BearSignal reads the entire U.S. and Hong Kong listed universe and surfaces the companies whose own numbers cannot all be true. We do it deterministically, and we show our work — every flag exposes the exact figures and the thresholds they crossed. This page explains the method and why it is defensible.
The engine reads at scale and publishes what clears certification. No watchlist, no sampling, no ranking by opinion.
Continuously read every U.S. and Hong Kong listed company — SEC EDGAR and HKEX filings, financial history, ownership records — reconstructed into a comparable financial picture per company, per fiscal period.
Run each company through independent forensic dimensions that test whether its own figures can all be true at once — accruals vs. cash, ratios vs. peers, price vs. fundamentals, disclosure clarity. A contradiction fires only on an absolute threshold, never on rank or opinion.
A gray flag is a starting point for investigation, not a conclusion. Every flag is certified against six structural checks, with evidence anchored to source filings, before publication.
A hard contradiction is rarely one number — it is a healthy-looking headline that several independent forensic lenses disagree with at once. Each dimension is computed from the company's own filings, with a fixed formula and an absolute threshold. When a dimension crosses its threshold, the exact value and direction are recorded.
How much of reported earnings is driven by accruals rather than cash. Earnings that outrun cash flow are lower-persistence and a classic contradiction signal.
A composite of accrual, margin, and leverage ratios. An elevated profile is consistent with an earnings-management-like pattern — surfaced for review, never asserted as intent.
The distance of a company's whole financial vector from its peer distribution. A large distance means the company is a statistical outlier against comparable firms.
Where market pricing diverges from the fundamental signal — the tape and the books telling different stories.
The readability of the annual-report narrative. Extreme obfuscation correlates with disclosure risk — clarity is a feature, not a footnote.
The dimensions combine through a martingale e-process with a false-positive certificate — so the bar to raise a flag is statistically controlled, not a hand-set score.
We never ask you to trust a score. Every gray flag shows the exact figure on each dimension that fired, the threshold it crossed, the direction, and the fiscal filing it came from — frozen at the moment of the flag (point-in-time). Below is the shape of what a gray-flag detail shows (illustrative).
A gray flag is what the engine raises. A red flag is what Ursa issues — and it is not an opinion, it is the output of a trained machine. Behind it sits the training record: independent practising accountants — never fewer than three — each writing their own reasoning chain on historical filings, blind to the others', cited line by line. Annotators teach the engine; they issue no opinion on any company.
Ursa reads the filings and issues the flag. No analyst opinion, no committee — a reasoning engine, and the record of how it learned.
Each historical case is worked by three annotators who cannot see one another's work, and every chain states its author's standing: "CPA · 12 yrs Big-4 audit · Annotator #0347." Annotators label the past; they issue no opinion on any company.
When annotators split, the dispute and its adjudication go into the training record. A judge trained only on sanitized agreement cannot be trusted.
Chains are timestamped and hash-anchored when written. Revisions are versioned; nothing is silently rewritten.
What a training record contains, in order: the engine trigger surface that raised the gray flag; every signed annotation chain, including any that dissented; the disagreement record; the lead reviewer's adjudication where one was needed; and the timestamp and hash of each chain as written. A flag whose record cannot show all of it does not ship.
Fixed formulas and absolute thresholds. The same filings produce the same flag — no rank, no opinion, no after-the-fact tuning.
Each flag records the values exactly as they stood when it fired — what we saw, when we saw it, auditable later.
Every contradiction is anchored to the source filing and fiscal period, built to withstand scrutiny.
The engine reads at scale and publishes what clears certification. A flag is the start of work, not the end.
The full forensic files — with the per-dimension figures behind each gray flag — are open to qualified institutions. Request access and we will be in touch.
Request Institutional Access View Investigations →