Case 01 · Institutional treasury forecasting
Nation-scale treasury cash-flow & liquidity forecasting
International Monetary Fund · Central Bank of Somalia
2023 – 2024
96%
Accuracy of the selected SARIMA model on held-out data at a six-month horizon, chosen from four univariate model families
Rubric first
Evaluation criteria were fixed before any model was trained, so the winning model was chosen on evidence, not preference
Client-owned
The central bank’s own engineers now run, maintain and extend the forecasting code. No permanent external dependency
- Context
- A central bank needed forward visibility into Treasury Single Account cash flows to plan national liquidity. The IMF commissioned an applied research and implementation effort with a clear constraint: transparent, auditable univariate models, and no large language models.
- Mandate
- Sole lead machine learning researcher and developer, from raw data through handover.
- Approach
- Cleansed and normalized the full transaction history of the Treasury Single Account and its ancillary accounts.
- Defined the evaluation rubric before modelling, then compared four model families, from classical statistical baselines to neural approaches.
- Documented every step in reproducible notebooks and a formal technical report for the IMF.
- Outcome
- The forecasting software was adopted by the central bank. Its Mogadishu-based engineering team was trained to operate, maintain and extend the models independently.
- Transferable to
- Corporate treasury and cash forecasting, public revenue forecasting, and any finance function that needs a baseline it can trust and own.
The technical report is held privately by the IMF; transaction data and figures are not disclosed. This engagement is independent of the Anominal AI platform.
