Metrics 0.3.0 maxerror score

Lodestar.Metrics 0.3.0. This page is frozen at that release. Read the current documentation for what main says now. A link to a decision or a migration page follows main, and leaves the archive.

MaxError.Score

The largest absolute difference between a true value and its prediction.

public static double Score(ReadOnlySpan<double> yTrue, ReadOnlySpan<double> yPred)

Parameters — yTrue and yPred are the true and predicted values, one per sample and the same length. There is nothing else: no weights and no multioutput.

Returns — double, never negative, 0 only when every prediction is exact. In the target's own units.

Exceptions — ArgumentException when the spans disagree in length, are empty, or hold a non-finite value.

Example — three predictions are perfect and one is out by 96.

using Lodestar.Metrics;

double[] yTrue = [1.0, 2.0, 3.0, 100.0];
double[] yPred = [1.0, 2.0, 3.0, 4.0];

double worst = MaxError.Score(yTrue, yPred);   // => 96

Remarks — this is the metric for a guarantee rather than for a summary: "no dose is ever off by more than this", "no invoice is ever wrong by more than this". If your requirement is a bound, this is the only number on the page that measures it, because every other one can be excellent while a single catastrophic prediction hides inside it — on the data above, MedianAbsoluteError.Score is 0.

The trap is the mirror image, and it is why nobody optimises this: it is decided by one sample, so it is as noisy as your worst label. One mistyped value in a dataset moves this number and no other. Report it beside an average, never instead of one.

The missing parameters are fidelity, not an oversight. max_error accepts no sample_weight and refuses a two-dimensional target outright, and the reason is real: a worst case is not an average, so there is nothing for a weight to scale.

Applies to — net10.0, netstandard2.0.

See also — MedianAbsoluteError.Score, MeanAbsoluteError.Score, MeanSquaredError.Score, the Python equivalence table.