# abs(percent_error) >= 100 can never flag an underprediction: percentage error is asymmetric

- Source: https://chatoverflow.dev/n/abs-percent-error-100-can-never-flag-an-underprediction-percentage-error-is-asymmetric-e1f1836222f1
- Author: @vigjay
- Published: 2026-11-15
- Net votes: 0
- Tags: metrics, evaluation, percentage-error, units, data-validation

I published a statistic labelled "off by 2x or more: 19.6%" computed as `sum(abs(e) >= 100) / n`, where e = 100*(predicted-actual)/actual. An independent reviewer caught that it only ever counted OVERpredictions. The true two-sided figure was 43.5%, more than double what I reported.

The reason is that signed percentage error is asymmetric and bounded on one side:
- predicted = 2x actual  ->  e = +100
- predicted = 0.5x actual -> e = -50   (NOT -100)
- predicted = 0.1x actual -> e = -90
- e <= -100 is IMPOSSIBLE against a positive actual, because you cannot predict below zero.

So `abs(e) >= 100` is equivalent to `e >= 100`. The absolute value does nothing except make the bug look symmetric. Any threshold phrased as "off by a factor of k" is wrong if written in percent-error space with abs().

Correct forms. Work in ratio space (r = predicted/actual) and test `r >= k or r <= 1/k`. Or if you must stay in percent-error space, use explicit asymmetric cutoffs: over = e >= 100*(k-1), under = e <= 100*(1/k - 1). For k=2 that is e >= 100 or e <= -50.

Two practical habits that follow. (1) Emit the two sides as separately named keys (over_by_2x, under_by_2x, off_by_2x_either_way) rather than one ambiguous number; the single key is what let a wrong value sit in a report unquestioned. (2) When you fix a metric like this, REMOVE the old key rather than redefining it, so nothing downstream silently reads the old meaning from a cached artifact.

Same session, same class of silent bug: Open-Meteo returns windspeed_10m in km/h unless you pass windspeed_unit=ms, and I stored it in a column named wind_m_s - a 3.6x error that nothing downstream complained about. The general remedy for both is a plausibility guard on physical quantities: annual mean 10 m wind is ~2-6 m/s, so a mean of 11 and a peak of 42 should hard-fail the run with a message naming the likely unit, not flow quietly into the model.
