Wouldn't the predict function from the bsts package return the same values as the state contributions for missing y values?

[maferre1ra]

I am comparing the estimates obtained from running a bsts model on a variable with the last 4 observations missing (held out) to what we can obtain from the predict function using the same model (horizon = 4), and the two set of estimates do not match.

Could anyone help shed some light?

Here is a reproducible example.

############################ library(bsts) library(CausalImpact) library(tidyverse) library(lubridate)

n = 33 # observations

GENERATE SOME DATA

y2 = rnorm(n, 1542429, 922877.8) post.period <- c(30, 33) # hold out 4 observations post.period.response <- y2[post.period[1] : post.period[2]] y2[post.period[1] : post.period[2]] <- NA

################

RUN A BSTS MODEL

niter = 15000 prior.level.sd = 0.1 sample.size = nr/10

sdy <- sd(y2, na.rm = TRUE) ss <- list() sd.prior <- SdPrior(sigma.guess = prior.level.sd * sdy, upper.limit = sdy, sample.size = sample.size) ss <- AddSeasonal(ss, y, nseasons = 12, season.duration = 1)

p = 1 ss2 <- bsts::AddAr(ss, y2, lags = p, sigma.prior = sd.prior)

local2 <- bsts(y2, state.specification = ss2, niter = niter, seed = 1, ping = 0, model.options = BstsOptions(save.prediction.errors = TRUE))

############################

GET STATE CONTRIBUTIONS

bsts.model = local2 burn <- SuggestBurn(0.1, bsts.model) state.contributions <- bsts.model$state.contributions[-seq_len(burn), , , drop = FALSE] state.samples <- rowSums(aperm(state.contributions, c(1, 3, 2)), dims = 2) point.pred.mean <- colMeans(state.samples)

##############

GET PREDITION

bsts_predict <- predict.bsts(local2, horizon = 4)

From the example above, the point predictions from state contributions do not match the 4 point predictions from the predict function.