Nick, Yes, that calculation is correct. I'm curious, for a typical datafile you're running, what does the window size calculate to? How does this compare to the "pattern window" (let's call these "frames" for clarity) size?
To get useful scores (i.e. scores that evaluate a real-time anomaly detector) from NAB it is required the detector under test evaluates point-by-point. I don't see this as a problem for your detector, provided you feed the data correctly -- one point at a time. Your detector must look at each subsequent data point as it comes in and calculate whether it is anomalous or not, before the next data point comes in. How the detector treats past data is up to you. That is, you can discretize the data into frames as it comes in (one point at a time). Note you cannot know the anomaly windows ahead of time; the frames your detector uses must be independent of the anomaly windows. Thank you for pointing out the discrepancy in Appendix F of the documentation. I will updated it accordingly. And yes, your approach for populating the results csv file values is correct. Best of luck with the defense! I'm sure a lot of us would love to see the presentation if you're able to share. Cheers, Alex Alexander Lavin Software Engineer Numenta
