Cleaning up chaos in data signals?

Ever noticed how data from sensors or apps can be full of noise? Think of an autoregressive model, a statistical tool that predicts future values based on past data. It works well, but when data is corrupted by strong or impulsive noise, predictions fail. Researchers developed an innovative method using self-supervised learning, a type of AI that learns on its own without labels. This approach, already used in computer vision, cleans autoregressive signals even under heavy-tailed noise like alpha-stable distributions. Tests showed excellent results on synthetic and semi-synthetic data, outperforming traditional methods. The goal? Recover the original signal and improve future forecasts. A step forward for weather, finance, and engineering applications.

The autoregressive (AR) model is a second-order stationary process widely used to model real-world phenomena. However, in the presence of additive noise, especially heavy-tailed or impulsive noise, classical estimation techniques show limited effectiveness. The proposed method draws inspiration from self-supervised neural networks, such as those used in Deep Image Prior, and does not require full knowledge of the noise distribution. Learning relies on a loss function designed to preserve the temporal structure of the clean signal. In infinite-variance cases, like alpha-stable distributions, the method remains stable due to robust regularization. Simulations show significant improvement in correlation coefficient and noise reduction compared to Kalman filters or robust regression. This adaptability to non-Gaussian scenarios makes the method highly relevant for applications in economics, geophysics, and telecommunications, where data often exhibits heavy tails.