Skip to content

Methodology

Models propose. A person publishes.

The machinery is an ensemble tuned to a client’s own history, plus the British calendar, plus a written challenge before anything leaves the studio. This page says what that means in practice.

Scenario bands drawn as a fan of curves on warm paper

Show the range

A point forecast is a comfort. We publish a base, a stress and an upside, and a band scaled from recent walk-forward error. If the band is embarrassingly wide, that is the finding.

Name the driver

Each release says which family of model carried the path, which feature moved it, and where an analyst overrode the machine. “The AI said so” is not a sentence we allow in a pack.

Keep the history honest

Skill is measured by walking forward through years the model had not yet seen, retraining only on what would have been known. A pretty fit on the full sample is treated as a warning.

The ensemble, in plain words

We do not bet a client’s year on a single architecture. Three families have to agree, or the disagreement itself is reported.

Gradient boosting on the drivers you can name

Price, mix, known contracts, promotions, headcount, and the lags that show up in your ledger. Useful when the business is a set of understandable causes rather than a pure wave.

A state-space model for level, trend and season

The slow structure: what is cyclical, what stepped up and stayed there, what is only a noisy month. This stops a one-off contract from being treated as a new run-rate.

A sequence model on timing

Where an order book, a pipeline or a till tape has a rhythm the other two miss: the weeks a framework order usually calls off, the delay between a proposal and an invoice. It is used as a witness, not as an oracle.

The British calendar is a feature

A model that treats 25 December like 25 March will be confidently wrong. We encode, and we say which version we encoded:

  • Bank holidays for England and Wales, with a flag if your sites in Scotland or Northern Ireland need their own set.
  • A fiscal year that starts in April, because a great many UK boards still think that way even when the statutory year does not.
  • VAT quarters and the payment dates you actually use, including the cash effect of the annual accounting scheme where it applies to you.
  • School holidays, where the client sells something families buy. We do not add weather unless the product genuinely moves with it. A wet Tuesday is not a strategy.

The human release

Before a pack is sent, an Acrisiaorix analyst reads it against three questions.

  1. Does the bridge from last month make sense in the client’s own words?
  2. Is there a single assumption which, if wrong, flips a decision? If yes, it is on the first page, not in an appendix.
  3. Did anyone override the model? The override is dated, named and kept. Silent edits are a fault.

If those questions cannot be answered, the release slips. A late honest pack is better than an early decorative one.

Limits we will say out loud

Short histories cannot support a clever model. A business that changed shape last spring — a disposal, a new channel, a lost customer who was a third of turnover — will be forecast with a heavier human weight and a wider band. We would rather mark the break than pretend the old season still governs.

Macro paths for Bank Rate or sterling are scenarios you choose with us, drawn as alternatives. They are not Acrisiaorix predictions of the Bank of England, and they are not to be read as a house view for traders.

Nothing in the method removes your duty to keep proper accounting records, to take advice where the law requires it, or to tell your board the truth when the range is ugly.

Talk through a first release