Back to Advanced Analysis
Chen 2024Paper replication packageChen et al. (2024), Management Science

MLP Trigger Model

Feed-forward neural network

A feed-forward model that maps weather inputs to a payout fraction.

Paper focus

Expected-Utility Neural Trigger Design for Index Insurance

This model uses many weather inputs at once instead of relying on one trigger variable.

Peril handling

Primary and secondary effects enter together through the weather feature vector.

Basis-risk role

It works as a benchmark when you want to compare a learned payout curve with simpler trigger rules.

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

This page trains from the local replication package for the Chen paper.

  • Uses the 72 monthly weather features and loss labels contained in the replication files.
  • Training split follows the paper package time windows instead of a random shuffle.
  • The page reads a saved checkpoint and shows live results.
Local implementation
  • Loads the Chen replication package directly from the backend and reconstructs the 72 weather inputs used for training.
  • Applies the paper-style temporal split so training, validation, and test windows follow the original replication package logic.
  • Learns payout fraction on the local machine, saves a versioned checkpoint, and serves case analysis from the trained artifact.
Model flow
Step 1
Replication package loader
Step 2
72 weather features + optional regime
Step 3
64-64-16 MLP w/ ReLU
Step 4
Sigmoid payout fraction
Step 5
Calibrated threshold θ*

Mathematical model

  • Output ŷ(x) = σ(W₃·ReLU(W₂·ReLU(W₁·x + b₁) + b₂) + b₃) keeps payouts in the [0,1] range.
  • The loss minimizes mean-squared error between ŷ and the realized payout so the network mirrors the expected-utility objective.
  • Attachment decisions sweep a calibrated threshold θ* over ŷ instead of using a single weather index.
Equation

ŷ(x) = σ(W₃·ReLU(W₂·ReLU(W₁·x + b₁) + b₂) + b₃)

ŷ(x) = σ(W₃·ReLU(W₂·ReLU(W₁·x + b₁) + b₂) + b₃)
Architecture presentation
Replication package loader brings 72 monthly ERA5 weather variables plus optional regime tags.
Dense stack (64 → 64 → 16) with ReLU before a sigmoid payout head keeps the output bounded.
Calibration stage derives θ* from the saved checkpoint so the payout curve matches the paper split.

Pros

  • Faithfully reproduces Chen’s dataset split while staying usable in ParaEval.
  • Live payout curve output lets you compare decisions against guardrail rules.
  • Supports HDP regime tags so the representation can capture compound perils.

Cons

  • Dense stack is harder to interpret than a single-index trigger.
  • Requires strict normalization of the 72 variables each time you retrain.
  • Still needs calibrated thresholds to turn continuous payouts into payouts.

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.

Boundary F1

—

Trigger quality after applying the attachment 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

Hong Kong Cold-Chain Warehouse — Typhoon Saola

Running model inference for the selected case.

Architecture

  • 72 weather features into an MLP with 64-64-16 hidden layers
  • ReLU activations in hidden layers and sigmoid payout head
  • Training target is payout fraction
  • Optional regime labels can be appended as extra inputs

Inputs

  • Event-level weather feature tables from ERA5 or similar sources
  • Historical loss or payout observations
  • Optional latent regime labels from the CRP/HDP pipeline

Outputs

  • Predicted payout fraction between 0 and 1
  • A trigger recommendation
  • Feature preview notes for the selected case