๊ณผ๊ฑฐ๊ฐ€ ๋ฏธ๋ž˜๋ฅผ ์˜ˆ์ธกํ•œ๋‹ค

4 minute read

๊ณผ๊ฑฐ๊ฐ€ ๋ฏธ๋ž˜๋ฅผ ์˜ˆ์ธกํ•œ๋‹ค

<Auto-Regressiave Model>, โ€œAutoโ€๊ฐ€ ๋ถ™์€ ๊ฒƒ์—์„œ ์•Œ ์ˆ˜ ์žˆ๋“ฏ, ์‹œ๊ณ„์—ด $\{ X(t) \}$์—์„œ ๊ณผ๊ฑฐ ์ž์‹ ์˜ ๊ฐ’์ธ $X(t-1)$, $X(t-2)$๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๋ชจ๋ธ์ด๋‹ค. โ€œ๊ณผ๊ฑฐ๊ฐ€ ๋ฏธ๋ž˜๋ฅผ ์˜ˆ์ธกํ•œ๋‹คโ€๋ฅผ ๋ชจ๋ธ๋ง ํ•œ ๊ฒƒ์ด๋‹ค.์‹์€ ์•„๋ž˜์™€ ๊ฐ™๋‹ค.

Definition. Auto-Regressive Model

\[X(t) = \phi_0 + \phi_1 X(t-1) + \phi_2 X(t-2) + \cdots + \phi_p X(t-p) + \epsilon(t)\]

Multiple Regression Model์ด์ง€๋งŒ, ์ž์‹ ์˜ ๊ณผ๊ฑฐ ๊ฐ’์„ ์‚ฌ์šฉํ•˜๊ธฐ ๋•Œ๋ฌธ์— โ€œAutoโ€๋ผ๋Š” ํ‘œํ˜„์ด ๋ถ™์—ˆ๋‹ค.

Hyper-parameter๋Š” ๋ช‡๊ฐœ์˜ Lag๋ฅผ ์“ธ ๊ฒƒ์ธ์ง€์— ๋Œ€ํ•œ $p$ ๊ฐ’์ด๋‹ค. ์ด๋ฅผ ๊ธฐ์ค€์œผ๋กœ $p$์ฐจ AR ๋ชจ๋ธ์€

\[\text{AR}(p)\]

๋ผ๊ณ  ํ‘œํ˜„ํ•œ๋‹ค.

AR(1): 1์ฐจ AR ๋ชจ๋ธ

\[X(t) = \phi_0 + \phi_1 X(t-1) + \epsilon(t)\]

์œ„์˜ ์ˆ˜์‹์—์„œ $\phi_0$, $\phi_1$์˜ ๊ฐ’์— ๋”ฐ๋ผ ๋‹ค์–‘ํ•œ ๋ชจ๋ธ๋“ค์ด ํŒŒ์ƒ๋˜๋Š”๋ฐ,

Definition. White Noise Model

$\phi_0 = 0$ and $\phi_1 = 0$

\[X(t) = \epsilon(t)\]

๊ณผ๊ฑฐ์˜ ๊ฐ’์œผ๋กœ ํ˜„์žฌ์˜ ๊ฐ’์„ ์˜ˆ์ธกํ•  ์ˆ˜ ์—†๋‹ค.

Definition. Random Walk Model

$\phi_0 = 0$ and $\phi_1 = 1$

\[X(t) = X(t-1) + \epsilon(t)\]

์˜ค๋Š˜์˜ ๊ฐ’ $X(t)$๋Š” ์–ด์ œ์˜ ๊ฐ’ $X(t-1)$๊ณผ ์˜ˆ์ธกํ•  ์ˆ˜ ์—†๋Š” ๋ณ€๋Ÿ‰ $\epsilon(t)$๋กœ ๊ฒฐ์ •๋œ๋‹ค. ์–ด์ฉŒ๋ฉด ์–ด์ œ์˜ ๊ฐ’์ด ์˜ค๋Š˜์˜ ๊ฐ’์„ ์˜ˆ์ธกํ•˜๋Š”๋ฐ, ๋ฒ ์ŠคํŠธ ๊ฐ’์ด๋‹ค.

<Markove Process>๋Š” ํ™•๋ฅ ์ด ์ง์ „ ์ƒํƒœ์—๋งŒ ์˜์กดํ•˜๋Š” ํ™•๋ฅ  ๊ณผ์ •์ด๋‹ค.

\[p(S_{t+1} \mid S_0, S_1, \dots, S_t) = p(S_{t+1} \mid S_t)\]

๋”ฐ๋ผ์„œ AR(1)์€ <Markov Property>๋ฅผ ๋งŒ์กฑํ•œ๋‹ค๊ณ  ํ•  ์ˆ˜ ์žˆ๋‹ค.

Definition. Random Walk Model with drift

$\phi_0 \ne 0$ and $\phi_1 = 1$

\[X(t) = \phi_0 + X(t-1) + \epsilon(t)\]

Drift $\phi_0$๊ฐ€ ์กด์žฌํ•˜๋Š” ๊ฒฝ์šฐ์ž„. ์‹œ๊ณ„์—ด์ด ๋šœ๋ ทํ•œ ์ถ”์„ธ๋ฅผ ๊ฐ€์ง€๊ณ  ์žˆ์Œ.

Definition. AR(1) with Stationarity

$\left| \phi_1 \right| < 1$

\[X(t) = \phi_0 + \phi_1 X(t-1) + \epsilon(t)\]

๋งŒ์•ฝ $\left| \phi_1 \right| < 1$๋ผ๋ฉด, ์‹œ๊ณ„์—ด์€ ์ •์ƒ์„ฑ์„ ๊ฐ€์ง„๋‹ค.

AR model with Stationarity

AR(1)์˜ ๋งˆ์ง€๋ง‰์—์„œ $\left| \phi_1 \right| < 1$๋ผ๋ฉด, ์‹œ๊ณ„์—ด์ด ์ •์ƒ์„ฑ์„ ๊ฐ€์ง„๋‹ค๊ณ  ํ–ˆ๋‹ค. ํ•ด๋‹น ๊ฒฝ์šฐ๋ฅผ ์ข€๋” ์‚ดํŽด๋ณด์ž.

๋งŒ์•ฝ, $\left| \phi_1 \right| > 1$๋ผ๊ณ  ํ•ด๋ณด์ž. ๊ตฌ์ฒด์ ์œผ๋ก  $\phi_1 = 2$๋ผ๊ณ  ํ•ด๋ณด์ž. ๊ทธ๋Ÿฌ๋ฉด, ๋งค์Šคํ…๋งˆ๋‹ค $X(t)$์˜ ๊ฐ’์€ 2๋ฐฐ์”ฉ ๋Š˜์–ด๋‚  ๊ฒƒ์ด๋‹ค: 2 โ†’ 4 โ†’ 8 โ†’ 16 โ€ฆ ๊ทธ๋Ÿฌ๋ฉด, ์‹œ๊ณ„์—ด์— ์ฆ๊ฐ€ ์ถ”์„ธ(T)๊ฐ€ ์žˆ๋Š” ๊ฒƒ์ด๊ธฐ์— ์‹œ๊ณ„์—ด์ด ์ •์ƒ์„ฑ์„ ๋„์ง€ ์•Š๊ฒŒ ๋œ๋‹ค.

Forecasting: Principles and Practices: ์ž๊ท€ํšŒ๊ท€ ๋ชจ๋ธ

๋‘ ๋ชจ๋ธ ๋‹ค $\left| \phi_1 \right| < 1$๋ฅผ ๋งŒ์กฑํ•œ๋‹ค.

AR(1)์ด $\left| \phi_1 \right| < 1$๋ผ๋ฉด ์–ด๋–ป๊ฒŒ ๋˜๊ฒ ๋Š”๊ฐ€? ์‹œ๊ณ„์—ด์ด ์‹œ๊ณ„์—ด์˜ ํ‰๊ท ์„ ์ค‘์‹ฌ์œผ๋กœ ์ง„๋™ํ•  ๊ฒƒ์ด๋‹ค. ์œ„์˜ ๊ทธ๋ฆผ์ด ์ด๋ฅผ ์ž˜ ํ‘œํ˜„ํ•˜๊ณ  ์žˆ๋‹ค.


๊ทธ๋Ÿผ ์ •์ƒ์„ฑ์„ ๊ฐ€์ง„ AR(1)์ด ์ •๋ง๋กœ ์ •์ƒ์„ฑ์„ ๊ฐ€์ง„ ๊ฑด์ง€ ์ •์ƒ์„ฑ ์กฐ๊ฑด์„ ์ฒดํฌํ•ด๋ณด์ž.

๋งŒ์•ฝ, AR(1) ๋ชจ๋ธ์˜ ์‹œ๊ณ„์—ด $Z(t)$์ด ์ •์ƒ์„ฑ์„ ๊ฐ€์ง„๋‹ค๋ฉด, ๊ทธ๋•Œ์˜ ๋ถ„์‚ฐ์€ ์–ธ์ œ๋“  ๋™์ผํ•  ๊ฒƒ์ด๋‹ค.

\[\text{Var}(Z(t)) = \text{Var}(Z(t-1))\]

์ด๋ฅผ ์ด์šฉํ•ด AR(1)์˜ ์ˆ˜์‹์— ๋ถ„์‚ฐ์„ ์œ ๋„ํ•ด๋ณด์ž.

\[\begin{aligned} \sigma_Z^2 &= \text{Var}(Z(t)) \\ &= \phi_1^2 \cdot \text{Var}(Z(t)) + \text{Var}(\epsilon(t)) \\ &= \phi_1^2 \cdot \sigma_Z^2 + \sigma^2 \\ \end{aligned}\]

์ด์ œ ์‹์„ ์ •๋ฆฌํ•˜๋ฉด ์•„๋ž˜์™€ ๊ฐ™๋‹ค.

\[\sigma_Z^2 = \frac{\sigma^2}{1 - \phi_1^2}\]

์ด๋•Œ, $Z(t)$์˜ ๋ถ„์‚ฐ $\sigma_Z^2$๋Š” ์–‘์ˆ˜์—ฌ์•ผ ํ•˜๋ฏ€๋กœ, ๋ถ„๋ชจ์˜ $1 - \phi_1^2 > 0$์—ฌ์•ผ ํ•œ๋‹ค. ๋”ฐ๋ผ์„œ

\[\phi_1^2 < 1\]

์—ฌ์•ผ ์ •์ƒ์„ฑ์„ ๋งŒ์กฑํ•œ๋‹ค!


๋‹จ, $\phi_1^2 < 1$ ์กฐ๊ฑด์€ AR(1) ๋ชจ๋ธ์—์„œ์˜ ์ •์ƒ์„ฑ ์กฐ๊ฑด์ด๋‹ค. ๋‹ค๋ฅธ ์ฐจ์ˆ˜ $p$์˜ AR ๋ชจ๋ธ์—์„œ๋Š” ์กฐ๊ฑด์ด ๋‹ค๋ฅด๋‹ค.

  • AR(1) ๋ชจ๋ธ
    • $-1 < \phi_1 < 1$
  • AR(2) ๋ชจ๋ธ
    • $-1 < \phi_2 < 1$
    • $\phi_1 + \phi_2 < 1$
    • $\phi_2 - \phi_1 < 1$

Reference