Responsible AI

Asotele is a forecasting and natural-language system built for high-stakes economic decisions. The failure modes are not hypothetical — a hallucinated number on a treasury desk or a confidently wrong inflation forecast on a market trader's screen can cause real harm. This page documents how Asotele handles those risks, what we measure, and where the limits are.

Risks specific to this system

  • Hallucination of facts. LLM-generated briefings can fabricate numbers, dates, or events. Highest stakes in monetary policy summaries and FX-rate descriptions.
  • False confidence. Point forecasts without uncertainty intervals invite over-reliance. Treasury and SME users may not natively interpret model uncertainty.
  • Bias toward English and the formal sector. Training data is heavier in English and in formal-sector financial reporting. Informal-economy signal coverage is improving but uneven, and non-English interfaces depend on smaller multilingual corpora.
  • Misuse as investment advice. Forecast outputs are decision-support inputs, not personalised investment advice. The line matters legally and ethically.
  • Stale or wrong source data. When upstream sources lag (NBS releases, CBN bulletins), the model can confidently reason from outdated inputs.
  • Concentration risk on a single forecast view. If a meaningful share of Nigerian SMEs make the same FX call because Asotele said so, that itself becomes a market-moving signal.

Mitigations in production

  • Source-grounded generation. Every quantitative claim is tied back to a specific dated source record in the data pipeline. The intelligence layer is instructed to cite, not invent.
  • Public evaluation harness. The asotele-eval-nigerian-economy dataset scores model outputs on factual recall, required citations, omission rate, and counterfactual reasoning. Scores are tracked monthly and published.
  • Confidence intervals on every forecast. Point forecasts are accompanied by GARCH-derived or model-implied uncertainty bounds. The UI surfaces uncertainty rather than hiding it.
  • Human review for public briefings. Daily public-facing briefings are reviewed before publication. The institutional API is not gated by human review and is documented as such — institutions know what they are integrating.
  • Use disclaimers. Every public output ships with a plain-language disclaimer that it is decision-support, not investment advice, and is not a guarantee.
  • Multiple redundant sources per signal. FX, oil, and equities each have multiple ingestion paths so a single bad source doesn't poison the output.
  • Regime-state communication. When the FX or oil regime is classified "crisis" or "stress," briefings explicitly flag that historical patterns may not hold.

What we measure and publish

  • Forecast MAPE per series per horizon, monthly
  • Regime-classification confusion matrix, monthly
  • Sentiment-classification F1 on Nigerian financial news, quarterly
  • LLM hallucination rate against the public eval set, monthly
  • Coverage gaps — which sources lagged, which signals were missing
  • Model card for each fine-tuned release, including known limitations

Open source as accountability

The pipeline code, forecast engines, evaluation harness, and model weights will ship under permissive open licenses. This is the structural commitment to auditability — a bank's risk team, an academic researcher, or a regulator can verify our claims rather than take our word for it. Source-code release is currently pending; institutional partners can request early access.

What is out of scope

  • Personalised financial advice tailored to an individual's portfolio.
  • Forecasts for individual securities or credit-decisions on individual borrowers.
  • Trading-signal generation at sub-daily frequencies.
  • Any use where forecast error directly causes irreversible action without human judgement in the loop.

Reporting concerns

If you encounter a hallucinated output, a misclassified regime, or any behaviour that looks unsafe, write to francis.oyakhire@gmail.com. We treat these as production incidents and respond within 7 days.

Last updated: 2026-06-14