What the system flagged on five historical episodes

The five episodes

Every number on this page comes from the published scorecard; the underlying per-run reports are research records and are not distributed. The tables are a projection of those published results, not a re-analysis.

• Cyclone Amphan: Bangladesh landfall, 20 May 2020: onset 2020-05-20, 14 districts named as affected.
• Cyclone Yaas: Bangladesh coast, 26 May 2021: onset 2021-05-26, 9 districts named as affected.
• Cyclone Mocha: Cox's Bazar coast and Teknaf, 14 May 2023: onset 2023-05-14, 4 districts named as affected.
• Eastern flash floods: Feni, Cumilla, Noakhali, 20–30 August 2024: onset 2024-08-24, 13 districts named as affected.
• Northeast and coastal monsoon floods: Sylhet, Sunamganj and the hill districts, 1 June 2025: onset 2025-06-01, 10 districts named as affected.

Read this first

• These are detection counts against recorded historical events, not a forecast-skill estimate.
• Five episodes is far too small to support a skill claim; the counts are honest but inconclusive.
• The classification model was not re-run for this evaluation; only published detection outputs were scored.
• No calibration accuracy is claimed.

These are five episodes, not a validation set. Detection is counted only over the districts the published assessments name: a district nobody named is unknown, not clear, so every number here is a ceiling on detection, not forecast skill.

Detection: did the system flag the districts the assessments name?

Each row is one episode, scored against that episode's own published assessment. "Flagged the class" counts only the class that occurred, which is the strict reading.

Totals: 50 of 50 named district-episode pairs carried a scored row and 13 were flagged under the class that occurred.

Per horizon: .

Scores: POD, FAR and CSI, and the rows where they do not exist

These are the standard verification scores, computed over the district-horizon samples in each episode's window. POD is "—" for amphan-2020, yaas-2021, mocha-2023, northeast-flood-2025: the published assessment names affected districts but records no dated outcome inside the prediction window, so there is no observed event to divide by. A false alarm ratio of 1.000 in that situation means "no negative sample existed", not "every alarm was wrong": with no named event there is nothing for an alarm to be right about.

The false alarm ratio is measurable only against districts where an event was recorded as absent. No district is treated as a confirmed negative, so treat the FAR column as a bound on the fraction of alarms that hit a district nobody reported as affected.

The alarm band

Threshold-sensitivity sweeps and driver diagnostics are research-private; only the headline detection counts above are published.

Alarm thresholds in force per episode: Cyclone Amphan: Bangladesh landfall, 20 May 2020 at 0.500 on the class severity score; Cyclone Yaas: Bangladesh coast, 26 May 2021 at 0.500 on the class severity score; Cyclone Mocha: Cox's Bazar coast and Teknaf, 14 May 2023 at 0.500 on the class severity score; Eastern flash floods: Feni, Cumilla, Noakhali, 20–30 August 2024 at 0.500 on the class severity score; Northeast and coastal monsoon floods: Sylhet, Sunamganj and the hill districts, 1 June 2025 at 0.500 on the class severity score.

Not published here

• Model code
• Dataset collection procedures
• Training and benchmarking details
• Severity derivation

Questions and answers