참고 : https://otexts.com/fppkr/

[ 계층적/그룹화된 시계열 ]

Ex) 자전거 판매 수

  • 산악 자전거
  • 로드 바이크
  • 어린이용 자전거


시계열 모음(collection)은 계층적인 합산 구조(hierarchical aggregation structure)를 따름

\(\rightarrow\) 계층적 시계열(hierarchical time series)

  • ex) 지역-주-매장


제품 종류와 지리적인 위치를 동시에 고려하여 분배(disaggregate) 가능

\(\rightarrow\) 그룹화된 시계열 (grouped time series)


1. 계층적 시계열

ex) \(K=2\) 수준 계층 구조

  • 수준 0 : TOTAL …. 수준 2 : 가장 하단
  • Notation : \(y_{j.t}\)

figure2


\(y_{t}=y_{\mathrm{AA}, t}+y_{\mathrm{AB}, t}+y_{\mathrm{AC}, t}+y_{\mathrm{BA}, t}+y_{\mathrm{BB}, t}\).

  • \(y_{\mathrm{A}, t}=y_{\mathrm{AA}, t}+y_{\mathrm{AB}, t}+y_{\mathrm{AC}, t}\).
  • \(y_{\mathrm{B}, t}=y_{\mathrm{BA}, t}+y_{\mathrm{BB}, t}\).


Matrix Formulation

  • \(\boldsymbol{y}_{t}=\boldsymbol{S} \boldsymbol{b}_{t}\).

  • \(S\) : 합산 행렬 (summing matrix)

\(\left[\begin{array}{c} y_{t} \\ y_{\mathrm{A}, t} \\ y_{\mathrm{B}, t} \\ y_{\mathrm{AA}, t} \\ y_{\mathrm{AB}, t} \\ y_{\mathrm{AC}, t} \\ y_{\mathrm{BA}, t} \\ y_{\mathrm{BB}, t} \end{array}\right]=\left[\begin{array}{ccccc} 1 & 1 & 1 & 1 & 1 \\ 1 & 1 & 1 & 0 & 0 \\ 0 & 0 & 0 & 1 & 1 \\ 1 & 0 & 0 & 0 & 0 \\ 0 & 1 & 0 & 0 & 0 \\ 0 & 0 & 1 & 0 & 0 \\ 0 & 0 & 0 & 1 & 0 \\ 0 & 0 & 0 & 0 & 1 \end{array}\right]\left[\begin{array}{c} y_{\mathrm{AA}, t} \\ y_{\mathrm{AB}, t} \\ y_{\mathrm{AC}, t} \\ y_{\mathrm{BA}, t} \\ y_{\mathrm{BB}, t} \end{array}\right]\).


2. 그룹화된 시계열

계층적 시계열보다 더 일반적인 합산 구조

ex) \(K=2\)-수준 그룹화된 구조

figure2


\(y_{t}=y_{\mathrm{AX}, t}+y_{\mathrm{AY}, t}+y_{\mathrm{BX}, t}+y_{\mathrm{BY}, t}\).

  • 표현 방법 1)
    • \(y_{\mathrm{A}, t}=y_{\mathrm{AX}, t}+y_{\mathrm{AY}, t}\).
    • \(y_{\mathrm{B}, t}=y_{\mathrm{BX}, t}+y_{\mathrm{BY}, t}\).
  • 표현 방법 2)
    • \(y_{\mathrm{X}, t}=y_{\mathrm{AX}, t}+y_{\mathrm{BX}, t}\).
    • \(y_{\mathrm{Y}, t}=y_{\mathrm{AY}, t}+y_{\mathrm{BY}, t}\).


\(\left[\begin{array}{c} y_{t} \\ y_{\mathrm{A}, t} \\ y_{\mathrm{B}, t} \\ y_{\mathrm{X}, t} \\ y_{\mathrm{Y}, t} \\ y_{\mathrm{AX}, t} \\ y_{\mathrm{AY}, t} \\ y_{\mathrm{BX}, t} \\ y_{\mathrm{BY}, t} \end{array}\right]=\left[\begin{array}{llll} 1 & 1 & 1 & 1 \\ 1 & 1 & 0 & 0 \\ 0 & 0 & 1 & 1 \\ 1 & 0 & 1 & 0 \\ 0 & 1 & 0 & 1 \\ 1 & 0 & 0 & 0 \\ 0 & 1 & 0 & 0 \\ 0 & 0 & 1 & 0 \\ 0 & 0 & 0 & 1 \end{array}\right]\left[\begin{array}{c} y_{\mathrm{AX}, t} \\ y_{\mathrm{AY}, t} \\ y_{\mathrm{BX}, t} \\ y_{\mathrm{BY}, t} \end{array}\right]\).


3. 상향식 접근 방식

\(\begin{aligned} \tilde{y}_{h} &=\hat{y}_{\mathrm{AA}, h}+\hat{y}_{\mathrm{AB}, h}+\hat{y}_{\mathrm{AC}, h}+\hat{y}_{\mathrm{BA}, h}+\hat{y}_{\mathrm{BB}, h}, \\ \tilde{y}_{\mathrm{A}, h} &=\hat{y}_{\mathrm{AA}, h}+\hat{y}_{\mathrm{AB}, h}+\hat{y}_{\mathrm{AC}, h}, \\ \text { and } \tilde{y}_{\mathrm{B}, h} &=\hat{y}_{\mathrm{BA}, h}+\hat{y}_{\mathrm{BB}, h} . \end{aligned}\).


\(\tilde{\boldsymbol{y}}_{h}=\boldsymbol{S} \hat{\boldsymbol{b}}_{h}\).

\(\left[\begin{array}{c}\tilde{y}_{h} \\ \tilde{y}_{\mathrm{A}, h} \\ \tilde{y}_{\mathrm{AA}, h} \\ \tilde{y}_{\mathrm{AB}, h} \\ \tilde{y}_{\mathrm{AC}, h} \\ \tilde{y}_{\mathrm{BA}, h} \\ \tilde{y}_{\mathrm{BB}, h}\end{array}\right]=\left[\begin{array}{ccccc}1 & 1 & 1 & 1 & 1 \\ 1 & 1 & 1 & 0 & 0 \\ 0 & 0 & 0 & 1 & 1 \\ 1 & 0 & 0 & 0 & 0 \\ 0 & 1 & 0 & 0 & 0 \\ 0 & 0 & 1 & 0 & 0 \\ 0 & 0 & 0 & 1 & 0 \\ 0 & 0 & 0 & 0 & 1\end{array}\right]\left[\begin{array}{c}\hat{y}_{\mathrm{AA}, h} \\ \hat{y}_{\mathrm{AB}, h} \\ \hat{y}_{\mathrm{AC}, h} \\ \hat{y}_{\mathrm{BA}, h} \\ \hat{y}_{\mathrm{BB}, h}\end{array}\right]\).

4. 하향식 접근 방법

\(p_1, \cdots p_m\) :

  • 시계열의 예측값이 어떻게 분배되는지 나타내는 분배비율(disaggregation proportion)의 집합


\[\begin{aligned} &\tilde{y}_{\mathrm{AA}, t}=p_{1} \hat{y}_{t}, \quad \tilde{y}_{\mathrm{AB}, t}=p_{2} \hat{y}_{t}, \quad \tilde{y}_{\mathrm{AC}, t}=p_{3} \hat{y}_{t}, \quad \tilde{y}_{\mathrm{BA}, t}=p_{4} \hat{y}_{t} \text { and } \quad \tilde{y}_{\mathrm{BB}, t}=p_{5} \hat{y}_{t} . \\ \end{aligned}\]
  • \(\tilde{\boldsymbol{b}}_{t}=\boldsymbol{p} \hat{y}_{t}\).
  • \(\tilde{\boldsymbol{y}}_{h}=\boldsymbol{S} \boldsymbol{p} \hat{y}_{t}\).


(1) 과거 “비율값”을 평균

top-down Gross-Sohl method A

\(p_{j}=\frac{1}{T} \sum_{t=1}^{T} \frac{y_{j, t}}{y_{t}}\).


(2) 과거 “평균값”의 비율

top-down Gross-Sohl method F

\(p_{j}=\sum_{t=1}^{T} \frac{y_{j, t}}{T} / \sum_{t=1}^{T} \frac{y_{t}}{T}\).


(3) 예측 비율값

위의 2 방법은, 비율이 시간에 따라서 변할 수도 있다는 것을 고려하지 못하기 때문에!

\(\rightarrow\) (과거 비율에 근거한) 하향식 접근 방식 < 상향식 접근 방식


(생략)