Abstract:
|
We discuss the construction of neural architectures for several tasks associated with change-point modeling in time series analysis: testing for the presence of change-points, determining the number of change-points, and estimating their locations. Working with both univariate and multivariate data, we cover changes in the level, variance and slope, amongst others. We make a number of observations regarding how the neural network appears to "learn" established statistics (e.g. CUSUM) in familiar simple cases, how it can be fooled, and how to prevent it from doing that. Finally, we discuss a simulation framework that enables the analyst to make use of this automatic change-point analysis engine even in the absence of natural training data.
|