Average Pinball Loss

Problem

Compute the mean pinball loss across several quantiles. forecasts holds one forecast list per quantile and quantiles the matching probabilities. For each pair let d = actual - forecast; add q * d when d is non-negative and (q - 1) * d otherwise, then average over all actual-forecast pairs.

Return 0.0 when there are no pairs.

Hint

Underprediction is penalized by q and overprediction by 1 - q.

Test cases

case 1

in: [[10,12,14],[[11,11,11],[15,15,15]],[0.5,0.9]]

expected: 0.5666666666666667

case 2

in: [[5],[[5]],[0.5]]

expected: 0

case 3

in: [[1,2],[[0,0],[3,3]],[0.1,0.9]]

expected: 0.15

case 4

in: [[1,2,3],[[2,2,2]],[0.5]]

expected: 0.3333333333333333

Starter codePython

def average_pinball_loss(actual, forecasts, quantiles):
    # Your code here
    pass

4 test cases · run your code against them in the editor