A1409
Title: A new motif discovery based method for forecasting and imputation of functional data
Authors: Jacopo Di Iorio - Emory University (United States) [presenting]
Abstract: Forecasting, involving the prediction of future values and/or the evolution of functional observations, has always been a major goal of functional data analysis. Given the increasing attention in the field of functional motif discovery, a method is proposed for functional forecasting that involves the identification of functional motifs, i.e., typical "shapes" or "patterns" recurring multiple times within a single curve and/or across misaligned portions of multiple curves. Portions characterized by the same motif are hypothesized to be more likely to evolve similarly; therefore, the identification of the last portion of a function before the prediction part as a portion can help in the forecasting exercise. Extensive diagnostics are performed to guide the user not only in tuning parameters but also in validating the aforementioned hypothesis, thus ensuring the applicability of the method. Method performance is assessed through simulations, and it is applied to a real-data case study. Similarly, the same methodology can be applied to the imputation of missing portions of a curve.
