GRX · Foundations
How the Index is constructed: normalisation, weighting, aggregation and uncertainty
Index construction in The Global Risks Index, step by step: framework, data, missing values, normalisation, weights, aggregation, uncertainty, validation and release, with every choice shown.
- As of
- 17 November 2026
- Bacaan
- 18 minutes
- Bukti
- 16 sources, cited in order
01definition
What index construction means here
A plain definition of index construction as the Index practises it, and what the draft status means.
Index construction is the set of choices that turn many measurements into a few numbers a reader can use. Which indicators go in? How are they put on a common scale? How much does each count, how do they combine, and how sure can anyone be of the result? This article sets out how those choices are made in The Global Risks Index, the open index family stewarded by The Global Centre for Risk and Innovation (GCRI), The Global Risks Forum (GRF) and The Global Risks Alliance (GRA). It follows the order of the Index's own methodology page.
Two limits come first. The methodology, structure and definitions are a preparatory working draft (method 1.0.0-draft) circulated for review and comment, not for implementation, and no index value appears until an edition is released from a governed evidence cut. The Baseline Edition 2026 has an evidence close of 30 September 2026 and its values are not yet published 1. Everything below therefore describes a method, not a result. The Index is research and risk-analytics information: it is not a credit or ESG rating, not official statistics and not advice 1.
The vocabulary of risk follows the public literature. The IPCC describes climate risk as the result of dynamic interactions between hazards and the exposure and vulnerability of the affected system 3. The Sendai Framework, which the UN General Assembly endorsed in 2015, is the global reference for disaster risk reduction 4, and the Index carries an indicator on national strategies aligned with it 5. The Index adds a fourth strand, the lack of coping and adaptive capacity, and builds all four into its structure 1. For the science behind those terms see the GCRI primer on foundations of risk science.
02why-it-matters
Why a composite has to show its working
Composite indicators compress information, so their choices must be visible.
A composite indicator is useful because it is short, and risky for the same reason. Any single number hides choices about scale, weight and aggregation that a reader cannot see. The Joint Research Centre of the European Commission runs a Competence Centre on Composite Indicators and Scoreboards that offers methodological support and carries out statistical audits of composite indicators 6. The Index sets out its construction choices on its methodology page 1.
The Index's first principle is a family of indices rather than one score. It publishes 1 twenty domain indices, four pillar indices, a composite and six further instruments, and the composite never appears without every pillar and domain beside it 1. The second principle is classes with intervals rather than league tables: every value carries a 90% interval and countries are listed by name or code, never by value 1. The sections below show how each of the ten construction steps makes those two principles real.
The Fundamental Principles of Official Statistics were adopted by the Statistical Commission in 1994, and the General Assembly endorsed them in resolution 68/261 of 2014 7. The Index is not official statistics 1, so it does not claim their standing.
03ten-steps
Ten steps from the OECD and JRC Handbook
The ten construction steps and where each is described in the Index method.
The Handbook on Constructing Composite Indicators was published by the OECD in 2008 with the European Commission's Joint Research Centre 2. The Index methodology states that its ten construction steps follow that Handbook and the statistical audit practice of the JRC Competence Centre on Composite Indicators, referenced for method only 1. The Handbook's text is open as a PDF from the OECD, and every statement about the Index's own steps is attributed to the Index method page.
The ten steps, in the Index's wording, are these: theoretical framework; data selection; missing data; multivariate analysis; normalisation; weighting and aggregation; uncertainty and sensitivity; back to the data; links to other indicators; and visualisation and dissemination 1. The table below sets out what the Index does at each step.
| Langkah | What the Index does at this step | Where it appears |
|---|---|---|
| 1 Theoretical framework | A risk chain from hazard to outcome, four pillars, twenty domains, and an explicit polarity and risk component for every indicator 1 | Pillars, domains, indicator definitions |
| 2 Data selection | Indicator definitions with SDMX-compatible keys, crosswalks to source code lists, an evidence class from A to E and a licence register 1 | Indicators, data and sources |
| 3 Missing data | Flagged imputation at the indicator layer only; modelled share capped; coverage minimum of two thirds 1 | Indicator layer |
| 4 Multivariate analysis | Correlation structure, principal components and Cronbach's alpha per sub-index; incoherent components are re-specified before release 1 | Sub-indices |
| 5 Normalisation | Min-max with fixed goalposts, percentile rank or identity on native 0 to 10 scales, with skew and kurtosis treatment 1 | Sub-indices |
| 6 Weighting and aggregation | Equal or expert weights recorded per composite; five aggregation shapes; a floor of 0.1 before geometric means 1 | Domains, pillars, composite |
| 7 Uncertainty and sensitivity | At least 1,000 Monte Carlo runs, 90% intervals, class probabilities and Sobol indices with every release 1 | Every published value |
| 8 Back to the data | Driver decomposition to indicator level; every value opens to its inputs, sources and transformation chain 1 | Profiles |
| 9 Links to other indicators | Published crosswalks and correlation checks against established indices, used as validity evidence and never as inputs to re-weight 1 | Validation |
| 10 Visualisation and dissemination | Classes and intervals, never league tables; API, SDMX keys, tiles and bulk files; corrections and versions public 1 | Profiles, data access |
04framework-and-data
Step 1 and 2: framework and data selection
How the framework fixes the structure before any data arrive.
A composite is only as defensible as its framework. The Index fixes the framework first: a risk chain from hazard through exposure, vulnerability, impact and likelihood to capacity, response, financing and outcome, with four pillars and 1 twenty domain indices placed on that chain. Every indicator declares its place on the chain and its polarity, so aggregation never mixes directions silently 1. The four pillars are hazard and threat exposure, human and systems vulnerability, macro-financial and structural fragility, and lack of coping and adaptive capacity 1. Their domains are split five, eight, five and two respectively, which sums to twenty 1.
Data selection comes second. The Index counts 451 indicator definitions, 106 composites and 501 crosswalks, and 205 registered sources of which 115 are admitted for public display 1,8. Each indicator has a unit, a polarity, a risk component, levels, sources and crosswalks 5. Identifiers follow the SDMX approach: SDMX is an international standard for statistical data exchange, meant for national statistical offices, central banks and institutions to standardise their data operations 9. The full list of indicators and their layers is on the indicators page, and a companion article covers the indicators in depth.
Hazards enter through the exposure pillar. The Index's Hazard Standard lists 14 hazard domains, 120 families and 905 hazards, including 281 hazards drawn from the 2025 hazard information profiles 1,10. See the hazards page and the companion article on the hazard catalogue.
05missing-data
Step 3: missing data is not zero
The Index shows missing, imputed and observed zero differently.
The third principle of the Index is that no data is not zero. A missing input is shown as missing, an imputed input is flagged, and a value is withheld when coverage falls below the method's minimum. An empty profile, an unavailable source and an observed zero render differently 1.
Imputation happens only at the indicator layer and is always flagged. The method names three techniques, dimension mean, indicator median and multiple imputation, and each indicator definition declares which it uses. Each composite caps the share of its weight that may rest on modelled or imputed inputs, and withholds its value below a coverage minimum of two thirds 1. A pillar likewise needs two thirds of its domain indices; if it has fewer it is withheld, and a composite whose pillar is withheld is withheld too, with the pillars shown alone 1.
Coverage is published as a vector per country and indicator, never as a single complete flag. Gaps cluster in fragile and small-island states, so their coverage is shown first 1. The reason is practical: a country with few data should not look safe because the data are few. This is also why status glyphs separate observed, reported, estimated, inferred, forecast, simulated, scenario and expert-input values 1.
06normalisation
Step 4 and 5: multivariate analysis and normalisation
How raw indicators are put on a common 0 to 10 scale.
Before indicators are combined, the Index checks whether they belong together. For each sub-index it examines the correlation structure, runs principal components and reports Cronbach's alpha; components that do not cohere are re-specified before release 1. There are 86 sub-indices in the draft structure 1, each formed from normalised, weighted and aggregated components inside a domain.
Normalisation puts different units on the same footing. The method allows three forms: min-max with fixed goalposts, percentile rank, or identity where an indicator is already on a native 0 to 10 scale. Skewed indicators are treated by winsorisation or a log transform above an absolute skewness of 2 and a kurtosis of 3.5 1. Fixed goalposts matter because they keep a value comparable across editions; moving goalposts would make a country look better or worse for reasons that have nothing to do with risk. The direction of each indicator, its polarity, is declared in its definition 1,5.
The effect of every choice here is tested rather than assumed. The Index Lab runs the published construction and validation methods in a browser on numbers a reader enters, and never shows an Index value 1.
07aggregation
Step 6: weighting and aggregation
Five aggregation shapes, and why the choice of shape is a choice about compensation.
Weights are equal or set by experts, and are recorded for every composite. The baseline gives equal weights at pillar level and records them per release 1. Aggregation then combines components, and the shape of the aggregation decides how far strength in one component can offset weakness in another. The Index uses five shapes and each domain declares which one it uses 1.
| Shape | Behaviour | Domain indices using it |
|---|---|---|
| Geometric mean | Weighted geometric mean after a floor of 0.1; partial compensation 1 | Nine: fiscal and public debt, climate physical hazard, water security, health security, socio-economic vulnerability, energy security, infrastructure fragility, cyber and lack of coping capacity 1 |
| Mazziotta-Pareto | Mean adjusted by the spread across components; unbalanced profiles read as riskier than balanced ones with the same mean 1 | Three: financial stability, food insecurity and conflict 1 |
| Arithmetic mean | Weighted arithmetic mean; full compensation, used where components are substitutes 1 | Six: external dependence, climate transition exposure, environmental pressure, displacement pressure, governance weakness and technology concentration 1 |
| Geometric complement | Complement of the geometric mean of complements; any single high exposure lifts the index 1 | One: multi-hazard exposure 1 |
| Product | Product of normalised components; risk exists only where all components are present 1 | One: space-weather exposure 1 |
08composite
The composite and the coupled composite
The composite formula, the floor, and the coupled composite, with arithmetic that shows what they do.
Each domain index feeds one pillar, and each pillar value is an expert-weighted geometric mean of its domain indices, with equal weights in the baseline 1. The composite is the fourth root of the product of the four pillars: composite = (HTE x VUL x MSF x LCC)^(1/4), with each pillar on a 0 to 10 scale and floored at 0.1 1.
A short piece of arithmetic on that formula shows why the form and the floor matter. This is an illustration of the formula, not an Index value. Take four pillars at 4.2, 5.1, 3.6 and 4.8, the starting positions of the simulator on the methodology page: their product is about 370.1 and its fourth root is about 4.39. Now take three pillars at 5 and one at 0. Without a floor the product is zero and the composite collapses to zero whatever the other three say; with the floor of 0.1 it is (0.1 x 5 x 5 x 5)^(1/4), about 1.88. The method states the floor of 0.1 for each pillar 1, and the arithmetic above shows what it does to a pillar at zero. The same logic applies at a smaller scale: two components at 8 and 2 with equal weights have an arithmetic mean of 5.0 and a geometric mean of 4.0.
The coupled composite adds systemic risk. Coupled = Composite x (1 + kappa x S), where S is a systemic coupling score between 0 and 1 computed on a typed dependency network and kappa is a coupling multiplier disclosed for each release, 0.25 in the baseline. It is reported beside the standalone composite and never instead of it 1. With S at 0.4 and kappa at 0.25 the multiplier is 1.10, so the illustrative composite of 4.39 becomes about 4.82.
09uncertainty
Step 7: uncertainty, intervals and classes
How the Index states how sure it is, and how a value becomes a class with a probability.
A single number conveys false precision. The Index runs at least 1,000 Monte Carlo draws over the declared alternatives for normalisation, weights, imputation and aggregation, and publishes 90% intervals, class probabilities and Sobol first-order and total-order sensitivity indices with every release 1. Each value therefore comes with the range that holds 90% of its draws, and with a measure of which choice moves it most 1.
Values are then placed in five fixed classes on the 0 to 10 scale: very low (0.0 to under 2.0), low (2.0 to under 3.5), medium (3.5 to under 5.0), high (5.0 to under 6.5) and very high (6.5 to 10.0). A class is printed with its probability, the share of draws inside it, and below 60% the adjacent class is printed too, for example high at 54% and very high at 41% 1. Fixed bounds keep classes comparable across editions. Intervals show when two countries cannot be told apart, and lists are ordered by name or code, never by value 1.
10validation
Steps 8 to 10: back to the data, validity and dissemination
How a value can be traced, tested and published.
Step 8 is to go back to the data: every value opens to a decomposition by driver, down to the indicator, with its inputs, sources and transformation chain 1. The chain is recorded with W3C PROV, under which provenance is information about the entities, activities and people involved in producing a piece of data, usable to assess its quality, reliability or trustworthiness 11. In the Index every value carries its source release, licence, retrieval time, method version and transformation chain, and a correction propagates to every dependent value with a public log entry 1.
Step 9 links the Index to other indicators. Published crosswalks and correlation checks against established indices are used as validity evidence and never as inputs to be re-weighted 1. The method also lists validation in public: back-testing of structural and outlook values against realised outcomes such as disaster mortality and losses, acute food insecurity phases, conflict onset, debt distress episodes and displacement; forecast verification with the Brier skill score, reliability diagrams, ROC area and CRPS; an independent statistical audit of each major methodology version; golden-value tests for every computation kernel; and a country right of reply with a public correction log 1. A rank-reversal test is part of each edition: adding or removing a country or an indicator must not reorder the others beyond the uncertainty already reported 1.
Step 10 is dissemination: classes and intervals rather than league tables, and machine access through an API, SDMX keys, vector tiles and bulk files 1. The FAIR principles call for data that are findable, accessible, interoperable and reusable, with machine-readable metadata 12. Each edition is deposited with its definitions, method version and inputs under a DOI, a digital identifier of an object 1,13, and the release run names a Zenodo deposit for tiles and bulk files 1. Zenodo is a repository launched in 2013 and provided by CERN 14. Developers can start at the developers page and the methods catalogue.
11edisi
Editions, revisions and reproducibility
How a change to a value or a method is made without erasing the past.
Method carries a semantic version. A change that breaks comparability is back-cast where inputs allow, and otherwise the series carries a declared break and comparisons across it are refused 1. Every edition and every correction is a vintage with its snapshot, frozen inputs and method version, and nothing is overwritten, so a study can be rerun on the information available at the time 1.
Each release publishes definitions, method versions, code paths, random seeds and content digests, so any value can be recomputed, and every kernel is tested against hand-derived golden values 1. An edition is also a workflow of 13 jobs with 2 approval gates, one for an independent statistical audit and one for steward review, before anything is published 1. A companion article covers the full life of an edition: editions, corrections and governance, and the governance page describes who holds the review gates 15.
12reuse
Reusing the method and the data
What the licences allow, and where to start.
Nexus-authored text is intended for CC BY 4.0 on release, a licence under which anyone may share and adapt material, even commercially, if appropriate credit is given 16. The terms are in the CC BY 4.0 deed. Every value carries the attribution of its sources, which keep their own licences 1. The Index's data page describes the licence gate that decides what enters a public value: nothing is shown whose licence does not admit that use 8. The companion article open data and licences goes through the sources, licences and the rules for reuse.
Anyone can comment on the method: the Index runs an open consultation on normalisation, weighting, aggregation, imputation, uncertainty and validation, and every comment receives a disposition in the public consultation record 1. Governance sets a consultation period of at least 30 days for method changes 15. To try the choices described here on your own numbers, use the Index Lab. The standards that tie the Index to the wider estate are on Nexus Standards.
13Open questions
What is still open
Where the method is unsettled and how comment is handled.
The method is version 1.0.0-draft, a preparatory working draft circulated for review and not for implementation 1. Normalisation, weights, aggregation, imputation, uncertainty and validation are all open for comment, and every comment receives a recorded disposition. Accepted changes are versioned and, where inputs allow, applied back to earlier editions 1.
Three parts of the method stand out. The method describes a coupling layer built on a typed dependency network with six layers 1. It lists forecast verification scorecards among the outputs of the Outlook instrument, which gives one to 36 month event probabilities 1. And it describes machine access through an API, SDMX keys, vector tiles and bulk files 1. No Index value is published yet, and the Index Lab runs the published methods on numbers a reader enters.
14Pengelolaan
One index, three institutions
Who holds which responsibility for the Index.
The Index is stewarded by three independent not-for-profit institutions of the Nexus Consortium, and technology and service providers work at arm's length with no role in stewardship 1,15. The governance page lists GCRI for evidence, methods, risk ontology, data discipline, verifiable records and correction; GRF for leadership, institutional dialogue and the stewardship of Nexus Universe; and GRA for finance, insurance and development readiness 15. Related reading from GRF is its article on risk governance. The Index feeds the Reports platform, and Risk Academy teaches how to read it.
This article describes a preparatory working draft (method 1.0.0-draft) of the Index method. It shows no Index value, because none is published, and its arithmetic examples use illustrative inputs and the published formulas only. The Index is research and risk-analytics information: not a rating, not official statistics, not an emergency warning, not a benchmark and not advice.
Pertanyaan
Pertanyaan dan jawaban
How does the Global Risks Index construct its composite?
It computes indicators, normalises them to 0 to 10, aggregates them into sub-indices and twenty domain indices, combines the domains into four pillars and takes the fourth root of the product of the pillars, each floored at 0.1 1.
What does the floor in the composite do?
Each pillar is floored at 0.1 before the fourth root, so a pillar score of zero enters the product as 0.1 and does not set the composite to zero. A composite whose pillar is withheld is itself withheld 1.
How is uncertainty reported?
Each value carries a 90% interval from at least 1,000 Monte Carlo runs over normalisation, weights, imputation and aggregation, plus a class probability and Sobol sensitivity indices 1.
Does the Index rank countries?
No. Countries are placed in five fixed classes with a stated probability and listed by name or code, never by value; below 60% class probability the neighbouring class is printed too 1.
What happens when data are missing?
Missing is shown as missing, imputation is flagged and only happens at the indicator layer, and a value is withheld below a coverage minimum of two thirds 1.
Can I recompute a value?
That is the aim: each release publishes definitions, method versions, code paths, random seeds and content digests, and kernels are tested against hand-derived golden values 1.
Are there Index values to download now?
Not yet. The methodology is a working draft and values are not yet published; the Baseline Edition 2026 has an evidence close of 30 September 2026 1.
Ketentuan
Glosarium
- Composite indicator
- A single number built from several indicators by normalising, weighting and aggregating them.
- Normalisation
- Putting indicators with different units on a common scale, here 0 to 10.
- Aggregation
- The rule that combines components into an index; the Index uses five shapes.
- Floor
- A minimum value, 0.1 for a pillar, applied before a geometric mean.
- Monte Carlo analysis
- Repeated calculation with varied choices to show how much the result depends on them.
- Sobol index
- A measure of how much of the variation in a result comes from each input or choice.
- Class probability
- The share of draws that fall inside the printed risk class.
- Komposit terkait
- The composite multiplied by (1 + kappa x S), where S is a systemic coupling score.
Terkait
Further reading across the network
- The Index methodology
- Data and sources, with the licence gate
- Governance, review and corrections
- The 451 indicators
- Domain bahaya
- Dua puluh indeks domain
- Index Lab: try the methods on your own numbers
- Katalog metode
- How the Index is built: method and limits
- Foundations of risk science
- Nexus Standards
Corrections and right of reply
Every value names its edition. Report a problem with a value or reply as a government through the forms of this site; replies are published beside the country page.
Sumber
References
16 sources, numbered in the order they are first cited
- The Global Centre for Risk and Innovation (The Global Risks Index) (Page GRX-MTH, method 1.0.0-draft, Baseline Edition 2026 (evidence close 2026-09-30, values not yet published)). The Global Risks Index: methodology (1.0.0-draft). risksindex.com. Accessed 2026-11-17.
- Organisation for Economic Co-operation and Development (OECD) and European Commission Joint Research Centre (2008-08-22). Handbook on Constructing Composite Indicators: Methodology and User Guide. doi.org. Accessed 2026-11-17.
- Intergovernmental Panel on Climate Change (2022-02-28). Climate Change 2022: Impacts, Adaptation and Vulnerability. Annex I: Global to Regional Atlas (Working Group II contribution to the Sixth Assessment Report). www.ipcc.ch. Accessed 2026-11-17.
- United Nations General Assembly (2015-06-03). Resolution 69/283: Sendai Framework for Disaster Risk Reduction 2015 to 2030 (A/RES/69/283). documents.un.org. Accessed 2026-11-17.
- The Global Centre for Risk and Innovation (The Global Risks Index) (Page GRX-IND, method 1.0.0-draft, Baseline Edition 2026 (evidence close 2026-09-30, values not yet published)). The Global Risks Index: indicators (Indicator Standard 1.0.0). risksindex.com. Accessed 2026-11-17.
- European Commission, Joint Research Centre (Knowledge4Policy page, current as of 17 November 2026). Competence Centre on Composite Indicators and Scoreboards (CC-COIN). knowledge4policy.ec.europa.eu. Accessed 2026-11-17.
- United Nations Statistics Division (UN Statistics Division page, current as of 17 November 2026). Fundamental Principles of Official Statistics. unstats.un.org. Accessed 2026-11-17.
- The Global Centre for Risk and Innovation (The Global Risks Index) (Page GRX-DAT, method 1.0.0-draft, Baseline Edition 2026 (evidence close 2026-09-30, values not yet published)). The Global Risks Index: data and sources. risksindex.com. Accessed 2026-11-17.
- SDMX sponsor organisations (SDMX website, current as of 17 November 2026). SDMX: the standard for statistical data and metadata. sdmx.org. Accessed 2026-11-17.
- The Global Centre for Risk and Innovation (The Global Risks Index) (Page GRX-HAZ, method 1.0.0-draft, Baseline Edition 2026 (evidence close 2026-09-30, values not yet published)). The Global Risks Index: hazards (Hazard Standard). risksindex.com. Accessed 2026-11-17.
- World Wide Web Consortium (W3C) (2013-04-30). PROV-Overview: An Overview of the PROV Family of Documents. www.w3.org. Accessed 2026-11-17.
- GO FAIR Foundation (GO FAIR page, current as of 17 November 2026). FAIR Principles. www.go-fair.org. Accessed 2026-11-17.
- International DOI Foundation (DOI Foundation page, current as of 17 November 2026). What is a DOI?. www.doi.org. Accessed 2026-11-17.
- European Organization for Nuclear Research (CERN) and OpenAIRE (Zenodo about page, current as of 17 November 2026). About Zenodo. about.zenodo.org. Accessed 2026-11-17.
- The Global Centre for Risk and Innovation (The Global Risks Index) (Page GRX-GOV, method 1.0.0-draft, Baseline Edition 2026 (evidence close 2026-09-30, values not yet published)). The Global Risks Index: governance. risksindex.com. Accessed 2026-11-17.
- Creative Commons (Licence deed, current as of 17 November 2026). Attribution 4.0 International (CC BY 4.0) Deed. creativecommons.org. Accessed 2026-11-17.
Catatan
Editorial record
Facts, counts and method states are given as of 17 November 2026, from method 1.0.0-draft and the Baseline Edition 2026 pages of the Index (evidence close 2026-09-30, values not yet published). This article is reviewed at least every 270 days and whenever the method version, an edition or a cited count changes.
Published by the Index data desk of The Global Centre for Risk and Innovation under the scientific and editorial integrity policy. Text, figures and data tables of this article are licensed CC BY 4.0; cited data keep their own licences. To report an error or request a correction, use corrections and concerns or the Index's right of reply; corrections are dated and listed with the page.
Cite this article
Index data desk, GCRI (2026, November 17). Index construction: how the Global Risks Index is built. The Global Risks Index. https://risksindex.com/index-construction-method/
@online{index2026indexconstructionmetho,
author = {{Index data desk, GCRI}},
title = {{Index construction: how the Global Risks Index is built}},
publisher = {The Global Risks Index},
year = {2026},
month = {11},
url = {https://risksindex.com/index-construction-method/},
language = {en}
} TY - ELEC AU - Index data desk, GCRI TI - Index construction: how the Global Risks Index is built PB - The Global Risks Index PY - 2026 DA - 2026/11/17 UR - https://risksindex.com/index-construction-method/ LA - en ER -
Republish
Republish or translate this article free of charge, in print or online, under CC BY 4.0: keep the byline and The Global Risks Index, link to the original and note any change. Figures, maps and twins made by the institutions carry the same licence; third-party data keeps its own credit line.
Attribution to paste
<p>How the Index is constructed: normalisation, weighting, aggregation and uncertainty — Index data desk, GCRI, The Global Risks Index. First published 17 November 2026 at <a href="https://risksindex.com/index-construction-method/">https://risksindex.com/index-construction-method/</a>. Republished under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/).</p>