Example: cross-temporal data
A two-level hierarchy with \(n = 8\) monthly time series. In the cross-sectional framework, at any time it is \(Tot = A + B + C\), \(A = AA + AB\) and \(B = BA + BB\) (\(nb = 5\) at the bottom level). For monthly data, the observations are aggregated to annual \((k = 12)\), semi-annual \((k = 6)\), four-monthly \((k = 4)\), quarterly \((k = 3)\), and bi-monthly \((k = 2)\) observations. The monthly bottom time series are simulated from five different SARIMA models. There are 180 monthly observations (15 years): the first 168 values (14 years) are used as training set, and the last 12 form the test set.
Cross-sectional hierarchy
Simulation
In the following script we simulate five independent monthly bottom time series, each of length 180 (15 complete years of monthly data).
library(FoReco)
library(forecast)
library(sarima)
values <- NULL
base <- NULL
residuals <- NULL
test <- NULL
bottom <- matrix(NA, nrow = 180, ncol = 5)
# Model definition
bts <- list()
#ARIMA(1,0,0)(0,0,0)[12]
bts[[1]] <- list(ar=0.31,
nseasons=12)
#ARIMA(0,0,1)(0,0,0)[12]
bts[[2]] <- list(ma=0.61,
nseasons=12)
#ARIMA(0,1,1)(0,1,1)[12]
bts[[3]] <- list(ma=-0.1,
sma=-0.12,
iorder=1,
siorder=1,
nseasons=12)
#ARIMA(2,1,0)(0,0,0)[12]
bts[[4]] <- list(ar=c(0.38,0.25),
iorder=1,
nseasons=12)
#ARIMA(2,0,0)(0,1,1)[12]
bts[[5]] <- list(ar=c(0.30,0.12),
sma=0.23,
siorder=1,
nseasons=12)
mm <- c(58.85, 60.68, 59.26, 35.47, 58.61)
set.seed(525)
for(i in 1:5){
bottom[,i] <- mm[i] + sim_sarima(n=180, model = bts[[i]],
n.start = 200)
}
colnames(bottom) <- c("AA", "AB", "BA", "BB", "C")
C <- matrix(c(rep(1,5),
rep(1,2), rep(0,3),
rep(0,2), rep(1,2), 0), byrow = TRUE, nrow = 3)
upper <- bottom%*%t(C)
colnames(upper) <- c("T", "A", "B")
values$k1 <- ts(cbind(upper, bottom), frequency = 12)
colnames(values$k1) <- c("T", "A", "B", "AA", "AB", "BA", "BB", "C")More precisely, AA is simulated from an AR(1) process, AB from an MA(1), BA from an ARIMA(0,1,1)(0,1,1), BB from an ARIMA(2,1,0), and C from an ARIMA(2,0,0)(0,1,1). The higher levels series in the hierarchy (T,A,B) are obtained by simple summation of the five bottom time series.
Then we compute the temporally aggregated series at annual \((k = 12)\), semi-annual \((k = 6)\), four-monthly \((k = 4)\), quarterly \((k = 3)\), and bi-monthly \((k = 2)\) frequencies.
# BI-MONTHLY SERIES
values$k2 <- ts(apply(values$k1, 2,
function(x) colSums(matrix(x, nrow = 2))),
frequency = 6)
# QUARTERLY SERIES
values$k3 <- ts(apply(values$k1, 2,
function(x) colSums(matrix(x, nrow = 3))),
frequency = 4)
# FOUR-MONTHLY SERIES
values$k4 <- ts(apply(values$k1, 2,
function(x) colSums(matrix(x, nrow = 4))),
frequency = 3)
# SEMI-ANNUAL SERIES
values$k6 <- ts(apply(values$k1, 2,
function(x) colSums(matrix(x, nrow = 6))),
frequency = 2)
# ANNUAL SERIES
values$k12 <- ts(apply(values$k1, 2,
function(x) colSums(matrix(x, nrow = 12))),
frequency = 1)
The first 14 years of each simulated series are used as training set,
and the last year as test set. The forecasts are obtained using the
auto.arima function of the forecast package
(Hyndman et al., 2020).
# MONTHLY FORECASTS
base$k1 <- matrix(NA, nrow = 12, ncol = ncol(values$k1))
residuals$k1 <- matrix(NA, nrow = 168, ncol = ncol(values$k1))
for (i in 1:ncol(values$k1)) {
train <- values$k1[1:168, i]
forecast_arima <- forecast(auto.arima(train), h = 12)
base$k1[, i] <- forecast_arima$mean
residuals$k1[, i] <- forecast_arima$residuals
}
base$k1 <- ts(base$k1, frequency = 12, start = c(15, 1))
colnames(base$k1) <- c("T", "A", "B", "AA", "AB", "BA", "BB", "C")
residuals$k1 <- ts(residuals$k1, frequency = 12)
colnames(residuals$k1) <- c("T", "A", "B", "AA", "AB", "BA", "BB", "C")
test$k1 <- values$k1[-c(1:168), ]The following plots show the actual values and the forecasts for the test year at any temporal aggregation level.
# BI-MONTHLY FORECASTS
base$k2 <- matrix(NA, nrow = 6, ncol = ncol(values$k2))
residuals$k2 <- matrix(NA, nrow = 84, ncol = ncol(values$k2))
for (i in 1:ncol(values$k2)) {
train <- values$k2[1:84, i]
forecast_arima <- forecast(auto.arima(train), h = 6)
base$k2[, i] <- forecast_arima$mean
residuals$k2[, i] <- forecast_arima$residuals
}
base$k2 <- ts(base$k2, frequency = 6, start = c(15, 1))
colnames(base$k2) <- c("T", "A", "B", "AA", "AB", "BA", "BB", "C")
residuals$k2 <- ts(residuals$k2, frequency = 6)
colnames(residuals$k2) <- c("T", "A", "B", "AA", "AB", "BA", "BB", "C")
test$k2 <- values$k2[-c(1:84), ]