Back to Advanced Analysis
SMARTSynthetic fallback datasetSMART Project (2021), GFDRR / World Bank

SMART Hybrid Model

CNN-LSTM multi-hazard sequence model

A sequence model for event windows with multiple weather channels.

Paper focus

SMART Multi-Hazard Time-Series Classification

This page treats an event as a short sequence instead of one row.

Peril handling

It is meant for cases where conditions build over time, such as rainfall followed by flood depth.

Basis-risk role

It is more useful than a tabular model when the order of conditions matters.

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 SMART page trains on synthetic multi-channel event windows built to the required tensor shape.

  • Uses fixed-length sequences so the backend can train, save, and serve the model end to end.
  • Captures the paper’s sequence structure even though the original dataset is not present in the repo.
  • The page still shows live metrics and inference from a saved checkpoint.
Local implementation
  • Builds a fixed-length hazard window for each event and feeds the sequence through the local LSTM training pipeline.
  • Uses the backend sequence artifact format so the model can be retrained and served like the tabular tracks.
  • Returns a live impact score for a selected case using the latest saved sequence checkpoint.
Model flow
Step 1
Hazard window
Step 2
Channel normalization
Step 3
Conv1d plus LSTM
Step 4
Impact score
Step 5
Trigger decision

Mathematical model

  • Inputs are tensor windows X∈R^{T×C} (e.g., T=10 time steps, C=5 weather channels).
  • Conv1d layers extract local temporal motifs before an LSTM encoder accumulates memory across the entire window.
  • Output y = σ(W h_T + b) becomes the cumulative peril impact score used for payoff decisions.
Equation

h_t = LSTM(Conv1d(x_t)), y = σ(W h_T + b)

h_t = LSTM(Conv1d(x_t)), y = σ(W h_T + b)
Architecture presentation
Fixed-length horizon (T=10) of multi-channel measurements feeds the model.
1D convolution per channel highlights short-term events before the LSTM reads the full history.
Sigmoid head outputs a hazard index that can be tuned via per-case thresholds.

Pros

  • Sequence-aware model reflects how rainfall builds into floods and droughts.
  • Combines local convolution with LSTM memory so it can detect secondary perils.
  • Fits into ParaEval’s training-and-serving pipeline via the sequence adapter.

Cons

  • Requires maintaining a sequence dataset instead of event-level rows.
  • More hyperparameters (window length, LSTM depth) makes tuning harder.
  • Still synthetic when the SMART dataset is not yet available.

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.

Performance Snapshot

End-user summary of how reliable the current model looks on held-out examples.

Trigger precision
—

When the model says payable, this estimates how often that call is right.

Trigger recall
—

When a payable event really happens, this estimates how often the model catches it.

Basis-risk exposure
—

Higher is better here. It reflects fewer false triggers and fewer missed payable events.

Payout fit
—

Higher is better here. It compresses payout error into a simple quality indicator for the UI.

Example cases

Live case inference

Jakarta Residential Asset — January 2020 Floods

Running model inference for the selected case.

Architecture

  • Input tensor shaped as sequence length by weather channels
  • Optional Conv1d front end for local temporal motifs
  • Two-layer LSTM encoder feeding a sigmoid impact head
  • Case-analysis output is returned as a payout-style score and recommendation

Inputs

  • Fixed-length event windows from ERA5 or other time-series sources
  • Per-channel normalization and channel metadata
  • Case-specific peril family for sensible defaults

Outputs

  • Cumulative peril impact score
  • Notes about the sequence window used for the case
  • A trigger recommendation