Skip to content

Latest commit

 

History

History
28 lines (28 loc) · 751 Bytes

ARIMA or ARIMAX models.md

File metadata and controls

28 lines (28 loc) · 751 Bytes
tags
mida

It is sometimes of interest to model non-stationary residual to describe a drift phenomenon $$ w(t) = w(t-1) + \eta(t) $$ where $w$ is called a random walk. One thing that is important to remember is that the variance is not constant $$ \text{Var} [w(t)] = \text{Var} [w(t-1)] + \lambda^{2} = \lambda^{2}t $$ and in operatorial notation it can be represented as $$ w(t) = \frac{1}{1-z^{-1}}\eta(t) $$ This is basically a discrete integral of the [[White noise]]. To model equation we have a noise term that becomes $$ C(z)w(t) = \frac{C(z)}{1-z^{-1}}\eta(t) $$ giving us $$ (1 - z^{-1})A(z)y(t) = (1 - z^{-1})B(z)u(t-k) + C(z)\eta(t) $$ that in normal notation is $$ A(z)\Delta y(t) =B(z)\Delta u(t) + C(z)\eta(t) $$