Abstract
The exponential proliferation of Internet of Things (IoT) devices and high-frequency sensor networks has precipitated a crisis in data storage and transmission bandwidth. Traditional lossless compression algorithms, which rely on rigid, hand-crafted statistical models, often fail to capture the complex, non-linear dependencies inherent in high-rate, stochastic time-series data. This paper introduces a novel neural compression framework specifically designed for high-rate time series, leveraging learned entropy models augmented by deep temporal priors. By utilizing a deep autoregressive network to approximate the conditional probability distribution of future samples based on historical context, our method minimizes the discrepancies between the estimated and actual data distributions. This allows for highly efficient entropy coding. We propose a hybrid architecture that combines causal dilated convolutions with a parameterized probability density function to achieve state-of-the-art compression ratios. Extensive experiments on diverse datasets, including seismic sensor streams and financial tick data, demonstrate that our approach significantly outperforms traditional codecs and existing deep learning baselines in terms of bits-per-dimension, while maintaining manageable computational latency.

This work is licensed under a Creative Commons Attribution 4.0 International License.
Copyright (c) 2025 William Sanchez, Mark White (Author)