Nothing
test_that("bias is estimated before the rows used to accept it", {
score <- expression(mean((predicted - actual)^2))
p <- rep(10, 8)
choose <- function(y, policy = "split", objective = "min", obj = score)
NNS:::.nns_arma_bias_choice(p, y, obj, objective, TRUE, policy)
persistent <- choose(rep(7, 8))
transient <- choose(c(rep(7, 4), rep(13, 4)))
expect_equal(persistent$estimate.rows, 1:4)
expect_equal(persistent$acceptance.rows, 5:8)
expect_equal(persistent$estimated.shift, -3)
expect_equal(transient$estimated.shift, -3)
expect_true(persistent$accepted)
expect_equal(persistent$applied.shift, -3)
expect_false(transient$accepted)
expect_equal(transient$applied.shift, 0)
expect_equal(choose(rep(7, 8), objective = "max",
obj = expression(-mean((predicted - actual)^2)))$applied.shift, -3)
expect_equal(choose(rep(7, 8), "none")$status, "disabled")
})
test_that("short validation windows do not claim separate bias validation", {
x <- NNS:::.nns_arma_bias_choice(4:6, 1:3, expression(mean((predicted - actual)^2)),
"min", TRUE, "split")
expect_equal(x$applied.shift, 0)
expect_equal(x$status, "insufficient_validation")
})
test_that("absolute seasonality scoring is invariant to interval-scale offsets", {
t <- seq_len(120)
x <- 15 + sin(2 * pi * t / 12) + .2 * cos(t)
a <- NNS.seas(x, plot = FALSE, dispersion = "absolute")
b <- NNS.seas(x + 273.15, plot = FALSE, dispersion = "absolute")
expect_equal(a$periods, b$periods)
expect_equal(a$all.periods, b$all.periods, tolerance = 1e-10)
})
test_that("Monte Carlo exponent and interval validation rejects invalid grids", {
for (e in list(0, -1, Inf, NA_real_, c(1, 2)))
expect_error(NNS.MC(1:10, exp = e), "exp must")
expect_error(NNS.MC(1:10, lower_rho = .8, upper_rho = .2), "lower_rho")
expect_error(NNS.MC(1:10, by = 0), "by must")
})
test_that("residual-scale expansion can be disabled independently", {
set.seed(5)
x <- rnorm(60)
run <- function(expand) {
set.seed(8)
NNS.meboot(x, reps = 8, rho = 0, drift = FALSE,
expand.sd = FALSE, force.clt = FALSE,
expand.residual.sd = expand)["replicates", ]$replicates
}
expect_gt(mean(apply(run(TRUE), 2, sd)), mean(apply(run(FALSE), 2, sd)))
})
test_that("VAR validation covariates cannot see observations after their origin", {
captured <- list()
testthat::local_mocked_bindings(
NNS.ARMA.optim = function(variable, h, ...) {
list(results = tail(variable, 1) + seq_len(h) / 10)
},
.nns_var_relevant = function(IVs.train, DV.train, ..., .validation = NULL) {
colnames(IVs.train)
},
NNS.stack = function(IVs.train, DV.train, IVs.test, ..., .validation = NULL) {
captured[[length(captured) + 1L]] <<- .validation
list(stack = rep(mean(DV.train), nrow(IVs.test)))
},
.package = "NNS"
)
t <- seq_len(40)
x <- cbind(a = sin(t / 5), b = cos(t / 7))
NNS.VAR(x, h = 4, tau = 2, ncores = 1, status = FALSE)
first <- captured
x[33:40, ] <- x[33:40, ] + 100
captured <- list()
NNS.VAR(x, h = 4, tau = 2, ncores = 1, status = FALSE)
expect_length(first, 2)
for (i in 1:2) {
expect_equal(captured[[i]]$IVs.train, first[[i]]$IVs.train)
expect_equal(captured[[i]]$IVs.test, first[[i]]$IVs.test)
expect_equal(captured[[i]]$DV.train, first[[i]]$DV.train)
expect_equal(captured[[i]]$DV.test - first[[i]]$DV.test, rep(100, 8))
}
# Later validation lags must also use forecasts, not realized heldout values.
expect_equal(first[[1]]$IVs.test$b_tau_0, x[32, "b"] + (1:8) / 10)
expect_equal(tail(first[[1]]$IVs.test$b_tau_2, 6), x[32, "b"] + (1:6) / 10)
})
test_that("VAR completes a mixed-frequency prefix before seeing later releases", {
captured <- list()
testthat::local_mocked_bindings(
NNS.ARMA.optim = function(variable, h, ...) {
list(results = tail(variable, 1) + seq_len(h) / 10)
},
# Missing releases are completed from the observed prefix alone.
.nns_var_index_fit = function(x, y, x.new, ...) rep(mean(y), length(x.new)),
NNS.reg = function(x, y, point.est, ...) {
list(Point.est = rep(mean(y), length(point.est)))
},
.nns_var_relevant = function(IVs.train, DV.train, ..., .validation = NULL) {
colnames(IVs.train)
},
NNS.stack = function(IVs.train, DV.train, IVs.test, ..., .validation = NULL) {
if (!is.null(.validation)) captured[[length(captured) + 1L]] <<- .validation
list(stack = rep(mean(DV.train), nrow(IVs.test)))
},
.package = "NNS"
)
t <- seq_len(40)
x <- cbind(a = 20 + sin(t / 5), b = 20 + cos(t / 7))
x[32, "b"] <- NA_real_ # trailing missing release at the validation origin
x[34, "a"] <- NA_real_
x[c(35, 38), "b"] <- NA_real_
out <- NNS.VAR(x, h = 4, tau = 2, ncores = 1, status = FALSE)
first <- captured
expect_equal(unname(out$validation$observed.targets), c(7L, 6L))
expect_length(first[[1]]$DV.test, 7L)
expect_length(first[[2]]$DV.test, 6L)
expect_equal(tail(first[[2]]$DV.train, 1), mean(x[1:31, "b"]))
x[33:40, ] <- x[33:40, ] + 100
captured <- list()
NNS.VAR(x, h = 4, tau = 2, ncores = 1, status = FALSE)
for (i in 1:2) {
expect_equal(captured[[i]]$IVs.train, first[[i]]$IVs.train)
expect_equal(captured[[i]]$DV.train, first[[i]]$DV.train)
expect_equal(captured[[i]]$IVs.test, first[[i]]$IVs.test)
}
})
test_that("VAR relevance coefficients never fit heldout rows in pipeline validation", {
trained_rows <- integer()
testthat::local_mocked_bindings(
NNS.reg = function(x, y, point.est = NULL, ...) {
trained_rows <<- c(trained_rows, length(y))
list(equation = data.frame(Coefficient = rep(1, ncol(x))),
Point.est = rep(mean(y), NROW(point.est)))
},
.package = "NNS"
)
t <- seq_len(40)
x <- data.frame(a = sin(t), b = cos(t), c = t)
y <- sin(t / 5)
fold <- list(IVs.train = x[1:32, ], DV.train = y[1:32],
IVs.test = x[33:40, ] * 0, DV.test = y[33:40])
NNS:::.nns_var_relevant(x, y, 8,
expression(mean((predicted - actual)^2)), "min", "cor", .validation = fold)
expect_true(length(trained_rows) > 0)
expect_true(all(trained_rows == 32L))
})
test_that("numeric two-level VAR targets remain forecast regressions", {
captured <- new.env()
captured$frames <- list()
suppressMessages(trace(".nns_var_relevant", where = asNamespace("NNS"),
print = FALSE, exit = bquote(assign("frames",
c(get("frames", envir = .(captured)), list(environment())),
envir = .(captured)))))
withr::defer(suppressMessages(untrace(".nns_var_relevant",
where = asNamespace("NNS"))))
local_mocked_bindings(
NNS.ARMA.optim = function(variable, h, ...) {
list(results = rep(mean(variable), h))
},
.package = "NNS"
)
prediction.reference <- list()
for (levels in list(c(0, 1), c(-5, 5))) for (method in c("cor", "NNS.dep")) {
captured$frames <- list()
t <- seq_len(40)
x <- cbind(a = rep(levels, 20), b = sin(t / 5))
out <- NNS.VAR(x, h = 2, tau = 1, ncores = 1, status = FALSE,
dim.red.method = method,
obj.fn = expression(mean((predicted - actual)^2)))
expect_identical(dim(out$ensemble), c(2L, 2L))
expect_true(all(is.finite(as.matrix(out$ensemble))))
expect_equal(unname(out$validation$observed.targets), c(8L, 8L))
# Relevance must extract real regression coefficients, not NULL from a
# list of nominal class equations and then silently retain every column.
fr <- captured$frames[[1L]]
expect_length(fr$coef, 3L)
expect_true(all(is.finite(fr$coef)))
if (method == "NNS.dep") {
expect_equal(fr$coef, c(0, 1, 0), tolerance = 1e-12)
expect_equal(na.omit(out$relevant_variables$a), "a_tau_1",
ignore_attr = TRUE)
# The first lag is observed at the origin; the seven later lags are
# prefix-mean forecasts. Their midpoint errors yield 7/8 * range^2/4.
expect_equal(fr$nns.ord, diff(levels)^2 * 7 / 32, tolerance = 1e-12)
expect_equal(out$multivariate$a, c(levels[1], mean(levels)),
tolerance = 1e-10)
} else {
expected <- apply(fr$CV.IVs.train, 2, stats::cor, y = fr$CV.DV.train,
method = "spearman")
expect_equal(fr$coef, unname(expected), tolerance = 1e-12)
expect_equal(na.omit(out$relevant_variables$a), c("b_tau_0", "a_tau_1"),
ignore_attr = TRUE)
expect_gt(fr$nns.ord / diff(levels)^2, .21)
expect_lt(fr$nns.ord / diff(levels)^2, .25)
expect_true(all(out$multivariate$a > min(levels) &
out$multivariate$a < max(levels)))
expect_lt(out$multivariate$a[1], mean(levels))
}
if (identical(levels, c(0, 1))) prediction.reference[[method]] <- out$multivariate$a else
expect_equal(out$multivariate$a,
levels[1] + diff(levels) * prediction.reference[[method]],
tolerance = 1e-10)
}
})
test_that("ARMA nonlinear fits and optimizer smoothing retain numeric responses", {
for (levels in list(c(0, 1), c(-5, 5))) {
y <- rep(levels, 15)
fit <- NNS.ARMA(y, h = 2, seasonal.factor = 3, method = "nonlin",
plot = FALSE, negative.values = TRUE)
expect_equal(fit, levels[1] + diff(levels) * c(10 / 9, -1 / 9),
tolerance = 1e-10)
optimized <- NNS.ARMA.optim(y, h = 2, seasonal.factor = 3,
negative.values = TRUE, ncores = 1, plot = FALSE, print.trace = FALSE,
obj.fn = expression(mean((predicted - actual)^2)), bias.correction = "none")
expect_equal(optimized$results,
levels[1] + diff(levels) * c(7 / 12, 5 / 12), tolerance = 1e-10)
}
})
Any scripts or data that you put into this service are public.
Add the following code to your website.
For more information on customizing the embed code, read Embedding Snippets.