Four audits, each run on your own data.

Every engagement below has the same shape. Somewhere in your operation a number was stated as a range, a band or a limit, and there is a record of what actually happened. The audit scores the stated claim against that record, strictly out-of-sample and by regime rather than in aggregate, and prices the misses in your own units. It runs on your infrastructure if you prefer. Your data is deleted on delivery and never used elsewhere, and a standard mutual NDA comes before anything moves.

Vendor Band Audit · the first step

One series and one stated band, audited end to end, so you can see whether the range you are quoted is true before committing to anything larger. It is the smallest useful engagement and the one to start with. A specimen of exactly this deliverable, run on a public German series, is published in full: read the specimen.

Its first two sections, series integrity and coverage against the stated level, you can run yourself on your own file with the Band Checker. It works in your browser, and your file never leaves your computer.

Useful for: any team that consumes a range someone else produced. Energy trading and scheduling desks, financial risk and treasury teams, demand and supply planners, logistics and operations control rooms.

Examples: a TSO's or vendor's day-ahead forecast band; a risk engine's VaR or expected-shortfall band; a demand-planning system's confidence interval; a carrier's delivery-time window; a weather service's prediction interval.

Input: one CSV with timestamp, forecast, actual, and the stated band if there is one. Twelve months or more.

You get: coverage against the stated promise, split by regime and by season; the width-tail accounting; a monetised shortfall estimate at your market's prices; a written verdict register; a 45-minute readout.

Timing: two weeks from data delivery.

Price: fixed, €3000.

Forecast Calibration Audit

The full portfolio version: every forecast you rely on, your own or a vendor's, audited exactly as the German grid paper does it. The question is the same for every series: when the forecast implied a range of outcomes, was the range honest, and did it stay honest in the regime where a miss is expensive?

Useful for: energy and utilities, financial risk and trading, retail and supply chain, logistics, insurance, and any operation that plans against a forecast with a stated range.

Examples: day-ahead load, solar, wind and price forecasts (the case in the grid paper); VaR and expected-shortfall backtests; demand forecasts feeding inventory and staffing; arrival and lead-time estimates; claims-frequency and loss forecasts.

Input: one CSV per series (timestamp, forecast, actual, stated band if any). Twelve months or more.

You get: the bias clock and coverage ladder for your portfolio; coverage by regime and by season; the width-tail accounting; a monetised shortfall estimate against your market's prices; an adaptive recalibration demonstration on your own data; a 45-minute readout.

Timing: two weeks from data delivery.

Price: fixed, set by the number of series, and quoted within one working day of a one-paragraph description. As a reference, a four-series portfolio runs about €9,500.

Process Calibration Audit

Any stated limit in a physical process, checked against the log of what the process actually did. A change interval, an alarm threshold, a control limit and a calibration interval are all the same claim: that reality stays inside a line someone drew. The audit finds out how often it did, by regime, and what the misses cost. Demonstrated on tool-wear records in the tool wear paper. A specimen of the deliverable, run on measured parts from a modern forge and including the day-2 data-structure check, is published in full: read the specimen. There is also a one-page summary for quality and process engineering.

Useful for: manufacturing and process industries. Machining and metalworking, chemicals and pharmaceutical batch production, food and beverage, and asset-heavy operations such as utilities and rail that run to maintenance thresholds.

Examples: tool change intervals (the case in the paper); predictive-maintenance alert thresholds; SPC control limits and process-capability claims; instrument calibration intervals; yield, scrap and cycle-time forecasts; inspection sampling plans and acceptance limits.

Input: the records where a limit, interval or threshold was stated and outcomes were logged: wear or dimensional measurements, sensor readings, maintenance and intervention history, scrap logs, process parameters. Six months or more, exported however they already exist, a spreadsheet included.

You get: the coverage picture for your own process, by regime rather than in aggregate; the threshold curve at your own limit with your own rates, whether that limit is a change interval, an alarm level or a control limit; and a data-structure check delivered on day 2, because if the records cannot support the analysis both of us should know inside 48 hours.

Timing: two weeks from data receipt.

Price: fixed, set by the size of the record set, and quoted within one working day. As a reference, one production line with two years of records runs about €6,500.

Model Health Check

One risk model and the backtest that grades it, reviewed together. The first half is the question every backtest asks: does the VaR breach as often as it states, do the breaches cluster, and what did the breach record do to the supervisory count. The second half is the one a backtest cannot ask of itself: whether the test's verdict comes from the model or from the reference the test was set up with. A specimen of exactly this deliverable, run on public prices, is published in full: read the specimen. The analysis behind it is the tail-risk backtesting paper, and there is a one-page summary for risk teams.

Useful for: a risk or model validation team that has to defend a VaR or expected-shortfall number to a committee, an auditor or a supervisor, and a consultancy that needs a second pair of hands on a validation.

Examples: a market-risk VaR and its backtest; an expected-shortfall model graded by a simulated threshold; a fund's VaR against its overshooting record; a challenger model whose improvement needs separating from a change to the test.

Input: one CSV with date, VaR, expected shortfall if you state one, and the P&L it is tested against, at each confidence level you report. Two years of daily history or more; a 99% VaR states only 2 or 3 breaches a year, so longer is better. For the second half: the specification of the test you run, and the change log of the model and of the test.

You get: a series-integrity check on day 2; breach frequency and clustering against each stated level, with intervals; the rolling 250-day count and time spent in each supervisory zone; the capital effect at your own book size where the rule applies to you; the ES verdict re-scored against alternative reference distributions and against models built to be broken; a findings register with severity derived by rule; an assumptions register; a 45-minute readout.

What it is not: a certification or a regulatory sign-off, and not a substitute for your own model risk function. It is evidence and analysis; the accountable decision stays with you.

Timing: three weeks from data delivery.

Price: fixed, set by the number of models, and quoted within one working day. As a reference, one model with its backtest runs about €6,500.

Terms, the same for all four

A fixed price and a fixed timeline, agreed before work starts. A standard mutual NDA before anything moves. The audit runs on your infrastructure if you prefer. Your data is deleted on delivery and never used elsewhere. And if your bands or your limits turn out to be well set already, the audit says so, which is a result worth having too.

First step

Send one paragraph describing the series or the records, or just the CSV, to matthew@mtanti.com. You will have a quote within one working day, and for the Vendor Band Audit a start date. Nothing is signed before the NDA. How I handle what you send is set out in the privacy notice.