mrv-lib: Model Risk Validator

Your model might be producing different outputs depending on which features you feed it, which seed you use, or how you bin the data: and your current validation doesn't catch this. mrv-lib tests whether your model outputs are stable across admissible specification choices, or silently depend on arbitrary modelling decisions.

mrv is a pure validation library: you supply labels from your own models, mrv measures how stable they are. Bank model risk management (OCC Bulletin 2026-13 -- the 2026-04-17 Revised Model Risk Management Guidance that supersedes SR 11-7) is the anchor application; the same tests deploy equally to production ML monitoring (route to fallback or human-in-the-loop when labels are unstable, regardless of domain).

What it does

Also includes: a business impact function (impact_fn), disagreement attribution (LOO / frequency-pair / temporal), and a specification-invariance report generator (report(): result JSON to LaTeX to PDF) covering both the representation and resolution tests.

Install

pip install mrv-lib

Quick start

The recommended public Python API lives at the top level of the mrv package. You supply a model_fn (features to integer labels) plus the admissible set of specifications, and mrv returns a typed result. mrv only measures agreement; it never fits a model itself.

Representation invariance (Paper 1) across feature representations:

import numpy as np
import mrv

rng = np.random.default_rng(42)
n = 200
base = rng.integers(0, 3, n)
labels_a = base.copy()
labels_b = base.copy()
flip = rng.random(n) < 0.10
labels_b[flip] = rng.integers(0, 3, flip.sum())   # small perturbation
returns = rng.standard_normal(n) * 0.01

result = mrv.rep_invariance_validator(
    model_fn=lambda x: x,   # passthrough: supply pre-computed labels directly
    admissible_class={"vol+dd+var": labels_a, "vol+var+cvar": labels_b},
    returns=returns,        # optional: enables the Spearman ordering check
    K=3,                    # number of regime states
)
print(result.summary())
print("mean ARI:", result.mean_ari["asset"])
print("partition passes:", result.passes_partition["asset"])

Already fit your own regime model? Wrap the labels with the passthrough model_fn=lambda x: x as above, or pass a real callable that maps a feature matrix to integer labels.

Resolution invariance (Paper 2) across frequencies:

import pandas as pd
import numpy as np
import mrv

rng = np.random.default_rng(0)
idx = pd.date_range("2026-01-05 09:30", periods=480, freq="5min",
                    tz="America/New_York")
labels_5m  = pd.Series(rng.integers(0, 2, 480), index=idx, dtype=int)
labels_15m = labels_5m.iloc[::3].copy()

result = mrv.res_invariance_validator(
    model_fn=lambda s: s,   # passthrough: supply pre-computed labels per frequency
    resolution_set={"SPY": {"5m": labels_5m, "15m": labels_15m}},
    spec=mrv.ResolutionSpec(freqs=("5m", "15m"), intraday_freqs=("5m", "15m")),
    run_permutation=False,
)
print(result.summary())
print(result.ari_matrix["SPY"].round(3))
print("overall mean ARI:", result.overall_mean_ari["SPY"])

The typed results (RepInvarianceResult / ResInvarianceResult) expose .summary() plus attributes such as .ari_matrix, .overall_mean_ari, .passes_partition, and .intraday_overall_ari_gap. To feed a real model, pass a model_fn that fits your regime model and returns integer labels. See examples/paper1_representation_invariance.ipynb and examples/paper2_resolution_invariance.ipynb for end-to-end walkthroughs.

Logging

mrv-lib uses Python's standard logging module with hierarchical names (mrv.validator.rep, mrv.validator.res, etc.). By default nothing is emitted.

import logging

# Show all mrv INFO+ messages
logging.basicConfig(level=logging.INFO)

# Show DEBUG for the representation validator only
logging.getLogger("mrv.validator.rep").setLevel(logging.DEBUG)

# Route mrv logs to a file
handler = logging.FileHandler("mrv_run.log")
logging.getLogger("mrv").addHandler(handler)

See src/mrv/utils/log.py and src/mrv/default_config.yaml for the YAML-based logging configuration used by the convenience pipeline.

Project layout

mrv-lib/
|-- config.yaml              # Configuration (for convenience pipeline)
|-- examples/
|   |-- quickstart.ipynb
|   |-- paper1_representation_invariance.ipynb
|   |-- paper2_resolution_invariance.ipynb
|   `-- example_california_housing.ipynb
|-- src/mrv/
|   |-- invariance/          # Recommended public Python API + typed results (rep, res)
|   |-- pipeline.py          # Internal labels-first backend behind the `mrv` CLI
|   |-- data/                # Data modules (optional)
|   |   |-- reader.py        # CSV / OHLCV loading
|   |   |-- factors.py       # Factor / feature engineering
|   |   |-- normalize.py     # Normalization (rolling z-score, minmax)
|   |   `-- download_yahoo.py # Yahoo Finance data download
|   |-- models/              # GMM/HMM fitting
|   |-- templates/
|   |   `-- template.tex     # Specification-invariance report template (rep + res)
|   |-- validator/
|   |   |-- base.py          # BaseValidator (subclass for custom tests)
|   |   |-- rep.py           # Representation Invariance (Paper 1)
|   |   |-- res.py           # Resolution Invariance (Paper 2)
|   |   |-- metrics.py       # ARI, AMI, NMI, Spearman, VI
|   |   |-- attribution.py   # LOO, frequency-pair, temporal hotspots
|   |   `-- report.py        # JSON -> LaTeX -> PDF
|   `-- utils/
|       |-- config.py        # YAML config loading
|       |-- download_ib.py   # IB data download
|       `-- log.py           # Logging setup
|-- reports/                  # Output (gitignored)
`-- tests/

Output

Each run creates a timestamped directory under reports/:

• result.json -- Complete data (reusable for report regeneration)
• report.pdf -- Report with cover page, dashboard, heatmaps, and remediation plan
• summary.txt -- Plain text quick view
• {asset}_ari_heatmap.png -- ARI heatmap per asset
• {asset}_timeline.png -- Regime timeline (res validator)
• pipeline_summary.csv -- Summary metrics per asset

Research

Based on the following PhD research:

• Zheng, Low & Wang (2026). Regime Labels Are Not Representation-Invariant (Paper 1). Finance Research Letters.
• Zheng, Low & Wang (2026). Regime Labels Are Not Resolution-Invariant (Paper 2). Finance Research Letters.

License

Dual-licensed.

• Open source: GNU Affero General Public License v3.0 or later (AGPL-3.0-or-later). See LICENSE. Free for academic research, teaching, and personal use. Note that the AGPL's network-use clause requires any modified version offered over a network to also offer its complete source.
• Commercial: Organizations that wish to use mrv-lib in proprietary or closed-source systems, or otherwise cannot meet the AGPL obligations, require a separate commercial license. See COMMERCIAL-LICENSE.md.

Maintainers

ModelGuard Lab -- Author: Kai Zheng.