Back to Advanced Analysis
Cesarini 2021Planned synthetic datasetCesarini et al. (2021), NHESS

Cesarini Weather Classifier

Rare-event weather classifier

A paper-correct track for multi-source flood and drought event classification using precipitation, soil moisture, and SPI-derived features.

Paper focus

The Potential of Machine Learning for Weather Index Insurance

This track models extreme-event detection directly, which is the paper’s actual problem setting, instead of forcing the task into payout-fraction regression.

Peril handling

The paper combines multiple precipitation products, ERA5 soil-moisture layers, and drought indices so the model can detect compound hydro-meteorological conditions.

Basis-risk role

Its role in ParaEval is upstream trigger screening: a well-calibrated rare-event classifier can reduce false alarms and missed events before payout logic is applied.

Trained run
Loading trained checkpoint metadata.
How To Read This Page
Start with Results to see the held-out performance that matters for this model family.
Use Live case inference to understand what the model would recommend for one insured event.
Read Training data and Local implementation for realism and replication caveats.
Training data

The first executable backend path will use a paper-shaped synthetic dataset for flood and drought event labels.

  • Synthetic rows will preserve the classification task and class-imbalance profile.
  • This will be used for training smoke tests and API wiring before real ingestion lands.
  • Results shown now are structured placeholders, not trained paper metrics.
Local implementation
  • A dedicated Cesarini backend trainer now saves versioned classifier artifacts using the shared model registry.
  • The adapter serves real calibration and case-level inference from those saved artifacts.
  • The current implementation is paper-inspired on feature engineering, while keeping the paper-correct classification framing.
Model flow
Step 1
Weather features
Step 2
Binary classifier
Step 3
Threshold sweep
Step 4
Trigger recommendation
Step 5
ParaEval decision

Mathematical model

  • Input x is a tabular vector built from multiple rainfall products, soil moisture, and transformed drought indicators such as SPI.
  • The paper studies neural classifiers for binary event detection, with best-performing configurations concentrated around deep ReLU networks and imbalance-aware sampling.
  • The deployed ParaEval path will expose threshold sweeps over predicted event probability so basis-risk tradeoffs remain explicit.
Equation

p(event | x) = σ(f_θ(x)), trigger = 1[p(event | x) ≥ θ*]

p(event | x) = σ(f_θ(x)), trigger = 1[p(event | x) ≥ θ*]
Architecture presentation
Weather-source fusion produces one event-level feature vector spanning rainfall, soil moisture, and drought indicators.
A deep classifier estimates event probability rather than payout fraction.
Calibration selects a trigger threshold that exposes the precision-recall basis-risk tradeoff clearly.

Pros

  • Matches the paper’s classification framing instead of distorting it into direct payout prediction.
  • Fits naturally with ParaEval’s trigger-calibration workflow.
  • Provides a clean insertion point for CRP posterior features.

Cons

  • Not a direct payout model, so a downstream trigger-to-payout mapping is still required.
  • Paper-faithful feature construction is more demanding than the current demo tabular schema.
  • Synthetic fallback data cannot fully reproduce the real environmental dependencies.

Results

Trigger Threshold

—

Event-probability boundary used for the trigger recommendation.

Precision

—

Share of predicted trigger events that are actually positive.

Recall

—

Share of true trigger events that the model successfully finds.

Basis Risk

—

Missed or false trigger rate after calibration.

Boundary F1

—

Balanced trigger quality at the deployed threshold.

Trigger Boundary View

This ruler shows where the current case sits relative to the deployed trigger threshold.

0%Threshold —100%
Case score
—
Decision boundary
—
Example cases

Live case inference

Jakarta Residential Asset — January 2020 Floods

Running model inference for the selected case.

Architecture

  • Multi-source weather feature builder
  • Binary event classifier with imbalance-aware training
  • Threshold calibration over held-out data
  • Optional CRP posterior features appended upstream

Inputs

  • Rainfall products and transformed hydrological indicators
  • Observed event labels for flood or drought occurrence
  • Optional regime labels or posterior probabilities from CRP/HDP

Outputs

  • Extreme-event probability
  • Calibrated trigger threshold
  • Case-level recommendation tied to the event score