How to detect outliers in your file
Upload a CSV, TSV, JSON, JSONL, .ndjson, .txt, .xlsx, or .xls file. The tool parses the file and calculates results in your browser, then provides flagged observations through Export CSV or Export JSON without modifying the source data. .ndjson is read as JSONL, while .txt is read as comma-delimited data.
Select the data and detection settings
After the file loads, choose the settings. Detection updates automatically whenever the selected columns, method, threshold, grouping, or row identifier changes:
- Under Columns to analyze, select at least one column. A column appears here when at least 60% of its present, non-missing values are numeric. Columns with more non-numeric values may not appear.
- Choose a Method: IQR (interquartile range), Z-score, or Modified Z-score. Their default thresholds are
1.5,3, and3.5, respectively. - Review Threshold. It must be a positive number.
- Use Group by (optional) to calculate statistics separately for each value in any available column. A categorical column is one practical use case, but the control is not limited to categorical columns. Missing group values are shown as
(missing). - Use Row identifier to choose the column shown as the record ID. If the selected value is missing, the result falls back to
row N; if no identifier column is selected, row numbers are used.
The methods use different scores. IQR reports distance beyond the lower or upper bound in IQR units. Z-score reports the absolute z-score. Modified Z-score reports the absolute modified z-score, using 0.6745. IQR suits skewed data, Z-score assumes roughly normal data, and Modified Z-score uses the median and median absolute deviation.
Review detection results
Each automatic result includes Row ID, Column, Group, Value, Method, Threshold, Score, and Severity, sorted by descending score. The summary counts rows, analyzed values, excluded values, flagged results, severity levels, and results by column.
A result is flagged when its score reaches the selected threshold. Severity compares the score with that threshold: Mild is at least 1 and below 1.5; Moderate is at least 1.5 and below 2.5; Extreme is at least 2.5.
The on-screen table shows up to 500 flagged results. If more are found, the page displays a truncation note. Export CSV and Export JSON include the complete flagged set.
At least one analysis column must be selected. If the file has no detected numeric columns, the analysis waits for a usable selection. The tool also waits when Threshold is not positive.
Export the complete flagged results
When results are ready, use Export CSV for a CSV file or Export JSON for a JSON file. CSV includes row_id, column, group, value, method, threshold, score, and severity. JSON includes the result object and flagged observations. Both exports include results beyond the 500-row display limit; they do not contain complete original source rows.
Supported file structures and limitations
- JSON must be an array of record objects, either as the root value or inside an object property.
- JSONL and
.ndjsonare read as one JSON object per non-empty line. A.txtfile is read as comma-delimited data. - For spreadsheets, only the first worksheet is read.
- Unreadable files and empty files produce an error.
- Missing, blank, and non-numeric values in analyzed columns are excluded from calculations and counted separately. Values are not treated as zero or filled in. Missing Group by (optional) values are assigned to the visible
(missing)group. - A column or group is skipped and listed when it has fewer than three valid values, or when the selected method has zero IQR, zero standard deviation, or zero median absolute deviation. These conditions do not require every value to be identical.
The excluded-value count is calculated during processing for each selected column. Because one source row can be excluded for more than one selected column, this count does not necessarily represent distinct source rows.
Processing and data handling
The local implementation parses the file and performs the calculations in the browser. The page code does not establish account requirements, usage limits, a formal privacy policy, or broader server-side security guarantees.
Frequently Asked Questions
Does this tool remove or change outlier values?
No. It only detects and reports outliers - it never deletes, caps, replaces, or otherwise modifies your data. If you want to remove or correct flagged records, use the Clean Tools section after reviewing this report.
Which detection method should I use?
IQR is robust to skewed data and doesn't assume a normal distribution. Z-score assumes roughly normal data and is sensitive to extreme values pulling the mean. Modified z-score uses the median and median absolute deviation, making it more robust than a standard z-score while still highlighting extreme points.
How are missing and constant values handled?
Missing, blank, and non-numeric values in analyzed columns are excluded from calculations and counted separately - they are never treated as zero or guessed. Missing values in Group by (optional) appear in the (missing) group. A column or group is skipped when it has fewer than three valid values or when the selected method has zero IQR, zero standard deviation, or zero median absolute deviation.
How is this different from the Dataset Profiler?
The Dataset Profiler summarizes distributions, types, and completeness across a whole file. This tool performs configurable, row-level outlier scoring and returns flagged observations with a method, threshold, score, and severity for each one.