How every number is made.

A deterministic rule engine reads SEC filings, exchange disclosures (NGX, NSE Kenya, JSE SENS) and the financial press in English, French, Portuguese and Arabic, and fires on patterns for eight signal types: M&A, credit risk, distress, earnings surprise, leadership change, regulatory action, debt restructuring and insider activity. A source must name the company, and the triggering phrase must be in a clause about it. The language model never creates, removes, retypes or re-grades a signal; it only explains confirmed ones.

Severity Design choice: the rule score, raised by independent corroborating sources, falls into a band for its signal type (for M&A: 0.85 and above is CRITICAL, 0.70 HIGH, 0.50 MEDIUM). CRITICAL needs at least two independent sources.

Confidence Measured: the share of past flags of the same type and severity that were followed within 90 days by a confirming filing, exchange disclosure or credible report, from a walk-forward backtest. For earnings and insider signals, where the disclosure is the event, it is the share whose market-adjusted move went the expected way. Small samples are smoothed toward the type-level rate:

The tables are not frozen at the backtest. Every hosted brief is labelled against what followed (price reaction after 12 days, confirmation after 95), and each hour the service pools those live outcomes with the backtest observations and recomputes every rate and move range. Once a signal type has 10 live outcomes in a market (US or Africa), the observed live rate for that market is used directly.

A 0 to 100 ranking of how actionable a signal is. The weights Design choice are stated, not fitted:

Severity weights are 1.00, 0.78, 0.52 and 0.28 from CRITICAL to LOW. Source credibility ranges from 1.0 for SEC and exchange filings to lower values for less established outlets. Corroboration is 1 − 1/(1 + 0.8n) for n sources. Recency halves every 14 days with a floor of 0.30. Liquidity comes from live market capitalisation in US dollars.

Whether the weights are useful is tested Measured: the backtest splits predictions into thirds by alpha and reports each third's confirmation rate, with the rank correlation. The evaluation page shows the result. The evaluation page states what the thirds show in the latest run. Even a rising confirmation rate would not show that high-alpha flags earn returns, which the evaluation page measures separately and finds indistinguishable from zero at the current sample size. Alpha is not a recommendation to buy or sell.

The move shown with a signal Measured is the distribution of market-adjusted returns over 10 calendar days after the same kind of flag in the backtest (close on the first trading day on or after day 10, against the first close on or after the flag; about 7 trading days) (benchmarks: S&P 500, Satrix 40, NGX ASI, NSE NASI). It is shown as the interquartile range with the median and the number of observations. With fewer than 10 observations the product says "insufficient history" rather than show a range. The direction is the source's own when it states one (an earnings beat, an insider purchase), else the measured majority.

A rubric of seven graded components Design choice: source reliability, corroboration, filing strength, historical precision, entity matches, recency and extraction, weighted 22, 20, 18, 15, 10, 10 and 5 percent. It describes how well a signal is supported, not the probability it materialises. Its historical precision component is measured.

Transactions are extracted by rules from headlines in African business and private-capital publications and from searches in every one of the 54 countries' news editions and languages, every day. Each record has its type, parties, amount, stake, round, stage, sector and country, with every source linked. Amounts are as reported and converted to US dollars at the deal year's average official rate Measured. Implied valuations are amount ÷ stake. Reports of one transaction merge by company and month.

Each record shows its own verification Measured. Every source keeps its headline, sources are grouped into independent stories (syndicated copies of one wire story and near-identical headlines count once), and each field is marked corroborated (stated by two or more independent stories), single source, derived (currency conversion, stake-implied valuation), classified (sector, by keyword rules), uncertain (sources disagree, a malformed party name, or a present-tense acquisition headline that does not confirm completion) or not disclosed. The first-reported date, the date a source first reported the deal as announced, and the date one first reported it completed are shown separately; none is a legal signing or closing date.

Accuracy is measured on labelled samples of raw headlines, with a held-out set drawn after all fixes and never used for tuning: 83% (95% CI 75–88%, n=120) of extracted records are genuine transactions, 92% (95% CI 85–96%, n=108) of deal-like headlines are caught, and 71% (95% CI 61–79%, n=99) of records have every field right. Field-by-field results, by country, source, language and deal type, are on the evaluation page.

Comparables are scored Design choice: 3 points for the same sector, 2 for the same type, 2 for the same country (1 for the same region), 2 for a size within 3× (1 within 10×); deals scoring 3 or more are shown with size and valuation quartiles.

NGX and NSE Kenya statements are parsed from the audited annual statements the companies file with the exchanges. Each figure is checked against accounting identities (gross profit = revenue − cost of sales; profit after tax = profit before tax − tax; assets = liabilities + equity) and marked verified, inconsistent or unchecked. Other listings use Yahoo Finance; for SEC filers they are cross-checked against the SEC's XBRL data.

Country risk Measured uses World Bank Worldwide Governance and Development Indicators; each component maps to 0 to 100 with a fixed formula Design choice, for example inflation ÷ 40% and (2.5 − political stability estimate) ÷ 5. The score is the weighted mean over the components with data, and the share available is shown. Currency risk combines last-year and five-year depreciation against the dollar and their volatility, reduced for pegged regimes. Sector risk is the smoothed share of adverse corporate events (credit, distress, restructuring, regulatory) against deal flow and other events over 90 days.

The question is parsed for countries, sectors, deal and investor types, companies and dates. Evidence is retrieved and numbered, and the language model answers only from it, citing each claim. Citations are checked afterwards: references to items that do not exist are removed, and sentences without a citation are listed with the answer.

A walk-forward replay: on each monthly date the engine sees only what was published in the previous 30 days, and each prediction is judged on the following 90 days. Lift is precision divided by the base rate, the chance that a random company-month had the event anyway. Hosted briefs are also labelled daily against later filings, credible news and abnormal price moves.

Two outcomes are kept apart throughout. Event confirmation asks whether the flagged event happened within 90 calendar days. Market reaction asks whether the market-adjusted share price moved the predicted way over 10 calendar days. A confirmed event can come with no price move, and neither is a trading result: no costs, sizing or portfolio construction are applied.

Time windows, reconciled: the backtest judges confirmation over 90 calendar days and price over 10 calendar days. Hosted briefs use the same windows but are labelled later, at 12 days for price (so the day-10 close exists even across a weekend or holiday) and at 95 days for confirmation (the 90-day window plus 5 days for filings and reports inside it to be indexed). Intervals are 95% Wilson intervals; the full labelling rules and every prediction are downloadable from the evaluation page.

Attribution, whether a news item's event belongs to the target company, is tested on real headlines only, each with its publisher, date and link: a random sample naming tracked companies, and a hard-case sample of denials, references back to the company, subsidiaries, several companies, rumours, history and short names. On the random sample the engine found 30% (95% CI 20–43%, n=60) of real events with precision 86% (95% CI 65–95%, n=21). Corroboration counts independent stories, not articles.

Every country gets news search in its own languages, deal and event extraction, regulator detection and country risk. Machine-readable filings and prices exist for NGX and NSE Kenya (tier A); vendor data for JSE, EGX and Casablanca (tier B); the other countries are covered through news, deals and risk data (tier C). The coverage page states this for each country.

• Measured rates rest on a backtest of 1,552 companies over 18 monthly dates, re-run monthly, plus 0 live outcomes so far; several signal types have small samples, shown with every figure.
• Every confirmable type beats its base rate in the backtest, but rare events (credit, distress, regulatory) still confirm under one time in ten and leadership changes beat chance only modestly. Market reactions are indistinguishable from zero. The evaluation page shows each figure with its interval.
• Deals come from public reporting: undisclosed private deals cannot appear. On the held-out set 71% (95% CI 61–79%, n=99) of records have every field right, so individual fields marked single-source or uncertain should be checked against the linked sources.
• Statement layouts vary, so a figure is only marked verified when an accounting identity confirms it; interim statements are linked, not parsed.
• World Bank data are annual and often a year old; sovereign ratings and bond spreads are licensed and not included.
• Detection is pattern-based: auditable and reproducible, but phrasings the rules do not cover are missed, and on news headlines most are. Filings and exchange notices are covered by stronger, item-code rules.