Verification Reports: Quant Finance Code Checked Against Ground Truth
A verification report is evidence that an implementation of a quantitative method does what the paper says — produced by running the code against inputs whose correct answer is known in advance, and recording what it got right, what it got wrong, and what was fixed.
The method behind every report, eight steps with the defect each one caught, is written up in How to verify a quantitative finance implementation.
Language models now produce a plausible implementation of almost any finance paper in seconds. What they cannot do is tell you whether it is correct. The first report below documents a VPIN implementation that returned zero toxicity for a perfectly one-sided tape — the opposite of the right answer — and looked fine until it was tested. As generating code becomes free, proving it correct is the part that keeps its value.
Reports
| Method | Record | Verdict | Defects found |
|---|---|---|---|
| VPIN — order-flow toxicity | v1.0 · 2026-08 | Passes after one fix | 1 — degenerate tape |
| Hierarchical Risk Parity | v1.0 · 2026-08 | Passes after two fixes | 2 — pandas 3 view; symmetry |
| Probabilistic Sharpe Ratio | v1.0 · 2026-08 | Published example was wrong; corrected | 1 — kurtosis term |
What a report contains
- Ground truth. How an input with a known correct answer was constructed — usually synthetic, from a fixed seed, because real market data never tells you what the right answer was.
- What was run. The exact commands, versions and seed.
- Results. The numbers the code produced, unedited.
- Defects. What was wrong, how it was detected, how it was fixed, and the regression test that now guards it.
- Limitations. What the verification does not establish. Passing on synthetic data proves the mechanism, not the market.
- Reproduce. Enough to get the same output on your machine.
Check it yourself, in one command
python quantmedia-research/verify_examples.py re-runs both
packages in a clean temporary directory and asserts that every shipped output
file matches to nine decimal places. If it passes on your machine, you hold the
same evidence this page describes. The runner and both packages are in the
public repository.
Verify your own implementation
If you have written or generated your own VPIN, HRP or PSR code, point it at the synthetic inputs described in each report and compare. If you would like a report written against your implementation — the same ground truth, the same discipline, a dated document you can cite — get in touch. That is the service this site is being built around.