Report card

Can I trust these numbers?

Most risk numbers are taken on faith. These come with a report card.

Every build of the live demo book is scored on five checks before it goes live: do the stocks move together the right amount, do they fall together, are moves the right size, did the range hold, and how often was the 1-in-20 loss line crossed. Every daily projection is scored again once its ten trading days have passed.

Did each build pass its five checks?

Before a build of the demo book goes live it is scored on past ten-day windows it never trained on. A failed check is shown as failed, and the build is published either way.

Loading the builds...

How much did we correct the range?

If reality has been running wider than the range, the range widens; if narrower, it narrows. The factor is set from earlier windows only, then applied to every number the book shows.

No correction yet: the numbers are uncorrected until the first calibration.

One point per daily projection of the demo book.

Did past ranges hold?

Every daily projection is scored once its ten trading days have passed. Daily origins overlap, so the breaches are not independent: read the counts as they accumulate, not a p-value.

Loading...

How did the range hold on the research books?

Research books, one year: five books of 8 to 50 US stocks, every two-week window of 2026, scored out of sample. The share of real fortnights that landed inside the 90% range, before and after the correction (set from earlier fortnights only), and the checks passed under the current engine.
Research books, one year
Research bookInside the 90% range, beforeAfter the correctionCorrection factorChecks passed
8 stocks91.0% (research books, one year: 18 two-week windows of 2026)89.4% (research books, one year: 18 two-week windows of 2026)x0.985 of 5
16 stocks83.0% (research books, one year: 18 two-week windows of 2026)91.8% (research books, one year: 18 two-week windows of 2026)x1.285 of 5
24 stocks81.5% (research books, one year: 18 two-week windows of 2026)92.0% (research books, one year: 18 two-week windows of 2026)x1.325 of 5
43 stocks82.6% (research books, one year: 18 two-week windows of 2026)92.7% (research books, one year: 18 two-week windows of 2026)x1.284 of 5
50 stocks81.9% (research books, one year: 18 two-week windows of 2026)93.1% (research books, one year: 18 two-week windows of 2026)x1.324 of 5

One year is 18 fortnights per book: about ±15 points of uncertainty on each coverage figure, narrowing to about ±7 with three years of history.

Audit your current numbers

Upload your vendor’s (or your own) daily loss line, 1 in 20 or 1 in 100, next to what the book actually did. You get the breach count, the two standard tests and the regulator’s traffic light. Nothing you upload is stored: the file is read, scored and discarded in the same request.
VaR 95% or VaR 99%; the column keeps the name var95 either way.

A CSV with the header date,var95,realised, one row per day, up to 200 KB. Fractions or percents (values above 1 are read as percents); losses negative, or a column of positive loss sizes.

What it cannot do

  • Forecast direction. The middle of any range called up or down about as often as a coin flip. We don't sell it, and the middle of the range is never a price target.
  • Use simulated paths to pick weights. Feeding simulated tails or volatility into an optimiser never beat plain risk parity out of sample. So the simulator measures risk and culprits; the weights come from measured co-movement and volatility.
  • Match the rarest half-hours. The most extreme half-hours appear less often in our paths than in real markets. In a liquidity panic losses can run past the bad-fortnight figure; the jump badges flag the stocks where this matters most.
  • Promise a one-day range. The first half-hour after the open is still under-ranged, so one-day bands are guides, not 90% promises. The ten-day range is the calibrated one.
  • Correct before reality moves. The range widens after the first stretched fortnight, not before it.
  • Escape a short history. The research is one year: 18 fortnights (about ±15 points on coverage) and 36 weekly rebalances (about ±0.6 Sharpe). A multi-year walk-forward is the next milestone.

The same report card for your book

Every book carries its own five checks, correction and scored fortnights.