09Honest limits
What it does not do yet.
Written down up front, not discovered in production.
01
Deals are what gets reported
Extracted from public reporting: undisclosed deals cannot appear. On a held-out test set, 83% (95% CI 75–88%, n=120) of extracted records are genuine transactions and 71% (95% CI 61–79%, n=99) have every field right. Each deal page shows which of its fields are corroborated, single-source or uncertain.
02
Coverage is uneven by country
Machine-readable filings and prices exist for NGX and NSE Kenya, vendor data for JSE, EGX and Casablanca. The other 49 countries are covered through news, deals and risk data.
03
Statements are parsed, then checked
PDF layouts vary, so each figure is checked against accounting identities and marked verified, inconsistent or unchecked. Interim statements are linked, not parsed.
04
Some signals beat chance, some do not
Across about 1,300 companies every confirmable type beats its base rate, but rare events (credit, distress, regulatory) still confirm under one time in ten, and common ones (leadership changes) beat chance only modestly. Confirming an event is not a profitable price move: measured market reactions are indistinguishable from zero.
05
Country risk uses public data
World Bank indicators are annual and often a year old; sovereign ratings and bond spreads are licensed and not included.
06
Patterns, not a trained model
Reproducible and auditable in English, French, Portuguese and Arabic, but phrasings the rules do not cover are missed: on a random sample of real news headlines about tracked companies the engine found 30% (95% CI 20–43%, n=60) of real events, with false alarms on 1% (95% CI 0–3%, n=342) of no-event headlines.
10Questions
Asked before you have to.
Does it need an LLM?
No. Detection, calibration, scoring, attribution and compound events are deterministic. Without a key the brief is built from those findings, labelled as such and flagged for human review. With a language model configured you also get analyst-language headlines, reasoning, an executive summary and an attribution check.
Can the LLM change a signal?
No. It only sees candidates the rule engine has already confirmed. It cannot add, remove, retype or re-grade them, and it cannot rate the brief below its most severe signal. If it omits a signal, the deterministic version is used. If it doubts attribution, that is shown and review is required.
Which model does it use?
Hosted language models from two providers, with no lock-in to either: a primary model, then a fallback; if one fails, the next answers. Each brief records whether the primary or the fallback model wrote its narrative.
Is this investment advice?
No. It is research tooling, not investment advice. Signals are grounded in linked source material, with evidence and attribution checks; corroboration and source-quality indicators help you assess support. When attribution is doubtful or a signal needs further review, the brief flags it. AI-written summaries are interpretations of that evidence, so inspect the linked sources before acting.
How do I know it is not hallucinating sources?
Every signal links to the filing or article it came from, with its publication date and publisher. Sources outside your lookback window are rejected, and the pipeline trace counts what was skipped and why.
Is the hosted app free?
Yes, and there is no account to create. Click Run your first brief and go: up to 1,000 analyses a day per visitor, subject to service capacity. Each brief gets its own link you can share, and your list of briefs is kept in your browser.
How do I know the numbers are right?
Every figure names its source, and the methodology page sets out how each score is computed. The evaluation page publishes how often past signals were right, weak results included, and the coverage page states how far each country's data can be verified.