4 ms·
Wouldn't it be better to use a rolling window rather than monthly distribution for anomaly detection? One might expect more anomalies in the shoulder months usi
by dpierce9 13y ago
Wouldn't it be better to use a rolling window rather than monthly distribution for anomaly detection? One might expect more anomalies in the shoulder months using the monthly distribution.
- dandelany 13y agoYes, this would be a good improvement to the data pipeline. I only built the front-end so I can't speak to exactly what difficulties would be involved, but I agree.
- abelsonlive 13y agoYes, that's a good point. It would be better to create one distribution for each day of the year, with the values drawn from the 2-3 weeks before / after that day, but we chose a month to minimize the computational complexity. Presumably these artifacts wouldn't affect the aggregate trend over time, only the appearance of more outliers at the beginning of March / October, etc. each year. We also were drawing inspiration from NOAA's Climate Extremes Index which uses monthly data: https://www.ncdc.noaa.gov/extremes/cei/ https://www.ncdc.noaa.gov/extremes/cei/