Skip to content

Risk aversion (CVaR)

Risk aversion makes the model weight high-cost scenarios more heavily instead of looking only at the expected cost. Novomodelo uses a combination of the expected value and the CVaR (Conditional Value at Risk), which is the expected cost over the worst fraction of scenarios.

The average can hide bad scenarios

Risk-neutral SDDP minimises only the expected cost, which can lead to policies that perform poorly in adverse scenarios. The risk-averse version protects against those tail risks and preserves the convexity properties needed to generate valid cuts.

CVaR is the average of the tail

CVaR with tail fraction 1 is the expected value, that is, risk-neutral. With fraction 0.5 it is the average of the worst 50% of outcomes; with 0.2, the worst 20%; with 0.05, the worst 5%. The smaller the fraction, the more risk-averse the stance.

Two parameters, per stage

The measure combines the expected value and the CVaR: one parameter is the tail fraction and the other is the weight on the CVaR. With weight 0 the measure is the expected value; with weight 1 it is pure CVaR. Each stage can have its own.

Figure: cost distribution, with the mean, the value at risk (VaR) and the CVaR marked and the tail of the worst outcomes shaded. The CVaR is the average cost inside the tail.
  • Weight of each opening in the cut aggregation
  • Risk-adjusted lower bound of the first stage $

Under risk aversion, cut aggregation stops using uniform probabilities and reweights the openings towards the worst outcomes. The model computes these weights; you provide only the measure of each stage.

  • The risk measure of each stage, in the risk_measure field of stages.json. The value "expectation", assumed when the field is absent, is risk-neutral. Any other string is rejected.
  • For CVaR, the two parameters, both required: the tail fraction (alpha, greater than 0 and up to 1) and the weight on the CVaR (lambda, from 0 to 1).
  • A zero weight is equivalent to the expected value, whatever the tail fraction.
  • The measure can vary from one stage to another.

A stage’s risk_measure weights the next stage’s openings in the cuts that stage stores. The first stage’s also weights the first stage’s own openings in the lower bound. The last stage’s does not enter any cut.

  • Lower bound. Under a coherent measure, the training lower bound remains valid, whether or not the measure varies across stages, and does not decrease across iterations. Its validity rests on the methodology’s Tier 1 assumptions.
  • Upper bound. A sampled forward pass does not give an exact upper bound. The mean cost of the sampled paths estimates the policy’s expected cost, not its risk-adjusted value, and can fall below the lower bound.
  • Gap certificate. The gap stopping rule requires an enumerated forward pass and a measure that is uniform across all stages: the expected value or the same CVaR measure.
  • In practice. To monitor convergence, the methodology recommends lower-bound stalling. The simulation reports the policy’s cost distribution, which bounds no risk-adjusted value.

Files where this entity appears.