Back to Advanced Analysis
Pang & ChoiSynthetic fallback datasetPang & Choi (2022), arXiv

Bayesian Trigger Model

MC Dropout / Bayesian neural network

A payout model that returns both a mean estimate and an uncertainty band.

Paper focus

Uncertainty-Aware Trigger Calibration for Parametric Insurance

This model shows the spread around the payout estimate instead of only showing one number.

Peril handling

It uses the same weather inputs as the MLP page but adds uncertainty to the final output.

Basis-risk role

It helps when a decision near attachment should show whether the model is confident or not.

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 current Bayesian page trains on synthetic weather-loss rows shaped like the backend tabular pipeline.

  • Synthetic rows match the event-level tabular payout workflow used by the backend.
  • The model reports both mean payout and uncertainty from repeated forward passes.
  • This is a practical approximation of the paper, not a dataset match.
Local implementation
  • Reuses the tabular payout pipeline and keeps dropout active during inference to produce multiple stochastic passes.
  • Aggregates mean and spread into one decision output so payout level and uncertainty are visible together.
  • Serves a live case-analysis response from the latest trained checkpoint rather than a hard-coded chart.
Model flow
Step 1
Weather features
Step 2
Dropout MLP
Step 3
Stochastic passes
Step 4
Mean ± variance
Step 5
Review posture

Mathematical model

  • Monte Carlo dropout approximates the Bayesian posterior by sampling dropout masks during inference.
  • Mean μ(x)=1/S Σ_{s=1}^S ŷ_s(x) and predictive variance σ²(x) are computed from the same stochastic passes.
  • The policy widens/narrows the trigger by inspecting μ ± k·σ so uncertain cases trigger review.
Equation

μ(x) = 1/S Σ_{s=1}^S ŷ_s(x), σ²(x) = 1/S Σ_{s=1}^S (ŷ_s(x) - μ(x))²

μ(x) = 1/S Σ_{s=1}^S ŷ_s(x), σ²(x) = 1/S Σ_{s=1}^S (ŷ_s(x) - μ(x))²
Architecture presentation
Same 72-feature loader as the Chen run but inserts dropout between each dense block.
Training turns dropout off, but inference runs S stochastic forward passes with dropout on.
Aggregator summarizes mean and variance to decide whether the payout should trigger or be reviewed.

Pros

  • Adds uncertainty bands to the tabular payout pipeline.
  • Highlights borderline events so reviewers can request additional evidence.
  • Uses MC dropout so it is easy to integrate with existing network definitions.

Cons

  • MC dropout is only an approximation to the paper’s deep sigma-point process.
  • Inference cost grows linearly with the number of stochastic passes.
  • Downstream policy editors must agree on what σ means before adjusting thresholds.

Results

Attachment

—

Decision boundary applied to the model output.

Payout MAE

—

Average absolute payout error on the held-out split.

Payout RMSE

—

Root-mean-square payout error on the held-out split.

Basis Risk

—

Threshold mismatch rate after calibration.

Test Uncertainty

—

Average predictive spread on the test split.

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

Hong Kong Cold-Chain Warehouse — Typhoon Saola

Running model inference for the selected case.

Architecture

  • Structured weather features into dense layers with dropout retained at inference
  • Monte Carlo samples aggregated into mean payout and uncertainty band
  • Decision policy uses the uncertainty band when the score is near attachment

Inputs

  • Same structured event features as the MLP baseline
  • Dropout rate and number of stochastic forward passes
  • Optional case-analysis settings such as sample count

Outputs

  • Mean payout estimate
  • Uncertainty interval and standard deviation
  • A recommendation for review when the interval overlaps attachment