Finance · Risk and validation teams
Is your backtest grading the model, or the reference inside the test?
A team that defends a VaR or expected-shortfall number reports a pass or a fail. That verdict can depend on a choice inside the test that nobody recorded as a modelling choice.
EvidenceAcross 5 models at 2 confidence levels, the conditional expected-shortfall test passes 0 of 10 under a Gaussian reference, 6 as shipped and 10 under a Student-t reference with 5 degrees of freedom. The models, the data, the VaR and the ES are the same in all 3 cases. Sample: 2,172 trading days on one three-asset book.
IllustrationA parametric 99% VaR was breached on 2.07% of days [1.43%, 2.77%]. On an assumed EUR 100 million book, the resulting addend raised the VaR component of the own funds requirement by EUR 1.6m on average [EUR 0.6m, EUR 2.6m]. It applies only to firms with internal-model permission, under the rule in force until 1 January 2027.
Evidence
- Specimen report Model Health Check: a three-asset book The finished deliverable, with the rule text and the working.
- One-page brief Tail-risk backtesting, in one page For a committee or a validation lead.
- Paper Tail Risk Backtesting: When the Test Improves and the Model Does Not KW-2026-04, for the method.
Offer
Model Health Check
- Input
- One CSV: date, VaR, expected shortfall if stated, and the P&L it is tested against. 2 years of daily history or more, plus the specification of the test you run.
- You get
- Breach frequency and clustering with intervals, the supervisory zone record, the ES verdict re-scored against alternative references, and a findings register.
- Timing
- 3 weeks from data delivery
- Reference price
- Fixed, set by the number of models. About EUR 6,500 for one model with its backtest.
Try it
The Band Checker counts breaches against a stated VaR level in your browser, on your own CSV or on the bundled finance sample (750 days of synthetic P&L against a 99% VaR), and your file never leaves your computer.
First step
Send one paragraph describing the series or the records to matthew@mtanti.com, and you will have a quote within one working day. A standard mutual NDA comes before any data moves, and your data is deleted on delivery and never used elsewhere.
What this does not support
One implementation on one three-asset book of public prices. It does not show that your backtest has this problem, only how to find out, and it does not show that the Acerbi-Székely tests are bad tests. The euro figure is an illustration on an assumed book: no real firm paid it, only the VaR component is computed, and the rule it rests on changes on 1 January 2027. Nothing here is a certification or a regulatory opinion.