HomeResearch › Research reproducibility

Research reproducibility

Every QuantMedia research claim that can reasonably be reproduced should be. This page indexes the implementations, example data and tools that back the research — and is equally explicit about which papers are currently research only, with no runnable material behind them.

Status definitions

Reproducible — runnable code, example data, expected output and tests exist, and the implementation follows the methodology in the paper. Interactive tool — a working calculator with a published formula. Research only — the paper stands on its own; no code has been released yet. Nothing is labelled reproducible on the strength of an intention.

Available implementations

VPIN — order flow toxicity

Research questionCan order-flow toxicity be estimated from a trade tape without quote data, and how sensitive is the answer to the classification method?
PaperVPIN & Order Flow Toxicity
ExplainerWhat is VPIN?
Codequantmedia-research/vpin-order-flow-toxicity/vpin.py, example.py
ImplementsEqual-volume bucketing with boundary splitting; Bulk Volume Classification (Student-t) and the tick rule; rolling VPIN
Example dataSynthetic tape, fixed seed 20260808, with a planted one-sided episode
Expected outputVPIN rises ~1.94x through the planted episode; BVC and tick-rule means differ (0.44 vs 0.18) on identical data
Tests15, covering bucket construction, volume conservation, both classifiers, bounds, degenerate tapes and validation
StatusReproducible

Hierarchical Risk Parity

Research questionDoes avoiding covariance-matrix inversion produce more stable out-of-sample allocations than mean-variance on correlated universes?
PaperHierarchical Risk Parity
ExplainerHRP vs mean-variance
Codequantmedia-research/hierarchical-risk-parity/hrp.py, compare_mvo.py
ImplementsCorrelation distance, hierarchical linkage, quasi-diagonalisation, recursive bisection; min-variance and shrinkage baselines
Example dataSynthetic 20-asset block-correlated panel, fixed seed 20260808
Expected outputOut-of-sample volatility drift: HRP +1.3% vs MinVar +13.6%; shrinkage narrows it to +7.7%
Tests13, including weights summing to 1, no negatives, a known 80/20 two-asset result and distance-metric properties
StatusReproducible

Probabilistic Sharpe Ratio

Research questionGiven a track record's length and return shape, how confident can you be that the true Sharpe beats a benchmark?
PaperProbabilistic Sharpe Ratio
ExplainerWhat is the Probabilistic Sharpe Ratio?
ToolPSR calculator — runs in the browser, no data leaves the page
VerificationWorked example published with every intermediate value, so the tool can be checked against an independent implementation
StatusInteractive tool

Research only

These papers have no released implementation. They are listed so the absence is explicit rather than inferred:

Running the code

cd quantmedia-research/vpin-order-flow-toxicity pip install -r requirements.txt python example.py cd ../hierarchical-risk-parity pip install -r requirements.txt python compare_mvo.py cd .. python tests/test_vpin.py # 15 tests python tests/test_hrp.py # 13 tests

Python 3.9 or later. Dependencies are numpy, pandas and scipy. No API key, no network access and no paid data subscription is required to run any of it.

On the sample data

Both packages ship synthetic example data generated from fixed seeds. It is not real market data and no conclusion about any real security follows from it. It exists so the implementations can be run and verified end-to-end without a tick-data subscription. Each package states this in its README, its module docstring and its console output.

Proprietary data

Separately from the research code, QuantMedia publishes two metrics computed from its own daily scan, with machine-readable history:

Not implemented, and why

Three things that would fit this site are deliberately absent:

Methodology versioning

The signal engine carries a version number, currently 2.0 (effective 2026-04-14), stamped into every scan output and history record. When production logic changes the version increments; historical records keep the version that produced them and are never retroactively rewritten. Papers describing earlier methodology are preserved as published rather than edited to match current production.