Risk-aware optimization

Purpose

The estate optimizer can evaluate risk-aggregated targets against a scenario ensemble through a fast account-overlay path. One deterministic account frame is evaluated under per-scenario account multipliers, the frames are aggregated with the existing risk layer, and the weighted-goal objective runs on the aggregated pseudo-accounts.

This is not a process-driven stochastic landscape simulation. The full process engine remains the auditable model; the overlay is a planning bridge until multi-fidelity ensemble evaluation exists.

Scenario file

The optimizer reads the existing ws4.scenario.v1 file. The overlay recognises account_multipliers:

schema_version: ws4.scenario.v1
scenarios:
  - scenario_id: low
    seed: 1
    overrides:
      account_multipliers:
        feature.Volume: 0.8
  - scenario_id: high
    seed: 2
    overrides:
      account_multipliers:
        feature.Volume: 1.2

Risk targets

Targets must use a non-single aggregation, such as mean, cvar_upper, quantile, or probability. The risk layer creates a risk.<aggregation>.<index>.<account> pseudo-account and the optimizer prices that target.

Flow targets cannot be risk-aggregated.

CLI

ws4 optimize ... --scenario-file scenarios.yaml

The optimizer loads the scenario set once, bypasses the per-move delta evaluator, and records the last risk evaluation for inspection.

Limits

  • Scenarios differ only by account multipliers; the estate simulation is shared.

  • Scenario seeds do not drive process draws in this path.

  • The spatial penalty is applied after the risk objective.

  • Full process-driven stochastic search and multi-fidelity evaluation are later work.

Multi-fidelity verification

When scenarios are active, the optimizer writes risk_verification.json comparing the best search objective with the final full-simulation risk objective. A non-zero gap is visible, not silently corrected.

--risk-validation-fraction splits the scenario ensemble into ordered training and validation subsets and writes per-scenario and aggregated risk files plus a risk_validation.json summary. The generalization gap is the validation objective minus the training objective. A one-scenario ensemble writes a skipped report with the reason.

The split is ordered, not random. Keep the scenario order intentional.