Outlier
Also known as: anomaly, data anomaly, outlier detection, anomaly detection, influential observation, extreme value
An outlier is a data point that falls far outside the typical range of values in a dataset -- either much higher or much lower than the rest of the data.
Outliers can arise from measurement errors (miscoded values, sensor malfunctions), data entry mistakes, genuine but rare extreme events (a single record-breaking sale, a catastrophic one-time loss), or they may represent genuinely different populations mixed into the data. Whether to remove or retain outliers depends on their cause: data errors should be corrected or removed; genuine rare events should be retained and analyzed. Outliers have outsized influence on statistical measures like the mean and regression coefficients.
Outlier, explained at three levels
At home
An outlier is a data point that stands out from the rest because it is much larger or smaller than typical values. If your monthly spending is usually between $2,000 and $3,000, a month where you spent $10,000 (because you bought a car) would be an outlier. Outliers matter because they can distort statistical summaries: if you calculated your 'average' monthly spending including the car purchase month, it would make your typical spending look much higher than it actually is. Identifying outliers helps you decide whether to treat an extreme value as unusual data or as a real but rare event.
At work
In business data, outliers appear frequently: a single unusually large customer order, an error in data entry that recorded a price as $10,000 instead of $100, a day when website traffic spiked 50x due to a viral post, or a week when production was halted for equipment maintenance. Each type of outlier requires a different response: data entry errors should be corrected; genuine unusual events (the viral spike, the maintenance week) should be retained but noted. The first step when you encounter an outlier is always to investigate its cause before deciding how to handle it.
At home
Methods for detecting outliers: the interquartile range (IQR) method defines outliers as values below Q1 - 1.5*IQR or above Q3 + 1.5*IQR, where Q1 and Q3 are the 25th and 75th percentiles. This is robust to the outliers themselves (unlike methods based on mean and standard deviation, which outliers distort). The three-sigma rule (values more than 3 standard deviations from the mean) is sensitive to outliers in the mean and SD calculations -- a single extreme outlier inflates both, making them less useful for detection. Z-score methods (measuring how many standard deviations from the mean) face the same issue. For strongly skewed data, log-transform before outlier detection or use IQR-based methods.
At work
Influence vs. outlier: in regression analysis, an outlier in the predictor variables (high 'leverage point') can have outsized influence on the regression line even if its y-value isn't extreme. A data point can have low leverage (not unusual in predictor values) but high residual (far from the regression line). 'Influential observations' (Cook's distance > 1) combine high leverage and high residual -- they substantially change the regression coefficients if removed. It is standard practice to examine influential observations individually: are they data errors, or genuine unusual cases that the model should reflect?
At home
Outliers as signals vs. noise: not all outliers are problems to be removed. In many applications, outliers are the most interesting data points. In fraud detection, fraudulent transactions are outliers compared to normal spending behavior -- the goal is to identify these anomalies, not remove them. In quality control, defective units are outliers in the production distribution. In customer analytics, the best customers (extremely high lifetime value) and worst customers (extremely high support costs) are outliers worth understanding and targeting differently. The decision about outlier treatment should always consider whether the outlier might be a signal carrying important information.
At work
Multivariate outlier detection: in datasets with many variables, a data point that is not extreme on any individual variable can still be an outlier in the multivariate sense -- it is an unusual combination of values. Mahalanobis distance measures how many standard deviations a point is from the multivariate mean (accounting for correlations between variables). Isolation Forest is an ML algorithm that detects outliers by noting that unusual data points are easier to 'isolate' (separate from other points) with random splits. These multivariate methods are essential for fraud detection and network anomaly detection, where the anomaly signal exists only in combinations of features, not in any single feature alone.
Real-world examples
At home
Marcus was analyzing his investment portfolio's monthly returns. He noticed one month showed a +45% return -- far outside his typical range of -3% to +5%. Investigating, he found this was the month he had received a large inheritance and deposited it into the account, not an actual investment return. He removed this data point from his return analysis (it was a data artifact, not an investment result) and recalculated his actual investment performance separately from the deposit.
At work
Jerome managed analytics for an online marketplace. His average order value analysis showed a mean of $250 -- much higher than the $75 median. Investigating, he found 50 orders of $50,000+ (representing a small number of wholesale bulk orders from businesses). These outliers were not data errors -- they were real, legitimate orders from a different customer type than the typical individual consumer. He segmented the analysis: individual consumer orders (median $72) vs. business bulk orders (median $15,000). The combined mean of $250 was meaningless, obscuring both customer types. The outlier investigation revealed an important customer segmentation insight.
Common confusions
- Many analysts assume outliers should always be removed from datasets to 'clean' the data — Removing outliers without investigation is a serious analytical error. Outliers may represent genuine rare events that are important (fraud transactions, exceptional customers, catastrophic failures), they may represent data from a distinct population that should be analyzed separately rather than combined, or they may be errors that should be corrected rather than deleted. Blindly removing outliers based on statistical thresholds risks: removing the most interesting data points, hiding evidence of problems (removing quality defects from production data), or improving model metrics on training data while hurting real-world performance on the genuine extremes the model will encounter.
- Some assume that a single outlier in a large dataset is statistically irrelevant because it represents such a small fraction of the data — A single extreme outlier can dominate statistical summaries despite representing one data point: one value of $1 billion in an otherwise $100-1,000 range data set will pull the mean dramatically upward. In regression analysis, a single high-leverage outlier can change the direction and magnitude of regression coefficients substantially. In machine learning model training, a few extreme outliers can cause training instability or cause the model to overfit to those points. The proportion of data that outliers represent is less important than their magnitude and the statistical measure's sensitivity to extreme values.
For analysts
Outlier analysis is a valuable early step in any data analysis engagement. When reviewing client data, an outlier investigation commonly reveals: data quality problems (miscoded values, system errors) that undermine the reliability of subsequent analysis; evidence of data from distinct subpopulations that should be analyzed separately; and unusual events (one-time losses, exceptional customers) that distort aggregate metrics. The practical protocol: visualize distributions before analysis, identify statistical outliers using IQR-based methods, investigate each outlier individually to determine cause, and document handling decisions (corrected, excluded, retained with explanation) in the analysis methodology.
History & origin
Statistical awareness of outliers dates to the 19th century, when astronomers first dealt systematically with anomalous measurements in astronomical observations. Methods for detecting outliers (Grubbs' test, 1950; interquartile range method) were developed in the 20th century. The 'three-sigma rule' (values more than 3 standard deviations from the mean are outliers) became a practical convention in many fields. The concept gained broad cultural relevance through Malcolm Gladwell's 'Outliers' (2008), which used the statistical term to describe people with exceptional performance (though in a different sense than statistical usage). In data science, outlier detection has become a distinct specialty: algorithms like Isolation Forest, Local Outlier Factor, and DBSCAN automate anomaly detection at scale for use cases like fraud detection, network intrusion detection, and quality control.