my-sire-that-is-an-outlier 0.1.0

Outlier detection for semi-normal distributions using the mirroring technique

mirror-mirror

Outlier detection for semi-normal distributions.

Most outlier detection methods assume your data is fully normal. But what if it's only half normal — a clear peak with one meaningful tail? mirror-mirror handles exactly that.

How it works

1. Find the peak of your distribution (via KDE or max value)
2. Take the tail-side data (everything at or below the peak)
3. Mirror it across the peak to create a synthetic full normal distribution
4. Compute mean and std of the synthetic distribution
5. Flag anything below mean − threshold × std as an outlier

Installation

pip install mirror-mirror

Quick start

from mirror_mirror import find_outliers

# Example: exam scores — most students score high, outliers at the low end
scores = [95, 97, 92, 88, 91, 99, 100, 94, 96, 72, 65, 55, 48]

result = find_outliers(scores, method="mode", threshold=2)

print(result["outliers"])   # array([48])
print(result["cutoff"])     # e.g. 62.4

API

find_outliers(data, method="mode", threshold=2.0)

Parameter	Type	Default	Description
data	list / np.ndarray / pd.Series	required	1-D numeric data. NaNs are ignored.
method	"mode" or "max"	"mode"	How to locate the peak of the distribution.
threshold	float	2.0	Sigmas below the mirror mean to cut off at.

Return value

A dict with:

Key	Description
outliers	np.ndarray of values identified as outliers
cutoff	The threshold value — points below this are outliers
mirror_mean	Mean of the synthetic mirrored distribution
mirror_std	Std dev of the synthetic mirrored distribution

method parameter

• "mode" (default) — uses KDE to find the most-dense region of the data as the peak. Best for real-world, noisy data.
• "max" — uses the maximum value as the peak. Useful when your data is naturally bounded at the top (e.g. a 100-point exam, property values with a known ceiling).

threshold parameter

Controls sensitivity. A value of 2.0 means any point more than 2 standard deviations below the mirror mean is flagged. Increase for fewer (more extreme) outliers; decrease for more.

When to use this

mirror-mirror works well when your data:

• Has a clear, single peak
• Has one natural tail (skewed left or right)
• Is not bimodal or uniformly distributed

Examples: exam scores, property values, response times, sensor readings with a hard ceiling.

License

MIT