conformal-finance 0.1.1

Uncertainty quantification for financial time series via conformal prediction

Conformal Finance

Distribution-free uncertainty quantification for financial time series.

Mission: Replace meaningless RMSE metrics with calibrated prediction intervals that maintain valid coverage even when financial regimes shift.

Finance ML usually reports point forecasts. Point forecasts hide tail risk. Conformal prediction gives non-parametric, finite-sample coverage guarantees — but almost no public code targets non-stationary financial data. ConformalFinance does exactly that.

What you can do with it

• Wrap any point forecast in valid prediction intervals using split, CQR, or rolling conformal prediction.
• Adapt to distribution shift online with adaptive conformal inference (ACI).
• Forecast returns, realized volatility, and VaR with calibrated intervals.
• Diagnose empirical coverage, interval width, and conditional coverage by regime.
• Load real market data (optional yfinance) or realistic synthetic AR/GARCH series.
• Visualize coverage and width traces (optional matplotlib/plotly).

Install

pip install conformal-finance

For development, docs, and optional data / plotting support:

pip install -e ".[dev]"

Quickstart

import pandas as pd
from conformal_finance.conformal.split import SplitConformalPredictor
from conformal_finance.data.synthetic import generate_ar_garch_returns

# Realistic synthetic returns
returns = generate_ar_garch_returns(n=2000, seed=42)

# Build a trivial point forecast: yesterday's return
y_true = returns.iloc[1:].reset_index(drop=True)
y_pred = returns.iloc[:-1].reset_index(drop=True)

# Split conformal intervals
cp = SplitConformalPredictor(alpha=0.1)
cp.fit(y_true_cal=y_true.iloc[:500], y_pred_cal=y_pred.iloc[:500])
intervals = cp.predict(y_pred_test=y_pred.iloc[500:])

print(intervals.head())

Design

• Core conformal methods are implemented from scratch so the API stays stable and lightweight.
• Optional extras (yfinance, matplotlib, plotly) are imported lazily and never required for import.
• scikit-learn-style interface: every predictor exposes .fit() / .predict() or equivalent streaming methods.
• Type hints and numpy/pandas idioms throughout.

Project layout

conformal_finance/
├── conformal/       # Split, CQR, ACI, Rolling conformal predictors
├── finance/         # Returns, volatility, VaR conformalizers
├── diagnostics/     # Coverage, width, and regime diagnostics
├── data/            # Synthetic AR/GARCH generators and yfinance loader
└── viz/             # Coverage/width plotting helpers

Documentation

Full docs are built with MkDocs and hosted at https://satyamdas03.github.io/conformal-finance/.

Local build:

mkdocs serve

Roadmap

Version	Focus	Status
v0.1.0	Split, CQR, ACI, rolling; returns, vol, VaR; diagnostics	In progress
v0.2.0	Realized ES, drawdown, and factor-return targets	Planned
v0.3.0	GARCH / historical-simulation benchmarks	Planned
v0.4.0	Portfolio sizing and vol targeting	Planned
v0.5.0	Research note + integration with FactorForge	Planned

License

MIT © 2026 Satyam Das