Outlier Detection and Profiling Analyst

Outliers are both a data quality concern and a source of genuine insight. A value far outside the expected range might represent a measurement error, a data entry mistake, a system glitch — or a genuinely exceptional observation that deserves its own analysis. Knowing which it is, and handling each type correctly, requires a systematic approach that goes far beyond simply flagging values beyond three standard deviations. This AI role provides that systematic, multi-method outlier detection and profiling capability.

The assistant applies a layered outlier detection strategy. For univariate outlier detection, it uses IQR-based fencing, z-score and modified z-score (using median absolute deviation for robustness), Grubbs' test for single outlier testing, and visual detection via box plots and violin plots. It explains the assumptions behind each method and which is most appropriate for your variable's distribution — standard z-scores, for example, are misleading for skewed distributions.

For multivariate outlier detection, where a combination of values is unusual even if each individual value is plausible, the assistant applies Mahalanobis distance for normally distributed data, Local Outlier Factor (LOF) for density-based detection, Isolation Forest for high-dimensional anomaly scoring, and DBSCAN for cluster-based outlier identification. Each method returns an outlier score or binary flag, and the assistant helps you set thresholds based on your business context rather than arbitrary cutoffs.

Critically, every detected outlier is profiled rather than simply flagged: What is the outlier's value? In what context (which rows, which combinations of other variables) does it occur? What is the most likely explanation — measurement error, legitimate exceptional case, data pipeline issue? What is the business or statistical impact of including or excluding it? This profiling informs a disposition decision for each outlier type.

Ideal for data scientists, quality assurance analysts, fraud detection teams, financial auditors, and researchers who need to make principled, documented decisions about anomalous observations.