Skip to content

Risk Academy, led by the Nexus Consortium · Risk Science and Methods · Department of Quantitative Risk Analysis

Learning from data without hiding the error

இடர் பகுப்பாய்வுக்கான AI மற்றும் இயந்திரக் கற்றல்

Machine learning entered risk analysis through its methods, not its headlines: emulators that stand in for slow hazard and catastrophe models, classifiers of damage from imagery, anomaly detectors on sensor streams and foundation models trained on satellite archives.

Map aI and machine learning for risk analysis on a real placeSee the courses

Illustrative: layers of units and their weighted connections, an activation front sweeping from input to output

01வரம்பு

What the discipline covers

The subtopics its courses and labs teach.

பாடநெறிகள்
In preparation
நிலைகள்
In development
ஆய்வகங்கள்
In development
  • Supervised learning for risk outcomes
  • Emulators and surrogates of physical and catastrophe models
  • Uncertainty quantification in machine learning
  • Conformal prediction and distribution-free coverage
  • Anomaly detection and change detection
  • Geospatial foundation models and satellite embeddings
  • Causal inference and machine learning
  • Validation, drift monitoring and model risk for machine learning
  • Fairness and bias in risk models

02வரைபடங்கள்

Build a twin of any place

Search any place or open the world; the layers and generators follow aI and machine learning for risk analysis on the planet map, the globe and in 3D.

The central Apennines: L'Aquila, Amatrice and Norcia: a flat plan of the faculty twin. Terrain 24 to 2,805 m; 3 reference earthquakes, M6.2 to M6.6; Earthquakes M4.5 and above, past week: read live when the 3D twin opens; USGS ShakeMap intensity contours: read live when the 3D twin opens. Contours every 500 metres from the elevation grid; reference earthquakes as marks.
The central Apennines: L'Aquila, Amatrice and Norcia, Italy. Flat plan of the same data: relief, the three reference epicentres with magnitude and depth, and the numbers below; the sampler's figure shows one precomputed field with its parameters. The 3D twin opens on devices that can draw it, once its data and code have been checked.
  • Terrain 24 to 2,805 m
  • 3 reference earthquakes, M6.2 to M6.6
  • Earthquakes M4.5 and above, past week: read live when the 3D twin opens
  • USGS ShakeMap intensity contours: read live when the 3D twin opens
Credits and dates per layer are listed under the stage. Relief in the twin is exaggerated ×1.5 for legibility.

Drag to move, right drag or Shift with a drag to turn and tilt, Ctrl with the wheel to zoom. Keyboard: arrows move, Shift with arrows turn and tilt, + and – zoom, 0 resets, V steps through the views, Escape clears.

Flat plan of the same data: relief, the three reference epicentres with magnitude and depth, and the numbers below; the sampler's figure shows one precomputed field with its parameters. The 3D twin opens on devices that can draw it, once its data and code have been checked.

The boundaries, names and designations shown do not imply any judgement on the legal status of any territory, or any endorsement or acceptance of such boundaries. Countries: Natural Earth (public domain).

03The subject

The subject, methods and measures and debates and bodies

The discipline

The field and its scope

For risk analysts and data scientists who use or review machine learning in a risk estimate and need to know when its outputs can be trusted and how its errors propagate.

Machine learning for risk analysis studies the use of learned models to classify and predict risk outcomes, to emulate slow physical and catastrophe models, to detect anomalies and change, and to quantify the uncertainty of those outputs so that they can enter a risk estimate honestly. The field sits between statistics, computer science and the domain sciences: a flood emulator, a default classifier and a satellite damage detector share the same questions about training data, generalisation, calibration and drift. Its users are modellers who need speed, analysts who need predictions where no physical model exists, and validators who must decide whether a learned model is fit for the decision.

The discipline is organised around the learning task (supervised, unsupervised, causal), the uncertainty tools (Bayesian and ensemble methods, conformal prediction with distribution-free coverage, calibration), the validation regime (held-out tests, temporal and spatial splits, drift monitoring) and the governance of fairness and explainability where models affect people. Geospatial foundation models and satellite embeddings have brought pre-trained representations into exposure and damage mapping. The field's central tension is between accuracy on the training distribution and reliability where risk lives: in the tail, after a regime change, and in places with no labels.

Measures

Quantities and units

QuantityUnit or scaleWhat to know
Out-of-sample error (RMSE, log loss, AUC)Variable's unit for RMSE; dimensionless for log loss and AUCPerformance on data not used in training. The value depends on how the split was made: random splits overstate performance when neighbouring points are correlated in space or time, so spatial and temporal blocking is the fair test for risk applications.
Calibration errorExpected calibration error, dimensionless; or reliability diagram deviationWhether a predicted probability of 0.2 verifies about one time in five. Guo and co-authors (2017) show modern neural networks are often over-confident; temperature scaling and similar methods correct it on the validation distribution only.
Prediction interval coverage and widthProportion of outcomes inside the interval at a nominal level; interval width in the variable's unitConformal prediction guarantees marginal coverage under exchangeability without a model assumption; coverage can still fail for subgroups or after a shift, and narrow intervals with correct coverage are the aim.
Emulator error relative to the full modelSame unit as the model output, often as a percentage of the output rangeHow closely a surrogate reproduces a physical or catastrophe model. The error is measured on held-out model runs, not on reality; the emulator inherits the full model's errors and adds its own, which are largest at the extremes of the parameter space.
Drift statisticsPopulation stability index, Kolmogorov-Smirnov statistic or embedding distance, dimensionlessShift in inputs or predictions since training. Input drift without outcome labels cannot say whether performance has fallen; it says only that the model is being asked questions it was not trained on.
Group performance disparityDifference or ratio of error rates, false positive rates or calibration between groups, dimensionlessFairness measures. Several definitions cannot all hold at once when base rates differ, so the choice of measure is a policy decision that should be stated.

முறைகள்

Methods and their limits

Each method is listed with what it assumes and where it fails.

முறைAssumesFails when
Supervised learning (gradient boosting, random forests, neural networks)That labelled examples represent the deployment population, that the relation between features and outcome is stable, and that the loss function matches the decision cost.When labels are scarce or biased (loss data recorded only where insurance exists), when the model learns a proxy that breaks under intervention, or when performance in the tail, where risk lives, is averaged away by the bulk.
Emulation and surrogate modelling (Gaussian processes, neural surrogates)That the full model's response is smooth enough to learn from a designed set of runs, and that the emulator's own uncertainty can be estimated and carried forward (Kennedy and O'Hagan, 2001).At the edges of the design space, for discontinuous responses (levee failure, phase changes), or when the emulator is used to explore extremes the full model was never run at.
Uncertainty quantification (deep ensembles, Bayesian approximations, dropout)That disagreement among models or posterior samples reflects epistemic uncertainty and that the aleatoric part can be learned from the data.When all ensemble members share the training data's blind spots, when approximate posteriors are over-confident, or when the uncertainty estimate is itself never validated against held-out coverage.
Conformal predictionThat calibration and test data are exchangeable, and that a reasonable score function is available from any base model.Under distribution shift that breaks exchangeability, for coverage conditional on a subgroup (only marginal coverage is assured), and when the calibration set is small so that intervals are wide.
Anomaly and change detectionThat normal behaviour can be characterised from history and that departures are rare, detectable and worth investigating.When the training period includes the anomalies, when the base rate of true events is so low that the false alarm count swamps analysts, or when the detector flags change in the sensor rather than in the world.
Causal inference with machine learning (double machine learning, causal forests)That the causal structure can be stated (a directed graph), that confounders are observed, and that treatment and outcome are measured without the model's own feedback.When unobserved confounding is plausible, when the positivity assumption fails for the groups of interest, or when the question is predictive and causal machinery adds assumptions without benefit.

Debates

What the field debates

Should machine learning emulators replace physical models in risk assessment?
Proponents point to speed gains of orders of magnitude that make probabilistic assessment with many scenarios feasible, and to benchmark results in weather forecasting. Critics reply that emulators cannot be trusted beyond the training envelope, that extremes are exactly where the training data are sparse, and that an emulator's error relative to the full model is not an error relative to reality; hybrid designs that keep physics in the loop are the current compromise.
Is distribution-free uncertainty (conformal prediction) enough, or are Bayesian models needed?
The conformal side holds that finite-sample coverage guarantees with no model assumptions are what a risk analyst needs and that Bayesian posteriors depend on priors and approximations that are hard to check. The Bayesian side replies that marginal coverage is weak, that decisions need full predictive distributions and conditional uncertainty, and that conformal intervals say nothing about why the model is uncertain.
Which fairness definition should a risk model satisfy?
Calibration within groups is defended as the condition for a score to mean the same thing for everyone. Equalised error rates are defended as the condition for the model to burden groups equally. Impossibility results show both cannot hold when base rates differ, so one position argues for context-specific choice disclosed in documentation, while another argues that certain uses should not be automated at all.

Bodies and data

Bodies that set practice, and the data they publish

Bodies and standards are named so learners can find and apply their published work; the Academy teaches their use and links to the publisher, and no affiliation is implied.

BodyRole in the field
National Institute of Standards and Technology (NIST)Publishes the AI Risk Management Framework and runs evaluations and measurement programmes for machine learning systems.
ISO/IEC JTC 1/SC 42 Artificial intelligenceMaintains the international standards on AI management systems, risk management, bias and data quality for machine learning.
International Machine Learning Society and the NeurIPS and ICML communitiesPeer-reviewed venues where the methods for uncertainty quantification, calibration and conformal prediction are published and benchmarked.
Royal Statistical SocietyProfessional body for the statistical foundations of learning, calibration and evaluation, with a data science and AI section.
IEEE Geoscience and Remote Sensing SocietyThe learned society for machine learning on Earth observation data, including the benchmarks for damage, flood and land cover mapping.
DatasetSteward
OpenML benchmark datasets and tasksOpenML Foundation
Prithvi geospatial foundation modelsIBM and NASA on Hugging Face
AI Incident DatabaseResponsible AI Collaborative
SpaceNet open satellite imagery datasetsSpaceNet

04முறை

From a trained model to an interval with assured coverage: residuals, the conformal quantile and the coverage a user can count on

A learned model gives a point prediction with no promise attached; holding out a calibration set, ranking its residuals and taking a quantile wraps the prediction in an interval whose coverage is assured without any assumption about the model, and the size of the calibration set sets how tight that guarantee is.

Sources: Vovk, Gammerman and Shafer, Algorithmic Learning in a Random World (2005); Angelopoulos and Bates, Conformal prediction: a gentle introduction, Foundations and Trends in Machine Learning (2023); Guo, Pleiss, Sun and Weinberger, On calibration of modern neural networks, ICML (2017); Gama et al., A survey on concept drift adaptation, ACM Computing Surveys (2014); Rudin, Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead, Nature Machine Intelligence (2019).

05At a glance

At a glance, hub and courses

3D வரைபடம்

இடர் அறிவியல் மற்றும் முறைகள் புலம்

265 nodes1,691 relations9 views

Explore the 3D graph: Faculty of Risk Science and Methods

06கட்டுப்பாட்டகம்

How AI and machine learning for risk analysis is taught, as a process you can run

Press Play to run this process from this page's own data.

இடர் பகுப்பாய்வுக்கான AI மற்றும் இயந்திரக் கற்றல்

How AI and machine learning for risk analysis is taught

5 subjects in scope, 3 key terms, courses in development, 1 lab, in the Quantitative Risk Analysis department of Risk Science and Methods: from the scope to the method and its terms, the courses that teach it, the labs on the twin, the assessment and the record.

  1. 01 Scope

    1. 01

      Supervised learning for…

      Start · the discipline

      Supervised learning for risk outcomes

    2. 02

      Emulators and…

      Start · the discipline

      Emulators and surrogates of physical and catastrophe models

    3. 03

      Uncertainty…

      Start · the discipline

      Uncertainty quantification in machine learning

    4. 04

      Conformal prediction…

      Start · the discipline

      Conformal prediction and distribution-free coverage

    5. 05

      Anomaly detection and…

      Start · the discipline

      Anomaly detection and change detection

  2. 02 Method and terms

    1. 06

      From a trained model to…

      Model · the learner

      A learned model gives a point prediction with no promise attached; holding out a calibration set, ranking its residuals and taking a…

    2. 07

      Emulator

      Record · the discipline

      A fast statistical or machine-learning model trained to reproduce the outputs of a slower physical or catastrophe model.

    3. 09

      Conformal prediction

      Record · the discipline

      A method that wraps any model to give prediction sets with a stated coverage under exchangeability.

    4. 11

      Distribution shift

      Record · the discipline

      A change between the data a model was trained on and the data it is used on, which can invalidate its accuracy.

  3. 04 Labs

    1. 08

      இரட்டை மாதிரி

      Exercise · the learner, on the faculty twin

      The method runs on the faculty twin and the admitted map layers.

  4. 05 Assess

    1. 10

      Module checks

      Gate · the course

      Each module ends with a short check against its outcomes.

    2. 12

      Assessed work

      Decision · the learner

      The assessed work that closes a course: a lab record, a case memo or an assignment, reviewed against the rubric.

  5. 06 Record

    1. 13

      நிறைவுப் பதிவு

      Outcome · the learner

      The record stacks into the discipline achievement.

    2. 14

      Discipline achievement

      Record · the Academy

      The discipline achievement follows the discipline rule, read from the records.

AI and machine learning for risk analysis · Quantitative Risk Analysis An outline of how the work is designed to run; it does not report the status of any decision or record.

07கற்றல்

Achievements, where this is practised, practised in and across the ecosystem

Achievements

Achievements in this discipline

  1. Course achievements

    One for each course with this as its primary discipline, earned when you pass the course's assessed activities.

  2. Department: Quantitative Risk Analysis

    Earn two discipline achievements in the Department of Quantitative Risk Analysis. In development.

An achievement is a record of the learning completed and assessed here. How recognition works.

Where this is practised

Where the subject is practised

A field platform page on this subject is in development.

Practised in

The two Foundry systems where it is practised

  • FoundryNXS-LAB systemMachine-learning methods are compared under transparent tests
  • NXS-NXQ systemModels have a lifecycle of fitness, validation and drift

Foundry system pages describe what each system and its rooms are designed to do; the courses teach the methods these systems apply.

Across the ecosystem

Across the ecosystem

The science, platforms and standards this discipline draws on and contributes to across the ecosystem.

Foundations in the STEM Commons

Pages of the ecosystem, cited so learners can follow the field into research, platforms, standards and national practice.

08பங்கேற்கவும்

Take part in Risk Science and Methods

நீங்கள் யார் என்பதிலிருந்து தொடங்குங்கள்.

Routes

  • வழித்தடம்

    தேசிய நிறுவனங்கள்

    Ministries, statistics offices and agencies: a national desk and chapter, and a seat in the working groups. National chapters

  • வழித்தடம்

    Sectors and professions

    Operators, regulators, banks, insurers and professional bodies whose work touches risk science and methods. For organisations

  • வழித்தடம்

    கற்றல்

    Courses and practice, with a record of every course completed. Free to study. Open learning

  • வழித்தடம்

    தரநிலைகள்

    Draft, review and adopt the methods and data standards with the working groups. Working groups

  • வழித்தடம்

    Communities and research

    Universities, civil society and individual experts: projects, data and review. Write to the Academy

Open calls

Open callSeeksCloses
Founding call for experts to the working groups CALL-FOUNDING-EXPERTSexpert, editor, chair2026-12-15
Open nomination for the three independent expert seats of the Global Standards Council CALL-GSC-INDமன்ற உறுப்பினர்2026-12-01
Nominations of interim regional members (applies only if the transitional provision is decided) CALL-GSC-REGமன்ற உறுப்பினர்2026-12-15
Review of the wave-1 risk core outlines (C1, C2, C3, C4, C8, C10, C11) CALL-REVIEW-WAVE1contributor2027-01-31
இடர் அறிவியலின் முதன்மைத் தரநிலைகளுக்கான நிபுணர்கள்: நிச்சயமின்மை மற்றும் சான்றுத் தரப்படுத்தல், அளவுகோல்கள், கட்டமைக்கப்பட்ட நிபுணர் தீர்ப்பு, சகிப்புத்தன்மை, உச்சநிலைகள் மற்றும் கூட்டு நிகழ்வுகள், பேரழிவு மற்றும் இருத்தலியல் இடர், நெறிமுறைகள் மற்றும் சமத்துவம், அலகுகள் மற்றும் அளவீடு (NRS-C17 முதல் NRS-C24 வரை) CALL-CORE-SCIENCEexpert, editor, chair2027-01-31
Independent reviewers for the families of the hazard atlas, three from at least two regions for each family (NHS2-R19), priority families first CALL-ATLAS-REVIEWநிபுணர்2027-03-31
இடைக்கால முன்னோடிப் பாதுகாப்பு மறுஆய்வுக் குழுவின் உறுப்பினர்கள் மற்றும் சிவப்புக் குழு மறுஆய்வாளர்கள் (உயிரியல் பாதுகாப்பு, AI பாதுகாப்பு, முக்கிய உள்கட்டமைப்புப் பாதுகாப்பு); விண்ணப்பங்கள் இரகசியமாகக் கையாளப்படும் CALL-FRONTIER-SAFETY-REVIEWநிபுணர்2026-11-30
தேவைப் பதிவேடு, NHS-2 இயந்திரப் பிரதிநிதித்துவம், இணக்கச் சோதனைத் தொகுப்புகள் மற்றும் இணக்க மதிப்பீட்டுத் திட்டத்திற்கான கருத்தியல், அளவியல், மென்பொருள் சரிபார்ப்பு மற்றும் இணக்க மதிப்பீட்டு நிபுணர்கள் CALL-CONFORMANCE-ONTOLOGYexpert, editor2027-01-31

Calls past their closing date are hidden automatically. Every call, proposal and route: open areas and calls.

Cells and CERTs

  1. Gather the cell

    Bring together three to seven people with complementary skills: data, field practice, policy and communication.

  2. Register it

    Write to the team with the cell's name, members, area and region. You receive a reference and a contact. Write to the Academy

  3. Choose a charter

    Pick one focus: research and evidence, data and maps, standards review, or community preparedness.

  4. Learn together

    Follow shared study paths at Risk Academy, with practice and a record of every course completed. Hazards and resilience

  5. Lead and report

    Hold regular sessions. Publish what the cell learns through Reports. Bring findings to the working groups. Working groups

  • Specialise as a Community Emergency Response Team (CERT)

    A CERT prepares its neighbourhood for the hazards it faces. It builds a local hazard profile from this site's map. It trains in household and street preparedness. It learns first aid and light search through recognised local providers, and it runs drills. A CERT works alongside the local emergency services and follows their direction. It never replaces them and is not an emergency service.

குறியீடு

What a national team or a local community can do on The Global Risks Index (NXS-GRX) today.

  • NXS-GRX

    Your country's profile

    Open risk profiles for 195 countries, with sources and licences on every value. Open the Index

  • NXS-GRX

    Correct or reply

    Correct a value or reply for your country. The change is logged with its effect. Open the Index

  • NXS-GRX

    Offer data

    Offer a dataset. It enters through the licence gate, with your credit. Open the Index

  • NXS-GRX

    Use the open data

    Layers, profiles and tiles through the open API. Open the Index

09References

Core references and where to begin

Start with Hastie, then Guo, then Angelopoulos.

  1. Hastie, T., Tibshirani, R. and Friedman, J., The Elements of Statistical Learning, second edition, Springer (2009). DOI
  2. Kennedy, M. C. and O'Hagan, A., Bayesian calibration of computer models, Journal of the Royal Statistical Society Series B (2001). DOI
  3. Vovk, V., Gammerman, A. and Shafer, G., Algorithmic Learning in a Random World, Springer (2005). DOI
  4. Angelopoulos, A. N. and Bates, S., Conformal prediction: a gentle introduction, Foundations and Trends in Machine Learning (2023). DOI
  5. Guo, C., Pleiss, G., Sun, Y. and Weinberger, K. Q., On calibration of modern neural networks, Proceedings of the 34th International Conference on Machine Learning (2017). proceedings.mlr.press
  6. Lakshminarayanan, B., Pritzel, A. and Blundell, C., Simple and scalable predictive uncertainty estimation using deep ensembles, Advances in Neural Information Processing Systems 30 (2017). arxiv.org
  7. Rudin, C., Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead, Nature Machine Intelligence (2019). DOI
  8. Reichstein, M. et al., Deep learning and process understanding for data-driven Earth system science, Nature (2019). DOI
  9. Pearl, J., Causality: Models, Reasoning, and Inference, second edition, Cambridge University Press (2009). DOI
  10. Barocas, S., Hardt, M. and Narayanan, A., Fairness and Machine Learning: Limitations and Opportunities, MIT Press (2023). fairmlbook.org

10கேள்விகள்

கேள்விகள், பதில்கள் மற்றும் முக்கியச் சொற்கள்

இங்கே படிப்பு இலவசமா?

ஆம். Nexus Consortium வழிநடத்தும் Risk Academy-இன் ஒவ்வொரு படிப்பையும் கற்றல் பொருளையும் இலவசமாகப் படிக்கலாம், சுய வேகத்திலோ ஒரு குழுவுடனோ. சாதாரணப் பங்கேற்புக்குக் கட்டணம் இல்லை; ஊதியம் பெறும் பணி, பயிற்சிப் பணிகள் மற்றும் ஆய்வு உதவித்தொகைகள் தனி ஏற்பாடுகள்.

ஒரு படிப்பை முடிப்பது எனக்கு என்ன தருகிறது?

A completion record: the course, its version, the date and what was assessed, private by default and shared by you. It stacks into discipline, department and programme achievements. A record states what was studied and assessed and does not certify competence; it is not a licence, a professional title or an employment status.

தரவும் இடங்களும் எங்கிருந்து வருகின்றன?

Every faculty teaches on a twin of a real place built from open data; each layer carries its provider, licence and date, and every result on the Academy states its source. Maps and figures are indicators for teaching, never an assessment of a place.

How do I take part beyond the courses in aI and machine learning for risk analysis?

Form a competence cell of three to seven people, study a shared pathway, take a supervised assignment in the National Working Groups, or bring a cohort from your institution. For organisations.

முக்கியச் சொற்கள்

Emulator
A fast statistical or machine-learning model trained to reproduce the outputs of a slower physical or catastrophe model.
Conformal prediction
A method that wraps any model to give prediction sets with a stated coverage under exchangeability.
Distribution shift
A change between the data a model was trained on and the data it is used on, which can invalidate its accuracy.
Calibrated classifier
A classifier whose predicted probabilities match observed frequencies.

அறிக்கைகள்பங்களிக்கவும்

இந்தப் பக்கத்திலிருந்து சான்றுகளைப் பங்களிக்கவும்.

Reports is the Consortium's quarterly publication. What you know about this subject can become evidence, a case or a correction in the next edition, credited to you and published under CC BY 4.0 where it is your own work.

Risk Academy

Never send personal data, sensitive infrastructure detail or anything you have no right to share. A contribution is reviewed before use; sending it confers no authorship or editorial rights, and sponsorship never buys a finding.

இடர் பகுப்பாய்வுக்கான AI மற்றும் இயந்திரக் கற்றல்

Prepare with a related course

Open courses in connected disciplines teach the methods this discipline will build on.

படிப்புப் பட்டியலை உலாவுங்கள்Back to Department of Quantitative Risk Analysis