quant-platform

A modular Python platform for systematic quantitative trading research — from point-in-time data ingestion, through an Advances in Financial Machine Learning (AFML) research library, to a backtester and a paper/live execution engine that share one code path.

Built as a uv-managed monorepo of five independently-tested packages with a strict, one-directional dependency graph.

Architecture

                ┌─────────────┐
                │  quantlake  │   point-in-time data substrate
                └──────┬──────┘   (bitemporal store, security master,
                       │           survivorship-complete universe, as-of joins)
                ┌──────▼──────┐
                │  quantdata  │   ingestion · catalog · corporate-action
                └──────┬──────┘   adjustment · survivorship-free panels
                       │
                ┌──────▼──────┐
                │  quantcore  │   AFML research library (the core)
                └──┬───────┬──┘
                   │       │
        ┌──────────▼──┐ ┌──▼───────────┐
        │  quantstrat │ │  quantengine │
        │ backtesting │ │  execution   │
        └─────────────┘ └──────────────┘

Dependencies flow downward only: quantengine and quantstrat build on quantcore; quantcore builds on quantdata/quantlake. Nothing upstream imports anything downstream.

Packages

Package	Role
quantcore	AFML research library: information-driven bars, fractional differentiation, triple-barrier labels, meta-labeling, purged K-fold CV, sample weights, bet sizing, conformal prediction, and portfolio allocation (HRP/NCO).
quantdata	Data layer: ingestion, catalog, corporate-action adjustment, and survivorship-free point-in-time panels.
quantlake	PIT cross-sectional data substrate: bitemporal store, security master, survivorship-complete universe/delistings, as-of joins, and a feature store.
quantstrat	Research backtesting: walk-forward engine, cross-sectional portfolio construction, transaction-cost and performance accounting, tearsheets.
quantengine	Execution layer: consumes signals from quantcore and emits broker orders with full accounting and paper/live parity.

Design principles

• Replay is live. quantengine's backtest replay runs the same execution loop as paper/live trading on a historical clock — if they diverge on identical inputs, replay is the bug. No separate vectorized backtester to drift out of sync.
• Leakage-aware by construction. Purged, embargoed walk-forward CV; point-in-time joins; survivorship-free panels. Lookahead is designed out, not patched after.
• AFML canon, tested. The research primitives follow López de Prado's Advances in Financial Machine Learning, each with its own test module.

Engineering

• ~1,650 tests (pytest) across the five packages — quantcore alone carries the bulk, with property-based (hypothesis) and parity/regression suites.
• Strictly typed — basedpyright in strict mode; ruff for lint + format.
• Reproducible — uv workspaces with pinned lockfiles per package.
• Python 3.11+.

Layout

quant-platform/
├── quantcore/     # research library  (src/ + tests/)
├── quantdata/     # data ingestion / adjustment
├── quantlake/     # PIT data substrate
├── quantstrat/    # backtesting
└── quantengine/   # execution

Each package is self-contained: cd <package> && uv sync && uv run pytest.

License & use

© 2026 Jules Verdez. All rights reserved.

This repository is published for portfolio and evaluation purposes. You are welcome to read and clone it to assess the author's work. No license is granted to use, modify, redistribute, or deploy the code, in whole or in part. It ships with no warranty and contains no proprietary data, credentials, or trading signals — only the engineering.