The package bayesRecon
implements probabilistic reconciliation of hierarchical time series forecasts via conditioning.
The main functions are:
reconc_gaussian
: implements analytic formulae for the reconciliation of Gaussian base forecasts;reconc_BUIS
: a generic tool for the reconciliation of any probabilistic time series forecast via importance sampling; this is the recommended option for non-Gaussian base forecasts;reconc_MCMC
: a generic tool for the reconciliation of probabilistic count time series forecasts via Markov Chain Monte Carlo.:boom: [2023-12-19] Added the vignette “Properties of the reconciled distribution via conditioning”.
:boom: [2023-08-23] Added the vignette “Probabilistic Reconciliation via Conditioning with bayesRecon”. Added the schaferStrimmer_cov
function.
:boom: [2023-05-26] bayesRecon v0.1.0 is released!
You can install the stable version on R CRAN
You can also install the development version from Github
# install.packages("devtools")
devtools::install_github("IDSIA/bayesRecon", build_vignettes = TRUE, dependencies = TRUE)
Let us consider the minimal temporal hierarchy in the figure, where the bottom variables are the two 6-monthly forecasts and the upper variable is the yearly forecast. We denote the variables for the two semesters and the year by \(S_1, S_2, Y\) respectively.
The hierarchy is described by the aggregating matrix S, which can be obtained using the function get_reconc_matrices
.
library(bayesRecon)
rec_mat <- get_reconc_matrices(agg_levels = c(1, 2), h = 2)
S <- rec_mat$S
print(S)
#> [,1] [,2]
#> [1,] 1 1
#> [2,] 1 0
#> [3,] 0 1
We assume that the base forecasts are Poisson distributed, with parameters given by \(\lambda_{Y} = 9\), \(\lambda_{S_1} = 2\), and \(\lambda_{S_2} = 4\).
lambdaS1 <- 2
lambdaS2 <- 4
lambdaY <- 9
lambdas <- c(lambdaY, lambdaS1, lambdaS2)
base_forecasts = list()
for (i in 1:nrow(S)) {
base_forecasts[[i]] = lambdas[i]
}
We recommend using the BUIS algorithm (Zambon et al., 2022) to sample from the reconciled distribution.
buis <- reconc_BUIS(
S,
base_forecasts,
in_type = "params",
distr = "poisson",
num_samples = 100000,
seed = 42
)
samples_buis <- buis$reconciled_samples
Since there is a positive incoherence in the forecasts (\(\lambda_Y > \lambda_{S_1}+\lambda_{S_2}\)), the mean of the bottom reconciled forecast increases. We show below this behavior for \(S_1\).
reconciled_forecast_S1 <- buis$bottom_reconciled_samples[1,]
range_forecats <- range(reconciled_forecast_S1)
hist(
reconciled_forecast_S1,
breaks = seq(range_forecats[1] - 0.5, range_forecats[2] + 0.5),
freq = F,
xlab = "S_1",
ylab = NULL,
main = "base vs reconciled"
)
points(
seq(range_forecats[1], range_forecats[2]),
stats::dpois(seq(range_forecats[1], range_forecats[2]), lambda =
lambdaS1),
pch = 16,
col = 4,
cex = 2
)
The blue circles represent the probability mass function of a Poisson with parameter \(\lambda_{S_1}\) plotted on top of the histogram of the reconciled bottom forecasts for \(S_1\). Note how the histogram is shifted to the right.
Moreover, while the base bottom forecast were assumed independent, the operation of reconciliation introduced a negative correlation between \(S_1\) and \(S_2\). We can visualize it with the plot below which shows the empirical correlations between the reconciled samples of \(S_1\) and the reconciled samples of \(S_2\).
AA <-
xyTable(buis$bottom_reconciled_samples[1, ],
buis$bottom_reconciled_samples[2, ])
plot(
AA$x ,
AA$y ,
cex = AA$number * 0.001 ,
pch = 16 ,
col = rgb(0, 0, 1, 0.4) ,
xlab = "S_1" ,
ylab = "S_2" ,
xlim = range(buis$bottom_reconciled_samples[1, ]) ,
ylim = range(buis$bottom_reconciled_samples[2, ])
)
We also provide a function for sampling using Markov Chain Monte Carlo (Corani et al., 2022).
mcmc = reconc_MCMC(
S,
base_forecasts,
distr = "poisson",
num_samples = 30000,
seed = 42
)
samples_mcmc <- mcmc$reconciled_samples
We now assume that the base forecasts are Gaussian distributed, with parameters given by
muS1 <- 2
muS2 <- 4
muY <- 9
mus <- c(muY, muS1, muS2)
sigmaS1 <- 2
sigmaS2 <- 2
sigmaY <- 3
sigmas <- c(sigmaY, sigmaS1, sigmaS2)
base_forecasts = list()
for (i in 1:nrow(S)) {
base_forecasts[[i]] = c(mus[[i]], sigmas[[i]])
}
We use the BUIS algorithm to sample from the reconciled distribution:
buis <- reconc_BUIS(
S,
base_forecasts,
in_type = "params",
distr = "gaussian",
num_samples = 100000,
seed = 42
)
samples_buis <- buis$reconciled_samples
buis_means <- rowMeans(samples_buis)
In the base forecasts are Gaussian, the reconciled distribution is still Gaussian and can be computed in closed form:
Sigma <- diag(sigmas ^ 2) #transform into covariance matrix
analytic_rec <- reconc_gaussian(S,
base_forecasts.mu = mus,
base_forecasts.Sigma = Sigma)
analytic_means_bottom <- analytic_rec$bottom_reconciled_mean
analytic_means <- S %*% analytic_means_bottom
The base means of \(Y\), \(S_1\), and \(S_2\) are 9, 2, 4.
The reconciled means obtained analytically are 7.41, 2.71, 4.71, while the reconciled means obtained via BUIS are 7.41, 2.71, 4.71.
Corani, G., Azzimonti, D., Augusto, J.P.S.C., Zaffalon, M. (2021). Probabilistic Reconciliation of Hierarchical Forecast via Bayes’ Rule. In: Hutter, F., Kersting, K., Lijffijt, J., Valera, I. (eds) Machine Learning and Knowledge Discovery in Databases. ECML PKDD 2020. Lecture Notes in Computer Science(), vol 12459. Springer, Cham. DOI
Corani, G., Azzimonti, D., Rubattu, N. (2023). Probabilistic reconciliation of count time series. DOI
Zambon, L., Azzimonti, D. & Corani, G. (2024). Efficient probabilistic reconciliation of forecasts for real-valued and count time series. DOI
Zambon, L., Agosto, A., Giudici, P., Corani, G. (2023). Properties of the reconciled distributions for Gaussian and count forecasts. DOI
Dario Azzimonti (Maintainer) dario.azzimonti@gmail.com |
Nicolò Rubattu nicolo.rubattu@idsia.ch |
Lorenzo Zambon lorenzo.zambon@idsia.ch |
Giorgio Corani giorgio.corani@idsia.ch |
If you encounter a clear bug, please file a minimal reproducible example on GitHub.