Analysis Results
| # | Values | Score | Lower | Upper | Status | Severity | Action |
|---|
Outliers are highlighted separately. Pair plots require two selected features.
| Method | Detected | Percentage | Flagged rows |
|---|
Formula Used
Z-score: z = (x - mean) / standard deviation Modified Z-score: M = 0.6745(x - median) / MAD IQR: lower = Q1 - k(IQR), upper = Q3 + k(IQR) Mahalanobis: D² = (x - μ)ᵀΣ⁻¹(x - μ) LOF compares local reachability density with neighbouring observations. Isolation Forest uses shorter random-tree paths for unusual observations.
How to Use
Paste values, tabular data, or upload a CSV file. Inspect columns, then select one or several numeric features. Choose a detection mode before adjusting method settings.
Prepare missing values, duplicates, scaling, and transformations carefully. Run the calculator and review every highlighted observation. Never remove valid data without checking its context.
Compare methods when results are uncertain or inconsistent. Export results, cleaned data, or outlier-only records. Keep the original dataset before applying any treatment.
Example Data
| Day | Temperature | Pressure | Expected interpretation |
|---|---|---|---|
| 1 | 21.2 | 101.2 | Normal |
| 2 | 21.8 | 101.0 | Normal |
| 3 | 22.1 | 100.9 | Normal |
| 4 | 49.5 | 96.1 | Potential multivariate outlier |
| 5 | 21.6 | 101.1 | Normal |
Understanding Outliers
An outlier is an observation that differs greatly from others. It may represent error, rarity, change, fraud, or discovery. Statistical distance alone cannot explain its real meaning.
Univariate methods inspect one feature at a time. Multivariate methods inspect unusual combinations across several features. Time-series methods also respect order, trend, and local behaviour.
Outliers can distort means, variances, regressions, and model boundaries. Robust methods reduce this influence without hiding useful anomalies. Domain knowledge should guide every final treatment decision.
Frequently Asked Questions
Is every outlier an error?
No. Some outliers are valid rare events and may contain important information.
Which method should I use first?
Use IQR for skewed data and Z-scores for roughly normal data.
Why do methods disagree?
Each method uses different assumptions, neighbourhoods, distances, and thresholds.
What contamination value is suitable?
Use a realistic expected outlier proportion. Test several values before deciding.
Should data be scaled?
Scale multivariate features when their units or ranges differ substantially.
Can this analyse time series?
Yes. Select a time column and use rolling Z-score detection.
What is a modified Z-score?
It uses the median and MAD, making it more resistant to extremes.
When should I remove outliers?
Remove them only when evidence shows measurement, entry, or process errors.
What does Mahalanobis distance measure?
It measures multivariate distance while accounting for feature covariance.
What does LOF detect?
LOF detects observations with lower local density than their neighbours.
Can I export cleaned data?
Yes. Choose a treatment, run analysis, then download cleaned CSV data.
Is this a replacement for expert review?
No. Treat calculator output as evidence supporting a documented decision.