Long-memory time series ensembles for concept shift detection

Usually time series are controlled by generative processes which display changes over time. On many occasions, two or more generative processes may switch forcing the abrupt replacement of a fitted time series model by another one. We claim that the incorporation of past data can be useful in the presence of concept shift. We believe that history tends to repeat itself and from time to time, it is desirable to discard recent data reusing old past data to perform model fitting and forecasting. We address this challenge by introducing an ensemble method that deals with long-memory time series. Our method starts by segmenting historical time series data to identify data segments which present model consistency. Then, we project the time series by using data segments which are close to current data. By using a dynamic time warping alignment function, we try to anticipate concept shifts, looking for similarities between current data and the prequel of a past shift. We evaluate our proposal on non-stationary and non-linear time series. To achieve this we perform forecasting accuracy testing against well known state-of-the-art methods such as neural networks and threshold auto regressive models. Our results show that the proposed method anticipates many concept shifts.