Is there a way to estimate the length of each period or seasonality in a time series in real-time? This is useful for adaptive and online time-series processing (streaming) where the entire data is not available a-priori.
Each observation may arrive sequentially or in a batch of observations. We want to be able to analyse observations as they come without having to store the entire historical data; though we may be able to store the most recent observations.
The natural way to assess the frequency of a periodicity is with a Fourier analysis. But your concern with not storing the sequence leads me to wonder whether you are looking for the frequency or are more concerned with detecting changes. Can you provide more information about what you want to know and what your data are?
Thank you Paul, yes I do not want to store the entire historical sequence but only the most recent K observations (K=100 for example). And yes, I am doing a forecast-based anomaly detection.
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics
- Statistical Techniques