Twenty Years of Time Series Classification: From Nearest-Neighbors to ROCKETs and Large Language Models

Time series classification, a key area within artificial intelligence and machine learning, has widespread applications across economics, finance, medicine, and engineering. Such applications include verifying signatures on touchscreens, recognizing physical activities using accelerometer data, identifying users through keystroke dynamics, and many more. This paper provides an overview of the most significant time series classification methods developed over the last two decades. These methods range from nearest neighbor models with dynamic time warping, to deep learning techniques like convolutional and residual networks (ResNet), and more recent approaches such as the Random Convolutional Kernel Transform (ROCKET). A notable insight from this review is that relatively simple models often perform remarkably well for time series classification tasks.