Why African data needs Africa-specific forecasting models

Most African economic forecasts are built on global models designed for developed markets. They assume stable institutions, deep financial markets, and consistent data. Africa has none of these uniformly — and the result is forecasts that look credible on paper but miss the structural realities of African economies.

At Pan Africa Data, we spent months confronting this problem directly. Every time we applied standard forecasting approaches to African data, the outputs were wrong in systematic ways. South Africa's unemployment forecast climbed to 37% by 2035 — implausible. Nigeria's inflation flatlined at 2% — impossible. Morocco's inflation was artificially lifted to match a global floor that simply doesn't apply to a central bank that targets 2%.

This piece explains what we found, why standard models fail for Africa, and what we built instead.

The six structural problems

Problem 1 — Inflation spike years distort the trend

The 2022–2023 inflation spike hit Africa hard. Ghana hit 54%. Ethiopia exceeded 35%. Even South Africa reached 7%. A standard 5-year trend model takes these spike years at face value and extrapolates a rising inflation trajectory for years afterwards.

The result: South Africa's inflation forecast climbs from 4.4% in 2024 toward 6%+ by 2029, even as the South African Reserve Bank has explicitly set a new midpoint target of 3% and inflation has already returned to 4.4% in 2024.

Problem 2 — A single global inflation floor doesn't work for Africa

Standard models apply a global inflation floor — typically 2% — to prevent forecasts from going negative. This makes sense for developed markets where deflation is the tail risk. For Africa, the problem is the opposite: applying a 2% floor artificially lifts forecasts for CFA franc zone countries (Senegal, Côte d'Ivoire, Benin, and others) that structurally operate at sub-2% inflation due to their currency peg.

Senegal's actual 2024 inflation was 0.8%. A global model floored at 2% would forecast it rising to 2.0% — not because the data suggests it, but because of a hard constraint designed for a different context.

We use country-specific central bank targets where they exist, and a lower global floor of 0.5% for countries without one.

Problem 3 — COVID distorts unemployment trends

South Africa's unemployment jumped from 29% to 34% between 2020 and 2021 due to COVID lockdowns. A trend model anchored on those years extrapolates that upward trajectory — projecting unemployment rising toward 37% or higher by 2035. This is not a forecast. It is an artefact of a one-time shock.

Structural unemployment in South Africa has been persistently high but relatively stable in the 28–33% range before and after COVID. Our model applies mean reversion toward this structural rate, producing a forecast that declines from 32.6% in 2024 to 28.0% by 2035 — still high, still realistic for South Africa's labour market, but not a runaway extrapolation of a crisis year.

We apply the same approach for Nigeria (structural rate 5.0%), Kenya (5.5%), Ghana (4.5%), and nine other countries where measured structural rates are available.

Problem 4 — GDP forecasts stop too early

The World Bank Macro Poverty Outlook, which we use as our primary institutional anchor, provides GDP forecasts to 2028. The IMF WEO extends slightly further. But professional clients doing market entry analysis or financial planning need a 10-year horizon — and there is no institutional source that covers Africa to 2035 with country-level granularity.

Our approach: use WB MPO as the anchor for 2025–2028, then extend to 2035 using GDP growth rates that converge from the MPO 2028 level toward long-run potential growth rates calibrated by country. East Africa's structural growth potential (Rwanda 7%, Tanzania 6%, Uganda 5.5%) is materially different from Southern Africa's (South Africa 1.8%, Namibia 2.5%). A global average would misrepresent both.

Problem 5 — National averages hide city-level reality

Nigeria's national Gini coefficient is 33.9. That single number is supposed to tell you something about the income distribution of 230 million people across 36 states and hundreds of cities. It doesn't. Lagos has a materially different income profile than Kano. Nairobi is not representative of Mombasa. Cairo operates in a different income tier than Alexandria.

Pan Africa Data has constructed proprietary sub-national income distribution data across two layers below the national level — city (338 metropolitan areas) and locality/suburb (2,044 localities of 50,000 people or more under the United Nations World Urbanization Prospects 2025 revision) — across 49 countries, the only dataset of this kind available for the African continent. For each geography we provide population by income class, income bounds in current USD, constant USD and PPP, and Gini coefficients, all derived from national-accounts-anchored data adjusted by sub-national income relativities.

Problem 6 — A single poverty line misrepresents income distribution across Africa's income spectrum

Most income distribution datasets apply one global poverty line to every country — typically the World Bank's $3.65/day lower-middle-income threshold. For Kenya or Nigeria, this is the right calibration anchor. For South Africa, Botswana or Namibia, it is systematically wrong. These are upper-middle-income economies where the World Bank itself uses a $6.85/day benchmark — and applying the lower line produces a distribution curve that materially understates the proportion of the population in poverty.

In South Africa's case, applying the $3.65/day calibration anchor produced a Marginalised population share of 31.2% — far below the World Bank PIP and Statistics South Africa figures of 55–64% below comparable poverty lines. The mismatch was not a data gap. It was a methodology error that only becomes visible when you validate the output against country-specific benchmarks.

This produces income distributions that are simultaneously comparable across countries (same income bands in PPP terms) and accurate within each country (calibrated to the correct World Bank poverty benchmark). South Africa's corrected Marginalised share moves to 37.5% — a substantial improvement validated against both World Bank PIP headcounts and Statistics South Africa's own poverty measurements.

Unlike a one-size-fits-all approach, this methodology correctly reflects that a $3.65/day calibration point is appropriate for Kenya but not for South Africa — and that the difference matters enormously for any market sizing, financial inclusion, or consumer segmentation analysis built on top of the data.

What this means in practice

These are not minor calibration differences. A global model applied to South African data produces inflation forecasts that diverge from SARB policy by 3 percentage points by 2030. An unemployment forecast that climbs to 37% when the structural rate is 28% gives the wrong signal for every consumer credit, retail expansion or insurance pricing model built on top of it.

The compounding effect across all five dimensions — inflation, unemployment, GDP, city income, and Gini — is the difference between data you can build on and data that looks plausible until someone stress-tests it.

Transparency as a feature

We document every methodology decision on our methodology page. Professional clients should know whether a forecast is a World Bank institutional projection, a PAD trend model anchored to a central bank target, or a proprietary derivation from city income relativities. Every record in our database carries a value type and source label — M for measured, TE for trend extrapolation, MODEL for proprietary derivations, CF for carried forward. No hidden assumptions.

This transparency is intentional. The value of data for serious analysis — investment due diligence, market sizing, financial inclusion research — depends entirely on whether the person using it understands what they are working with.