In many teams, KPIs are everywhere, but understanding when something is actually wrong is still a challenge. In People Analytics, for example, a sudden change in turnover, engagement, or role movement may point to a real problem, or it may simply reflect normal variation in noisy, human-centered data.
In this talk, I will walk through how to detect anomalies in KPIs using Python, drawing from practical experience working with real organizational metrics. We will start by building a clear intuition around what data anomalies are, and what they are not, and why simple rules or fixed thresholds often fail when metrics exhibit seasonality, trends, and structural changes.
From there, we will explore statistical, time-series, and machine learning approaches to anomaly detection, with concrete Python examples using libraries such as pandas, scikit-learn, and PyOD. I will show how these techniques can be applied to People Analytics metrics like employee turnover, engagement trends, and role rotation, and how the same challenges appear across many other business KPIs.
In some cases, I will also show how generative models can be used to support anomaly detection workflows, especially for summarizing patterns, providing context, and helping humans interpret unusual behavior in metrics. The focus of this talk, however, remains on practical decision-making. We will discuss how to choose the right method for your data, how to evaluate whether an anomaly is meaningful, and how to design monitoring that supports human interpretation instead of creating alert fatigue. Attendees will leave with a solid mental model for anomaly detection, reusable Python patterns, and a clearer understanding of how to apply these techniques to their own metrics.