Cut-off and channel-stuffing flags
How isolation-forest and sequence models catch quarter-end spikes that do not match the surrounding cadence of sales — the kind of timing drift that precedes many restatements.
Editorial desk · Wanhua, Taipei
Ion Mesh Investment Review
Independent editorial articles on how machine-learning models surface anomalies in financial statements — written from a small Taipei desk for readers who want to understand the method, not buy a product.
Taiwan lists over 2,000 companies on its exchanges and files every annual report in both Chinese and English. That bilingual, rule-based disclosure culture makes the island a useful observation deck for the patterns machine-learning models flag in financial statements — from revenue cut-off timing to sudden reserve releases. Our articles read those filings the way an anomaly detector would: line by line, looking for the line that does not fit its neighbours.
What the articles cover
Three lenses on statement anomalies
Each focus area pairs a documented failure pattern with the detection method that exposes it. Nothing here is investment advice.
Counterparty graph outliers
Reading the related-party note as a graph: which entities sit unusually close to management, and which clustering measures surface circular flows buried in footnotes.
Reserve-release sequencing
Tracking the direction of provision movements across periods — the accrual-quality signal that anomaly detectors weight most heavily when earnings land exactly on guidance.
How an article is built
The editorial process behind each piece
01
Pull the primary filing
Every article starts from the issuer's own annual report on the Market Observation Post System, not from a derivative data feed. We quote page numbers so a reader can open the same PDF and disagree with us.
02
Run the anomaly pass
A documented, reproducible model run — usually an isolation forest over standardised line items — produces a ranked list of unusual observations. We publish the parameters and the limits of the method, not just the headline.
03
Write for the reader, not the trade
The piece is written in plain English for an investor who reads footnotes. We state what the model noticed, what it cannot prove, and where a human auditor would still need to look. No buy or sell framing is attached.
This desk publishes articles and answers questions. It does not sell software, run a paid newsletter, place trades, or give personalised investment advice. If a piece has made a financial statement clearer to you, the only next step is an inquiry — and an inquiry is just a question.