A Perturbative Approach to Novelty Detection in Autoregressive Models

We propose a new method to perform novelty detection in dynamical systems governed by linear autoregressive models. The method is based on a perturbative expansion to a statistical test whose leading term is the classical F-test, and whose O(1/n) correction can be approximated as a function of the number of training points and the model order alone. The method can be justified as an approximation to an information theoretic test. We demonstrate on several synthetic examples that the first correction to the F-test can dramatically improve the control over the false positive rate of the system. We also test the approach on some real time series data, demonstrating that the method still retains a good accuracy in detecting novelties.