In this talk we motivate the use of a multi-scale embedding of time series to localize anomalies in the given time series, in a online streaming setup. The reconstruction error produced by the projection of the new incoming window on to the principal component at each scale serves as the anomaly score for a given scale or window size. The aggregation of errors from multiple scales into a final anomaly score is also explored. The method was evaluated on the Yahoo! and Numenta datasets for streaming anomaly detection.