Data & methodology
Every score shows its workings.
An unexplained number is not usable in an institutional process, however accurate it turns out to be. This is how ours are built, and what we claim for them.
Principles
Six commitments we hold ourselves to.
Say what the score is
Kilwa runs two kinds of score and labels which is which. The scores inside published reports — the Takeoff Gap Score, the Repatriation Risk Score and their peers — are weighted composites of house assessments, structured risk rankings rather than validated predictive models. ISI and METI are modelled: a hierarchical gradient-boosted ensemble and a hybrid time-series forecast. Using predictive methods is not the same as publishing a validated track record, and we do not claim one we have not published.
Flag every input
Each input carries a provenance flag — verified or estimate. Estimates are never presented as measurements, and the full input table with flags is published in the appendix of each flagship report.
Publish the robustness work
Re-weighting tests and 1,000-draw Monte Carlo simulations are run against every flagship score and published, so a reader can see how much of the rank order survives a different set of assumptions.
Decompose every result
SHAP-based attribution returns the specific indicators that drove a score and the direction of each contribution. A number no one can interrogate cannot carry an allocation decision.
Document what would make us wrong
Model cards record scope, training data, known weaknesses and invalidation conditions. Each flagship report publishes its own signposts, and we grade ourselves against them.
Compute in the source language
Sentiment is extracted from French, Arabic and Swahili sources directly rather than translated into English first, because translation pipelines are where policy nuance is most reliably lost.
Inputs
90 indicators, four families.
ISI synthesises 90 macroeconomic, governance, regulatory and market-structure indicators into a sector-aware score out of 100, refreshed as the inputs move rather than on a publication calendar. Every metric, source and language is registered in Kilwa's Golden Source Data Dictionary, which is versioned and dated.
- Macroeconomic
- GDP and growth series, inflation, FDI flows, current account, reserves and import cover, debt and debt service.
- Governance & institutions
- Regulatory quality, rule of law, policy predictability, contract enforcement and institutional stability indicators.
- Market structure
- Financial market depth, currency convertibility and cover availability, exit routes and listed-market turnover.
- Policy & news sentiment
- Regional media and official policy sources in English, French, Arabic and Swahili, processed natively.
Score bands
How to read a Kilwa risk tier.
Our published risk scores run on a common tiering so a reader can move between reports without relearning the scale. Higher means more risk, less cover, or fewer working routes — depending on what the score measures.
| Tier | Band | Interpretation |
|---|---|---|
| Critical | 70 and above | No working route, or no meaningful cover, at any price. Size the exposure accordingly. |
| High | 55 – 69 | One conditional route, usually dependent on a single counterparty or a development balance sheet. |
| Elevated | 40 – 54 | Real but expensive and episodic. Ladder the exposure; do not assume continuity. |
| Moderate | Below 40 | Working routes through market depth, pegs, or both. The exception, not the rule. |
Tier boundaries are fixed across reports so scores remain comparable. The underlying score name differs by report — for example the Hedging Gap Score in The Missing Market and the Exit Liquidity Score in The Long Hold.
Coverage
What we mean by a market.
Three different numbers describe Kilwa's reach, and they are not interchangeable. We publish all three rather than pick the flattering one, because a reader who checks should find the basis rather than a discrepancy.
- 54Research
- Sovereign African states, as recognised by the United Nations. This is the frame our flagship research is scored on, and the number our published reports use.
- 55Registry
- Country and territory profiles in the ingestion registry — African Union membership, which adds Western Sahara to the 54. Each profile carries a minimum of six source slots: the monetary authority, the statistics office, the executive, the legislature and regional media.
- SubsetLive scoring
- Live ISI and METI scoring covers a subset of the registry and expands with each release. A market is only scored once its sources clear the data-quality thresholds, so the live count is smaller than the registry by design rather than by omission. The current figure is stated in-product.
Transparency in numbers
What we publish alongside every score.
Methodology, the complete input table with provenance flags, the re-weighting tests and the Monte Carlo results all appear in Appendix A of each flagship report.
Browse the research90
indicators per country score
1,000
draw Monte Carlo robustness tests per flagship score
100%
of inputs flagged verified or estimate
85
languages in the registry
- What is the Investment Suitability Index (ISI)?
- The Investment Suitability Index is Kilwa's country and sector scoring model for frontier markets. It synthesises 90 macroeconomic, governance, regulatory and market-structure indicators into a single score out of 100 for every market Kilwa covers, and it is refreshed as the inputs move rather than on a publication calendar.
- What is the Market Entry Timing Index (METI)?
- The Market Entry Timing Index is Kilwa's timing model. It fuses time-series analysis with multilingual news and policy sentiment to forecast entry windows of one to six months, returning an ENTER, HOLD or WATCH posture with a confidence indicator for each covered market.
- Are Kilwa scores predictive models?
- Some are and some are not, and the distinction matters. ISI and METI are models: ISI is a hierarchical gradient-boosted ensemble that clusters comparable markets so a thinly-covered economy can borrow statistical power from its peers, and METI is a hybrid forecast that injects the Sentiment Pulse delta into a time-series model as an exogenous variable. The scores inside published reports are different — weighted composites of house assessments against published weights, with every input flagged verified or estimate. Neither carries a published out-of-sample track record, so we call neither a validated predictive model. Using predictive methods and having demonstrated predictive accuracy are separate claims, and we make only the first.
- How many indicators go into a Kilwa country score?
- Ninety, across four families: macroeconomic, governance and institutions, market structure, and policy and news sentiment. Every input carries a provenance flag marking it as verified or estimate, and the full input table is published in the appendix of each flagship report.
- Which languages does Kilwa process?
- Kilwa's language registry covers 85 languages and variants across the continent, tiered by how each is handled. Tier A languages — including English, French, Arabic, Portuguese and Swahili — run on native models. Lower tiers use dialect adapters or translation with a reviewed financial lexicon. Sources are processed in their original language rather than translated into English first, because translation pipelines are where policy nuance is most reliably lost.
- How many sources does Kilwa ingest?
- The Sentiment Pulse source registry holds 416 records across all 55 African Union country and territory profiles, with a minimum of six source slots per geography covering the monetary authority, the national statistics office, the executive, the legislature and regional media. A separate structured source master holds 99 records.
- How does Kilwa test the robustness of a score?
- Every flagship score is subjected to re-weighting tests and a 1,000-draw Monte Carlo simulation, and the results are published in Appendix A of the report so a reader can see how much of the rank order survives a different set of assumptions.
- Can a Kilwa score be explained?
- Yes. SHAP-based attribution returns the specific indicators that drove a score and the direction and size of each contribution, and model cards document scope, training data, known weaknesses and the conditions that would invalidate the result.
Disclosure
Kilwa research is analysis, not investment, legal or tax advice. Scores published by Kilwa are structured risk rankings, not validated predictive models, except where explicitly stated otherwise. Past performance of any signal is not a guarantee of future results. Questions about methodology are welcome at research@kilwa.io.
Get started
Put the methodology in front of your risk team.
We're happy to walk your due-diligence process through the models, the inputs and the limits.