Autoregressive Models

Abstract

Autoregressive models take advantage of the correlation between errors across time periods. Basic linear regression views this autocorrelation as a negative statistical property, a bias in error terms. Such bias often arises in cyclical data, where if the stock market price was high yesterday, it likely will be high today, as opposed to a random walk kind of characteristic where knowing the error of the last forecast should say nothing about the next error. Traditional regression analysis sought to wash out the bias from autocorrelation. Autoregressive models, to the contrary, seek to utilize this information to make better forecasts. It doesn’t always work, but if there are high degrees of autocorrelation, autoregressive models can provide better forecasts.