An AR(p) model, or autoregressive model of order p, is a time series forecasting technique where the current value is regressed on its previous p values. It assumes a linear relationship and is used to capture temporal dependencies in data, making it suitable for stationary time series analysis.