APPLIED AI / REGIONAL FORECAST LAB

Explore a future.
Inspect the model behind it.

An interactive application of my AI and data practice: train a revenue model, test it against held-out years, and explore population, youth, economic output and resource scenarios.

Evidence edition: 25 October 2026 · runs locally in your browser · no API key required

Learn from observations

Ridge regression learns nominal revenue patterns from 259 annual observations across 36 states and the FCT.

Keep assumptions visible

Population, youth and GDP use transparent scenarios. They are not census results or independently validated ML forecasts.

Test before trusting

Chronological backtests compare revenue predictions with actuals and a last-year baseline. Forecast ranges are illustrative, not calibrated confidence intervals.

Loading the local evidence files…

Build your scenario.

StateExploreHorizonChart currency

Revenue modelRegularisation strengthAdditional annual revenue change (%)

This adjustment is your scenario, not an effect learned from a policy intervention.

Population assumptionsAnnual population growth (%)Additional net migration per 1,000 residents each year

The starting growth rate is calibrated between the 2006 census and 2023 projection. Additional migration changes that scenario; it is not a measured state flow.

Youth assumptionsYouth share in the 2023 scenario baseline (%)Annual share change (percentage points)

The default uses the selected state’s 2022 NPC projected share aged 15–24. A fixed or adjusted share is not a full cohort-component projection.

Economic output assumptionsAnnual nominal GDP growth (%)

The 5% starting value is an editable assumption, not an estimated growth rate. The base is NGF’s source-specific 2023 GDP series.

Segilola · Osun onlyAnnual production change after 2026 (%)Illustrative gold price (£ per troy ounce)

£2,500 is a user-editable illustration, not a market quote or price forecast. Output × price is gross illustrative sales, before costs, taxes and recovery adjustments.

Currency assumptions

NGN per US dollarGBP per US dollarAnnual NGN depreciation from 2025 (%)

Defaults are 2025 annual-average World Bank / IMF reference rates. Forecast FX changes are your assumptions. Historical rows use the same reference rates for comparability, not historical conversions.

LEARNED REVENUE MODEL

Revenue outlook

Source historyModel / scenarioIllustrative range

Exact values and evidence type

VALIDATION / WALK-FORWARD TESTING

How well did the revenue model predict?

For each test year, training stops at the previous year. The model predicts 2023, 2024 and 2025 without seeing that year’s revenue. The last-year baseline predicts no change. Lower error is better.

Inspect the fitted model and its features

The target is the logarithm of annual IGR. Inputs are the logarithm of previous-year IGR, calendar year and jurisdiction indicators. Ridge regularisation shrinks coefficients; scaling uses training data only. The intercept is unpenalised. Exponentiation returns a central log-scale estimate, not a bias-corrected arithmetic mean.

Coefficients are statistical associations, not causal effects. Changing a slider does not demonstrate that a policy will produce the selected outcome.

MODEL CARD / EVIDENCE LIMITS

One lab, several kinds of evidence.

Revenue: fitted and backtested

2019–2025 NBS IGR data supplies 259 annual observations and 222 consecutive-year training pairs. The model pools 37 jurisdictions. There are only seven years per jurisdiction and three holdouts each. Inflation, exceptional receipts and changes in tax collection can dominate nominal growth. Kwara’s 2023 figures include exceptional pilgrims-fee collections.

Trying several regularisation settings makes the displayed validation exploratory. It is not an untouched final test. The forward error band expands historical log-error with the square root of the horizon; it is not a probabilistic guarantee.

Population and youth: explicit scenarios

The 2006 census is observed history. The 2023 total and 2022 age structure are NPC projections from different vintages. The lab does not treat those projections as independent observations to train and “validate” a population ML model. It cannot predict an official census result.

Population ranges vary annual growth by ±0.5 percentage points. Youth ranges also vary the youth share by ±2 points. Migration is an additional scenario adjustment. Age structure, mortality and fertility would need fuller modelling for an operational demographic forecast.

GDP: a source-specific base

NGF’s 2023 state GDP profiles do not specify price basis or estimation method; Lagos’s own series differs. The lab applies an assumed nominal growth rate, with a ±2-point growth range. GDP per resident combines this output scenario with the population scenario; it is an indicative ratio, not a welfare or “economic power” score.

Resources: production is not geology

The resource case study is Segilola mine in Osun, not all Osun production or a Yorùbá ownership claim. It shows reported gold poured in 2022–2025 and the operator’s revised 2026 guidance of 70,000–72,000 ounces. Subsequent years are user-defined output scenarios with a ±20% illustration, not operator guidance or a mine plan.

Mine life, grades, reserves, recovery, permissions and costs can invalidate extrapolation. No credible state-wide annual resource-production panel was verified for the other states. Mineral occurrence and installed capacity are not substituted for production.

All computations run in your browser. Your slider settings are not sent to a prediction service. Downloaded CSV files record assumptions, model version, evidence types and currency rates. This is an educational applied analytics project, not an official planning forecast.

Open the evidence.

Explore the state, Benin and Togo evidence dashboard → · Explore the natural-resource catalogue →