Methodology

How Asotele ingests, models, and explains Nigerian economic conditions. This page is updated as the system evolves.

Data sources (17)

  • Energy: Brent crude (daily)
  • Currency: CBN official rate, parallel market (Aboki), CoinGecko USDT/NGN crypto-derived rate
  • Central bank: CBN MPR, foreign reserves, CPI
  • Equities: NGX All-Share Index + 9 sector indices + market breadth (34,349 historical points)
  • Real economy: WFP food prices, fuel (PMS + diesel), cement, employment (Jobberman), telecom subscribers (NCC)
  • Sentiment: Financial news via FinBERT, CentralBankRoBERTa, InflaBERT; Google Trends 33-query consumer panel
  • Macro: 12 World Bank indicators; USDT cross-border flows via Tronscan API

Forecast engines

ARIMA

Univariate baselines for each series. Auto-activates per series at 20+ observations. Used as a sanity-check benchmark for more sophisticated models.

Markov regime-switching + GARCH(1,1)

Three-regime classification (boom / stress / crisis) applied to oil prices, with GARCH(1,1) volatility modeling for the FX series. EGARCH auto-upgrade at 200+ observations to better capture asymmetric shocks.

VAR + Granger causality

Multi-variable vector autoregression for FX, oil, NGX, and inflation interactions. Activates at 30+ observations per series. Granger causality testing identifies which signals predict which.

Foundation models

Asotele uses three model tiers:

  • Tier 1 (live): Ollama-hosted Qwen 3 14B, Mistral Small 24B, Devstral 24B for daily briefing generation
  • Tier 2 (live): Chronos-2 and BISTRO time-series foundation models for forecasting
  • Tier 3 (in progress): Qwen 3.6 35B-A3B and Gemma 4 26B-A4B, fine-tuned on Nigerian economic corpus + CBN/NBS document archive

Evaluation

All forecasts are evaluated against realized outcomes monthly. Accuracy reports published openly. Benchmark suite covers: forecast error vs realized values, factual recall on Nigerian economic facts, sentiment classification accuracy on financial news, and counterfactual reasoning.

Code, weights, datasets

All released under permissive licenses. See the GitHub repo for code (Apache 2.0), the HuggingFace org for model weights and datasets.