An accurate probabilistic step finder for time-series analysis.

Many time-series data sets exist which are challenging to analyze with current state-of-the-art methods, either because they contain excessive noise or because they violate one or more assumptions inherent to the chosen analysis method. To our knowledge, BNP-Step is the first time-series analysis algorithm which leverages Bayesian nonparametrics to learn the number and location of transitions between states and the emission levels associated to each state, while providing rigorous estimates of uncertainty for the learned quantities. We anticipate our algorithm will allow analysis of data sets at levels of noise or sparsity beyond what current state-of-the-art methods allow, and could potentially reveal previously unknown features in data sets analyzed using existing methods.