Autoregression
Autoregression is a mathematical technique where a value is predicted using its own past values. The term appears across distinct fields with different meanings and applications.
Meanings
- Autoregression (statistics) — a Time series modeling approach where future observations depend on previous ones
- Autoregression (machine learning) — a generative approach in neural networks that generates sequences one element at a time
- Autoregression (linguistics) — sound change patterns where linguistic features influence neighboring sounds
In statistics, autoregressive models (AR models) assume that correlations exist between sequential data points—like how today's weather partly depends on yesterday's. This is foundational to Time series forecasting.
In machine learning, autoregressive models generate text, images, or audio by predicting the next token conditioned on all previous tokens. This powers many modern language models.
The concept reflects a deep pattern: systems often contain echoes of themselves, where the past shapes the present. Understanding autoregression means recognizing how distributions evolve through time, a principle relevant across nature, Economics, and Cognitive science.
Related
Time series, Probability, Generalization (machine learning), Sequence modeling, Forecasting