A score is a starting point.
The method is the context.

What the index measures

This project describes food-system vulnerability through trade dependencies, projected exposure and response capacity. It organizes seven indicators into three dimensions: sensitivity (five food-system indicators), exposure (the projection-based climate indicator, called “hazard” in the dataset), and coping capacity (ND-GAIN readiness).

This is the project’s composite framework. In IPCC AR6 terminology, hazard is a separate component of risk, rather than a part of vulnerability itself. The grouping should not be read as an official IPCC definition or an endorsed index.

The data describe 2014–2023. The index does not predict when a disruption will occur or how many people would go hungry.

The seven indicators

Why dependency can exceed 100%

Gross dependency divides imports by domestic supply: production plus imports minus exports. Processing, re-exports and stock flows can produce values above 100%. Each valid item ratio is bounded before averaging. Nonpositive supply is undefined; coverage below 80% is insufficient evidence for a scored input, not low dependency.

The food basket uses positive food kcal/person/day weights; feed uses positive feed tonnes. Group contributions sum item contributions rather than recomputing group ratios. Net dependency + self-sufficiency = 100% for valid ratios; gross dependency + self-sufficiency need not. These baskets do not allocate imported units to final food or feed use.

From indicators to scores

Normalization uses pooled reference-imputation p01/p99 winsorization before goalpost clipping, scaling and direction adjustment. Some goalposts are fixed; fitted component anchors are unavailable in the publication. Readiness is reversed into a deficit; climate components are normalized separately, then averaged.

The floor is affine, not simply max(0.01, t). The lab distinguishes known stages from unavailable fitted anchors and cell-level MICE masks.

Weights sum to one. Equal weights assign 1/7 to each indicator. Hierarchical weights assign 1/15 to each of five sensitivity indicators and 1/3 each to exposure and coping. Exposure-only assigns 1/6 to each retained indicator. PCA weights are not supplied. Weighted log terms are not additive percentages of a composite score.

A geometric mean reduces compensability compared with an arithmetic average, but it is not a worst-case rule: low vulnerability on one indicator can still offset higher vulnerability on another.

The published headline is the equal-weight score averaged across 20 imputations. It has no single reference-input decomposition. The hierarchical method is an alternative, not the calculation behind that headline.

Compare the four reference methods to isolate weighting choices. Comparing the headline mean against a reference method also changes the imputation summary, so that difference is not purely a weighting effect.

Frozen HHI precision

The publication exporter rounded supplier_hhi_norm to one decimal because its HHI-name rule ran before its normalized-field rule. That loses information on a 0–1 field. Raw HHI and full-precision evidence have different precision; an inverse formula cannot recover the damaged normalized values or exact composite contributions.

Country bars retain the frozen values. The exporter ordering is corrected for future exports, but this evidence release does not rewrite the published scores or ranks. Read raw reconstruction, rounded publication agreement and historical input-hash agreement separately.

Reading supplier connections and unit values

Group HHI sums positive mapped product values by supplier before squaring shares. Its full denominator is independent of top-N arcs. “Other” is a display remainder; its HHI is the sum of omitted supplier terms, not its squared aggregate share. The country headline is the diet-calorie-weighted mean of group HHIs, not whole-basket HHI or an average of HS6 HHIs.

USD/kg = matched nominal USD / (1,000 × positive reported tonnes). Both sums use the same eligible rows; all-value totals and quantity coverage remain separate. Same-year HS6 comparisons are descriptive. Group unit values mix products and do not establish savings, inefficiency, landed prices, tariff effects or quality-adjusted differences.

Arcs are supplier-to-importer relationships, not shipping routes or calorie flows. Coordinates never decide economic inclusion. Native BACI 490 retains the executed Taiwan (TWN) proxy. Food/feed links share only a coarse FBS/BACI group concordance.

Source vintages are separate evidence

The 2014–2023 frozen publication, current local FAOSTAT artifacts and August 2026 ND-GAIN archive are separately identified and hashed. Retrieval time and export build time are not source release dates. Local food-import-burden observations differ from frozen values in 379 matched rows; Egypt’s 2023 row is local 38% for period 2022–2024 versus frozen 37%. Three-year periods remain visible.

ND-GAIN’s 2026 economic-readiness component transitions from Doing Business to Business Ready (B-READY), backfilled to 2022, with provider regional estimates where inputs are pending. A continuous 0–1 scale does not establish like-for-like trends across that break. The lab links the provider codebook and release note; country-specific placeholder status is not inferred.

Observed-row flags mean an input row exists, not that it was directly measured. Provider estimation, inferred zeros and frozen project imputation are different stages. The historical IDR configuration hash differs from the current registered configuration; numerical agreement does not erase that provenance mismatch. Inspect both hashes in the evidence quality panel.

Missing data and uncertainty

The pipeline fills missing indicator observations using multiple imputation. The exported headline summarizes 20 completed datasets. Minimum and maximum scores form an imputation range, not a confidence interval and not a complete account of methodological uncertainty.

The site displays this range only with the mean equal-weight method. Alternative reference methods have no corresponding imputation bands in this snapshot. Missing exported values appear as unavailable, never as zero. Raw and normalized fields may reflect different pipeline stages and rounding, so the normalized score should not be reconstructed from a rounded raw display.

Checking the cereal calculation

FAO publishes a cereal import-dependency series. The standalone model recomputes that cereal measure and compares it against FAO’s published series. This checks a specific dependency calculation; it does not validate the full composite or its ability to predict food insecurity.

The 95th-percentile absolute error is 12.56 percentage points. Tail discrepancies remain despite the strong ranking agreement.

Research limitations

The repository’s September 2026 methods review calls for major revision. The website presents the existing exported results so they can be inspected, while keeping these limitations visible:

• Outcome validation needs revision. The review identifies dropped censored undernourishment observations and use of the same outcome during model development. It cannot be treated as a clean held-out validation.
• Exposure is projection-based. Cereal-yield projections include extreme values; their normalization can move rankings substantially. Population change is demographic pressure, not climate hazard.
• Indicators are correlated. Exposure and coping contain related development information. Nominal weights do not necessarily describe their effective influence.
• Production concentration is provisional. Its tonnage weighting makes heavy products dominant and is not a direct measure of ecological crop diversity.
• Supplier aggregation matters. Diet-calorie weighting can emphasize concentration even where imported shares are small.
• Time resolution varies. Food import burden comes from three-year averages; projection-based climate exposure contributes little annual change. An annual row should not imply every indicator was independently observed that year.
• Uncertainty is incomplete. Supplied ranges cover imputation variation for one method. They do not cover alternative normalization, weighting, source revisions or structural assumptions.

These limitations are reasons to examine components and compare methods, rather than interpret a single rank as a precise forecast. Research updates will require a new versioned data export.

Sources and provenance

Exported 2026-08-28 from pipeline revision 860875e. Coverage: 178 countries and 1,780 country-years.

• CEPII BACI: bilateral trade, HS12, V202601.
• FAOSTAT Food Balance Sheets: supply, production, food and feed baskets.
• FAOSTAT Food Security: import burden and cereal replication reference.
• ND-GAIN Country Index: exposure projections and readiness.

The source repository is currently private. The published data snapshot and field guide are available on the downloads page.