Time Series Anomaly Detection (TSAD) is the task of monitoring a time series, say an ECG, or the pressure in an industrial boiler, while attempting to recognize when there has been an anomalous event. The anomalies could be the beginning of heart attack, or a leak in the boiler that will cause the industrial product to spoil. TSAD is a commercially important problem, by most estimates worth billions of dollars each year. In the last few years there has been an explosion of interest in TSAD, with dozens of papers appearing in the top conferences and journals each year. 

In this talk I will make a surprising claim. At least 95% of the papers on TSAD are deeply flawed, and are at best unreliable. These flaws come from two major sources, using unsuitable metrics of success, and testing on flawed benchmark datasets. I will demonstrate my claims with forcefully, simple, visual examples. I will then go on to suggest some simple fixes for these issues. I will conclude by briefly touching on two interesting meta-questions. How did the research community not spot these issues before? And, what does it say that when these issues are pointed out, most of the community offers no counterarguments, but just ignores the problem (the head-in-the-sand response).
 
Joint work with Renjie Wu.

Dr. Eamonn Keogh