Methodology

Scoring Methodology

How platform scores and rankings are calculated, normalized, weighted and adjusted for confidence.

Overview Scoring Confidence AI Assessments Revisions

Prototype Notice:This prototype currently uses mock and illustrative data. Source records, update schedules, confidence scores and methodology examples are placeholders for the future production system.

Score Design Principles

Score Objective

Each score states what it measures and for whom.

Score Direction

All scores increase with the better outcome after direction adjustment.

Indicator Selection

Indicators are chosen for coverage, comparability and stability.

Update Frequency

Scores recompute on the slowest input's cadence.

Revision Policy

Scores are recomputed when methodology or inputs change.

Version Control

Each score carries a version tag and effective date.

Ranking Tie Rules

Ties share the same rank; the next rank skips accordingly.

Confidence Adjustment

Scores are down-weighted when coverage or freshness is weak.

Scoring Pipeline

1
Raw Source Value
Original value as received from the source record, with timestamp and units preserved.
2
Validation
Range, unit, type and date checks; suspect values flagged for review.
3
Unit Normalization
Convert to platform-standard units (USD, %, index points, persons).
4
Period Alignment
Align to a common observation period (monthly, quarterly, annual).
5
Direction Adjustment
Invert indicators where lower is better so all scores point the same way.
6
Indicator Score
Indicator-level 0–100 score from the normalized value.
7
Dimension Score
Weighted aggregation of indicators within a dimension.
8
Composite Score
Weighted aggregation of dimension scores into the overall score.
9
Confidence Adjustment
Down-weight scores where coverage or freshness is weak.
10
Published Platform Score
Final score with confidence band, version tag and revision history.

Normalization Methods

Min-Max Scaling

Linear scaling between observed minimum and maximum.

Percentile Ranking

Position within the cross-country distribution.

Z-Score Standardization

Distance from mean in standard deviations.

Threshold Bands

Discrete bands tied to policy or risk thresholds.

Direction Reversal

Inverts lower-is-better indicators.

Log Transformation

Compresses heavy-tailed distributions.

Winsorization

Caps extreme values at chosen percentiles.

Peer-Group Normalization

Scales within an income or regional peer group.

Prototype methodology options — not production claims.

Illustrative Prototype Weights

Mock

Score	Dimension	Weight	Type	Rationale	Confidence	Version
Country Risk	Macro Stability	25%	Expert Weight	Anchors solvency view.	High	0.2.0
Country Risk	External Position	20%	Expert Weight	FX and reserves matter for shocks.	High	0.2.0
Country Risk	Public Finance	20%	Expert Weight	Debt sustainability.	High	0.2.0
Country Risk	Political	15%	Hybrid Weight	Institutions and stability.	Moderate	0.2.0
Country Risk	Financial Sector	10%	Expert Weight	Banking system buffers.	Moderate	0.2.0
Country Risk	Structural	10%	Data-Driven Weight	Long-term capacity.	Moderate	0.2.0
Business Environment	Contract Enforcement	30%	Expert Weight	Operational certainty.	Moderate	0.1.0
Business Environment	Property Rights	25%	Expert Weight	Investment protection.	Moderate	0.1.0
Business Environment	Access to Finance	25%	Hybrid Weight	Working capital availability.	Moderate	0.1.0
Business Environment	Regulatory Burden	20%	Expert Weight	Compliance cost.	Moderate	0.1.0

Missing Data Rules

Never silently replace missing values with zero.
Display 'Unavailable' where appropriate.
Reduce coverage percentage when fields are missing.
Reduce confidence when important fields are missing.
Avoid high-confidence composite scores when coverage is weak.
Use estimates only when explicitly labeled.
Show the last available observation date.
Distinguish delayed data from unavailable data.

Outlier Rules

Extreme values are flagged for review, not silently dropped.
Suspected unit errors trigger mandatory review.
Sudden revisions trigger mandatory review.
Conflicting sources are visible to the user.
Outliers are never auto-deleted without a recorded reason.