Financial data engineering · Applied ML · 2026
IDA Financial Data
Control Tower
An interpretable anomaly-detection and analyst-review layer for public World Bank IDA commitment and gross-disbursement data.
Live read-only deployment using the current public-data run. Independent portfolio prototype, not a World Bank product.
The mandate
Make unusual financial records visible before they become reporting risk.
IDA reporting spans thousands of country, aggregate, annual, and quarterly observations. A record can be structurally valid and still deserve attention: a sharp year-over-year movement, an unusual financing mix, or a total that no longer reconciles to its components.
The design brief was not to label anomalies as errors. It was to create a reproducible control layer that prioritizes evidence for a human analyst, combining financial rules, statistical comparison, and multivariate detection without hiding judgment behind a score.
My contribution
Designed and built end to end.
I independently defined the control philosophy, built the ingestion and feature pipeline, calibrated the segmented Isolation Forest models, implemented the FastAPI review service, designed the analyst interface, wrote the evaluation harness and technical documentation, and translated the local system into an Azure production architecture.
Control logic
Three independent detector families. One review queue.
Deterministic controls
Missing fields, invalid categories, duplicate grain, negative values, and component-to-total reconciliation.
Statistical comparison
Comparable-period and year-over-year movements, segmented so annual and quarterly records are never mixed.
Isolation Forest
Four segmented models surface unusual multivariate combinations across magnitude, component shares, change, and ratios.
End-to-end system
From public financial data to auditable analyst action.
The local implementation is complete. The Azure layer is a documented target design and has not been deployed.
- 01Ingest
Paginated API retrieval, declared-count validation, and raw persistence.
- 02Standardize
Financial categories, periods, entity types, and comparison features.
- 03Detect
Rules, statistical controls, and segmented Isolation Forest scoring.
- 04Consolidate
Agreement and materiality inform severity; ML alone cannot create critical status.
- 05Review
Analyst states, outcomes, notes, evidence, version checks, and append-only audit events.
Evidence
Test the control behavior, not a claim of perfect prediction.
The fault-injection harness verifies selected software and control behavior. It does not establish production accuracy or a real-world false-positive rate. Precision remains intentionally suppressed until enough cases have been resolved by human review.
Responsible use
A prioritization instrument, not an automated verdict.
The anomaly score is normalized within each run and segment; it is not a probability of error. The 2% contamination assumption manages analyst workload, it does not imply that 2% of the data is wrong.
Country records enter the statistical and ML models; aggregates remain available to deterministic controls. Low-denominator ratios are suppressed, and critical severity requires more than an ML-only signal.
This independent prototype uses public World Bank data. It is not affiliated with, endorsed by, or deployed for the World Bank Group, and it does not access internal IDA systems.
Technical evidence