ThinkHazard! v3.0 scores review

[matamadio]

THINKHAZARD UPDATE TO v3.0

The new version of ThinkHazard is soon to be released!

Several things have been changed:

• Updated global ADM 2 boundaries
• Added "Urban" boundaries for about 3000 major cities
• Updated Hazard datasets (see docs)
• Scoring logic improved to account for both value and area threshold (see docs)
  • For an overview of each score computation, see TH scores update: OVERVIEW #959
• Updated documentation and API tool.
• Updated page content: description, recommendations, contacts

The preliminary scores and website changes can be inspected at:

https://int.thinkhazard.org/

The ADM boundaries including scores can be downlodad from WB OneDrive.

---

REVIEW PROCESS

🔎 Review Process for Preliminary Results & New Functionalities

We are sharing preliminary results and new functionalities and inviting collaborators to review and provide feedback.

---

🎯 What we are looking for

We welcome feedback on:

1. Hazard scores

• Do results look plausible in your region of expertise?
• Any clear inconsistencies or surprising outputs?
• Comparison with known events or datasets

2. Methodological choices

• Thresholds and classification logic
• Use of input datasets
• Any assumptions that should be revisited

3. New functionalities

• Usefulness for decision-making
• Clarity of outputs
• Missing features or improvements

---

🧪 How to review

Please try to:

• Focus on specific countries / regions / hazards
• Provide concrete examples (e.g. “Flood risk in X seems underestimated”)
• When possible, include:
  • References (papers, datasets, reports)
  • Screenshots or comparisons
  • Suggested alternatives

---

💬 How to contribute

You can contribute by:

• Commenting directly in this discussion
• Replying to existing comments to build on feedback
• Opening a separate issue if:
  • You identify a clear bug
  • You propose a substantial methodological change

---

🏷️ Suggested comment structure (optional)

To keep things structured, you can use:

Location:
Hazard:
Type of feedback: (data / methodology / functionality)

Observation:
Why it matters:
Suggested improvement (if any):

---

⏱️ Timeline

We’ll keep this review open until: End of October 2026

After that, we will:

• Consolidate feedback
• Prioritize and deploy changes
• Move into production

---

⚠️ Important note

These results are preliminary and subject to change based on your feedback.

[matamadio]

Summary of comments from review kick-off (25/09/2026)

1. Reviewers appreciate to get a comprehensive review of scores and thresholds against empirical evidences > produce full technical note
   1.1 Provide threshold analysis and justification, and comparison against previous scores
   1.2 Justify the 50 cm flood threshold

To be discussed, most current thresholds are the same of, or based on, the legacy approach.

2. Water scarcity is an odd mix with the other hazards as it depends very much on demand and isn't temporary - but a baseline condition. Using it as an indicator for drought, which is a temporary hazard, can be problematic.

Partially addressed in the Water scarcity issue, which compares results from SPEI and PDSI. To be re-discussed with drought experts and to evaluate alternatives (CDI?).

3. Earthquake: does the new version factors GEM models?

No, since the latest GEM only releases one return period for free and we need at least 3. We tried to request more to GEM, but was rejected - not available for free. Earthquake scores are compared to old ones.

4. Scores per hazard: have you been considering the key thresholds used by the insurance industry (as some of the commercial tools out there do?)

To be discussed. In my experience, industry often uses very different (much higher) RP as reference, e.g. 500 or 700 years.

5. Risk tiering: for the public factor material, I'd suggest explaining in more layperson / practical terms what these means in terms of probability of occurrence and likelihood of impact.

6. ThinkHazard would benefit from being much clearer on what it is presenting – this is using observational data reflecting ‘current’ expected hazards. Not past, not future.

7. Overall, use of global layers to make statements on highly localized hazards for small aggregations, ADM2, city-scale… is very risky and can easily be misinterpreted by users. What steps are being put in place to make more responsible use and understanding of what is presented?

These 3 comments are about the message vehiculated by the score and the uncertainity behind it. To be discussed.

8. Use of EmDat records: EmDat data comes from country reporting, with no scientific review of reported events. Need to be careful when using as a baseline as accuracy has been found to be questionable for reported events. Additionally, there are known issues for reporting for flooding, for example, between meteorological vs hydrologic flood.

EMDAT is far from perfect but remains the most complete reference of historical hazard events. Is there anything better we can use? Happy to integrate alternative sources!

9. Using RP1000 from FATHOM has significant uncertainties – how is this included/ reflected in data presentation for user interpretation. Otherwise this is irresponsible and can lead users to significant misinterpretation. RP10,000 for STORM TC does not make sense considering the production of the dataset.

These extreme layers are only used as boundary conditions to identify the domain of the hazard, i.e. if an area might be affected at all by those kind of hazards.

10. WBGT – is considered as a theoretical understanding of heat risk – given highly, highly localized instrumentation measurements and input data known quality (radiation, surface wind, etc). As such this theoretical metric is not considered appropriate for localized risk metric for planning or applicable risk assessments.

We can acknowledge this in our documentation; happy to include any better global data that might have become available in the meanwhile.

11. Using area thresholds for classification (sometimes very low %), means admin-2 can receive a high score even if no population lives close to the location where hazard thresholds are exceeded. This high score is then applied to all admin levels above due to the max aggregation rule. Even though you are clear that the platform doesn’t consider population exposure, I think it is easy for users to misinterpret this downstream. For example, counting how many people live in admin-1 with low and high hazard classifications could give a misleading exposure metric, especially for the localized hazards like flooding.

Thinkhazard does not include exposure or vulnerability by design. It is not a risk assessment tool, we can discuss how to avoid misinterpretation by users.

[matamadio]

Review issues

• Extreme heat category – Las Vegas #1068
• Coastal megacities #1070