AutoGen Casualty ILS: Multi-Agent Tail Risk Structuring System
Structuring casualty insurance-linked securities requires collaborative analysis of long-tail risk distributions and trigger conditions across counterparty agents. Build an AutoGen multi-agent conversational framework that models loss distributions and achieves accurate secondary market yield predictions within 5% error.
What you are building
Build an AutoGen multi-agent system comprising Risk Modeler, Structurer, and Investor agents to design casualty ILS instruments.
How work is evaluated
Evaluates multi-agent conversation termination and output yield pricing error against ground truth actuary benchmarks.
Datasets
Public datasets, plus the private datasets you have access to.
How submissions are scored
market_clearing_pass
Validates that agents reached consensus on bond structure.
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
yield_prediction_error_pct
Percentage error between calculated yield and ground truth benchmark. • target: 0.04 • range: 0-1
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
Learning goals
Set up an AutoGen multi-agent group chat environment.
Assign specialist roles (ActuaryAgent, StructuringAgent, RatingAgencyAgent).
Simulate negotiation of coupon yield based on long-tail casualty loss distributions.
Produce final bond terms with target yield prediction error under 5%.
Resources
Dataset of 40 historical and synthetic casualty ILS issuance parameters and pricing outcomes.
Challenge setup
Evaluates multi-agent conversation termination and output yield pricing error against ground truth actuary benchmarks.
Challenge input
JSON containing expected_loss_pct, attachment_point, tail_factor
AutoGen
Microsoft framework for orchestrating conversational agents.
Evaluated output
JSON containing agreed_coupon_bps, rating_grade, market_clearing_status
- Validates that agents reached consensus on bond structure.
- Percentage error between calculated yield and ground truth benchmark. • target: 0.04 • range: 0-1
- Benchmark: Multi-Agent Consensus Benchmark
- Yield Prediction Error Pct target: 0.04
- 1 public reference case
- Python execution harness
- Python sandbox (unavailable on Versalist)
Shows the first task, up to three tools, and up to three evidence items.
View technical recipe
Configured tools
- AutoGen · Required
- Novita AI · Optional
- Lingo · Optional
Evaluation contract
- market_clearing_pass · Weight 1
- yield_prediction_error_pct · Weight 1
Metrics
- yield_prediction_error_pct · Target 0.04
Benchmarks
- Multi-Agent Consensus Benchmark
- general
Reference cases
- 1 public reference case
Workflow tasks
- Task 1
Input: JSON containing expected_loss_pct, attachment_point, tail_factor
Output: JSON containing agreed_coupon_bps, rating_grade, market_clearing_status
Execution harness: Python
Evaluation tools
- AutoGen · Multi-Agent Framework
Recipe freeze marker
No recipe freeze time is recorded.