Focus Areas

The desk publishes free editorial articles on how machine-learning models surface anomalies in financial statements. The pages below describe the recurring focus areas, the detection methods each one relies on, and the limits a reader should keep in mind. Nothing on this page is a paid service, a subscription tier, or investment advice.

Revenue recognition: cut-off and channel-stuffing flags

The first focus area reads the timing of recognised revenue against its own history. When a company books a disproportionate share of a quarter's sales in the final weeks, the cadence of daily or weekly revenue stops matching its earlier shape. An isolation forest trained on standardised daily revenue lines ranks the periods whose end-of-quarter mass sits furthest from the typical distribution; a short sequence model then checks whether those spikes reverse in the opening weeks of the next quarter. The article series explains which filings show the pattern, what the model can and cannot infer from it, and why a human auditor still has to walk the cut-off evidence before any conclusion is drawn.

Related-party transactions: counterparty graph outliers

The second focus area treats the related-party note as a small graph. Each disclosed counterparty is a node; the disclosed transaction types and balances are edges. A reader can usually follow one relationship at a time, but a clustering pass over the whole note surfaces entities that sit unusually close to management — for example, a supplier whose address matches a director's registered office, or a customer whose balances move in lockstep with a guarantor. The articles walk through how a simple graph is built from the note, which centrality measures are meaningful on a graph this small, and where the method is simply amplifying what the footnote already says, more loudly.

Reserves and accruals: reserve-release sequencing

The third focus area follows the direction of provision movements across reporting periods. A warranty, litigation, or obsolescence reserve that falls in a year when the underlying risk has not visibly improved is a classic accrual-quality signal; anomaly detectors weight it heavily when reported earnings land exactly on guidance. The desk's articles trace the line items through consecutive annual reports, show the model's ranked output, and separate the cases where a release is defensible from the cases where it is merely convenient. The framing is always descriptive — what the line did — never a recommendation to act on it.

What this page is not

It is not a catalogue of paid reports, a tier of premium analysis, a managed-account offering, or a tool you can license. There is no checkout, no member area, and no booking calendar. If a focus area is relevant to a question you are working on, the natural next step is to write to the desk and ask. An inquiry is answered by email; it does not open a sales process, because there is no sales process to open.