs: Dashboard / Predictor / Backtest / IPO Table / Model Info)
| Choice | Why |
|---|---|
| XGBoost (regressor + classifier) | Tabular data with nonlinear feature interactions (e.g. GMP momentum only matters conditional on subscription level) — gradient boosting handles that natively without manual interaction terms. |
| Two separate models (regression + classification) | A point estimate and a direction probability answer different questions — "how much" vs. "how confident am I it's positive" — and XGBoost's classifier gives a better-calibrated probability than deriving one from regression residuals. |
TimeSeriesSplit (not random k-fold) |
IPO market conditions drift over time; random k-fold would leak future market regime information into past-fold training and overstate accuracy. |
| SQLite | Single-file, zero-setup, sufficient for an 800-row dataset that's read far more than written. |
| Streamlit | Fastest path from a trained model to an interactive, shareable dashboard — no separate frontend needed for this scope. |
| Confidence band = ±1 residual std | Cheap, interpretable uncertainty estimate without the complexity of quantile regression or conformal prediction — adequate for a portfolio-stage signal, not a production trading system. |
Python 3.11+. No API keys or external services — everything runs offline against the synthetic dataset.
pip install -r requirements.txt
python scripts/seed_data.py # generates data/ipo_gmp.db (deterministic, seed=42)
python src/model.py # trains + saves models/*.pkl, models/meta.json (~10s)streamlit run src/app.py
pytest tests/ -v # 19 testsGMP momentum (day-1 → final), not just final GMP. A GMP that rose from
₹20 to ₹80 reads as a stronger bullish signal than one that started at ₹100
and fell to ₹80, even though they end at the same level — momentum is
(gmp_final - gmp_day1) / |gmp_day1|, clipped to ±5 to bound the effect of
near-zero day-1 GMP.
QIB/NII subscription ratio as a feature, not raw subscription counts. Institutional (QIB) demand and leveraged HNI (NII) demand say different things about an IPO's quality — a high QIB/NII ratio suggests informed institutional conviction rather than leveraged retail-adjacent speculation. Modeling the ratio directly, rather than leaving XGBoost to infer it from the two raw columns, made the signal show up cleanly in feature importance.
Backtest reuses the training data (explicitly flagged as in-sample in the UI). A true walk-forward backtest — refitting the model at each point in time using only data available then — is the honest version of this and isn't implemented yet; see "What I'd improve."
5-fold TimeSeriesSplit cross-validation, 800 IPOs:
| Metric | Value |
|---|---|
| Directional accuracy | 91.2% |
| Mean absolute error | 12.7% |
| R² | 0.871 |
| Backtest: signal avg return | 55.2% vs. 34.0% all-IPO baseline |
Feature importance is dominated by gmp_pct (59%) and gmp_rs (17%) —
expected, since those two are the closest proxies to how the synthetic
target was generated. Everything past the top two contributes single-digit
percentages.
Beyond the synthetic-data caveat above (which applies to every number in this README): see "Limitations" below for the specific gaps in this build.
- Real data. This is the actual gap, not a nice-to-have — replace
seed_data.pywith scraped/purchased historical IPO GMP + subscription + listing data (e.g. Chittorgarh, Groww IPO pages) and re-run the identical pipeline. Every metric above should be treated as provisional until this happens. - True walk-forward backtest instead of reusing training data — refit at each historical point using only prior information, matching how the model would actually have been deployed.
- Quantile regression or conformal prediction for the confidence band instead of a flat ±1 residual std, which assumes symmetric, homoscedastic errors that IPO returns (fat-tailed, especially for SME) probably don't have.
- Feature: sector-relative GMP (an IPO's GMP% vs. its sector's trailing average) — right now sector is only a flat encoded feature, not interacted with GMP.
- llm-regression-detector — CI/CD regression detection for an LLM classifier's eval suite.
- rag-hybrid-search — hybrid dense+BM25 RAG pipeline with citation verification.