🤖 AI Summary
This study addresses the common pedagogical disconnect in time series analysis, where autocorrelation function (ACF) and partial autocorrelation function (PACF) are often taught in isolation from regression principles. The paper proposes a unified regression-based perspective, systematically interpreting ACF as the ordinary least squares (OLS) coefficient from a simple linear regression on a lagged variable, and PACF as the partial regression coefficient of an additional lag in a multivariate regression framework. It further reveals that the Durbin–Levinson recursion fundamentally embodies a sequence of partial regression steps. Through illustrative AR(1) and MA(1) examples grounded in OLS theory and stationarity assumptions, the work clarifies the regression origins of ACF’s tailing-off behavior and PACF’s cutoff property, substantially lowering conceptual barriers. Practical teaching recommendations are offered to bridge the gap between regression and time series curricula.
📝 Abstract
The autocorrelation function (ACF) and partial autocorrelation function (PACF) are foundational tools for identifying autoregressive moving-average (ARMA) models, yet they are often introduced to students as computational recipes disconnected from the regression framework students already know. This note develops a unified, regression-based instructional framework for the ACF, the PACF, the Durbin-Levinson recursion, and one-step-ahead prediction in stationary time series. We show that the ACF is precisely the coefficient from a simple linear regression of a stationary process on one of its own lags, that the PACF is the coefficient on the newest lag in an expanding multiple regression model, and that the Durbin-Levinson recursion is nothing more than a sequence of partial regressions, each expressible using ordinary least-squares identities from a first course in regression. Worked examples for the MA(1) and AR(1) processes illustrate how the tailing-off and cutting-off behavior of the ACF and PACF emerge directly from this viewpoint. We argue that this framing lowers the conceptual barrier between regression courses and time series courses and provide classroom-oriented suggestions for instruction.