We aim at developing a high-performance implementation of Empirical Dynamic Modeling (EDM), an emerging framework for non-linear time series analysis. EDM enables a variety of analyses such as short-term forecasts, quantification of non-linearity, and causal inference. Although EDM is a generic modeling method for time series data, it was originally developed in the field of ecology, where available datasets are relatively small. Thus, the current libraries for EDM are not designed with performance in mind, and the scale of datasets that can be analyzed ar limited. To enable large-scale analysis using EDM, we have been developing a high-performance EDM implementation. In this research, we continue this effort by (1) porting time-consuming kernels in EDM to the SX-Aurora TSUBASA Vector Engine (2) enhancing the scalability of EDM by utilizing approximate algorithms (3) analyzing neural activity datasets to evaluate the performance of the ported implementation.

• kEDM (our EDM implementation)