mrv-lib | Model validation library for financial AI

What it is

A model validation library

Runs specification-invariance validators that test whether a model's regime labels survive an equivalent change of inputs.

How it works

Refit, compare, gate

Refit under equivalent inputs, measure agreement, and gate the model when it disagrees with itself.

Why it is required

Model risk regulation

OCC 2026-13, Basel, and the EU AI Act require evidence a model is sound. This produces it.

Methodology

Inference Collapse in AI Models

Theory, Evaluation, and Fail-Safe Design for Financial Systems. Two published results, each shipped as a validator.

Paper 1 · validate_repPublished

Regime labels are not representation-invariant

Kai Zheng, Rand Low, Ruili Wang · Finance Research Letters (2026) · DOI 10.1016/j.frl.2026.110383

Swap economically equivalent risk features and the regime labels change: cross-representation ARI 0.34 to 0.43, below the 0.65 threshold.

Paper 2 · validate_resAccepted

Regime labels are not resolution-invariant

Kai Zheng, Rand Low, Ruili Wang · Finance Research Letters (accepted) · SSRN 10.2139/ssrn.6531002

Regime labels disagree across data frequencies (5m / 1h / 1d).

Examples

From install to a signed report in three commands.

Run from the CLI, or drive the pipeline from Python. Both produce the same verdict and PDF.

CLI

# install
pip install mrv-lib

# download data, then validate + report
mrv download config.yaml
mrv run config.yaml rep      # representation
mrv run config.yaml res      # resolution
mrv report                   # regenerate PDF

Python

from mrv.pipeline import run, download, validate_rep, report
from mrv.utils.config import load

download("config.yaml")
run("config.yaml", "rep")    # validate + PDF report

# or call a validator directly
cfg = load("config.yaml")
verdict = validate_rep(labels=labels, cfg=cfg)
report(verdict["json_path"], cfg=cfg)

Each run returns a plain dict (verdict, ari, spearman, json_path) and a PDF report. Interactive notebook (no market-data account needed).

Regulations

The mandates that require model validation.

Every major supervisor requires evidence that a model is sound under equivalent inputs. mrv-lib produces it. It supports these obligations; it does not certify compliance.

Jurisdiction	Regulation	Requirement	Source
🇺🇸 US	OCC Bulletin 2026-13 (SR 26-2)	Model risk management; supersedes SR 11-7	occ.gov
🌐 Intl	Basel FRTB (BCBS d457)	Robustness of internal market-risk models	bis.org
🇪🇺 EU	EU AI Act, Art 15	Accuracy and robustness of high-risk AI	eur-lex.europa.eu
🇬🇧 UK	PRA SS1/23	Model risk management principles for banks	bankofengland.co.uk
🇺🇸 US	NIST AI RMF 1.0	MEASURE function for AI risk	nist.gov
🇳🇦 NZ	RBNZ BS2A	Model validation independence, monitoring	rbnz.govt.nz
🇸🇬 SG	MAS AI MRM Guidelines	AI model risk management for FIs	mas.gov.sg
🇦🇺 AU	APRA CPS 220	Risk management, model oversight	apra.gov.au

mrv-lib provides a quantitative diagnostic that makes the judgements in these standards more defensible and auditable. It does not certify that a model meets any requirement, replace a qualified validator, or provide a legal opinion.

Roadmap

From invariance tests to a governance layer.

Released capability tracks published research; what comes next is a direction, not a promise. Current release v0.6.1.

now

Invariance validators Released

Representation (validate_rep) and resolution (validate_res) tests. Papers 1 and 2.

Fail-safe governance Direction

Turn the two-layer invariance verdict into a gate that flags or suspends a model before an unstable regime label drives a live decision.