We could modify the code above to be a loop that uses the model to make forecasts and uses those forecasts as inputs to make more forecasts:
new_data <- rbind(
tail(r_enwiki, 22),
data.frame(
date = seq(
as.Date("2017-05-17"),
as.Date("2017-05-17") + 90,
"day"
),
pageviews = NA,
normalized = NA
)
); rownames(new_data) <- NULL
for (d in 23:nrow(new_data)) {
new_mlts <- mlts_transform(
new_data[(d - 22):d, ],
date, normalized, p = 21,
extras = TRUE, extrasAsFactors = TRUE,
granularity = "day")
new_mlts <- cbind(
new_mlts[-1, ], # don't need to forecast known outcome
dummy(
new_mlts[, c("mlts_extras_weekday", "mlts_extras_month", "mlts_extras_monthday", "mlts_extras_week"),
drop = FALSE],
object = mlts_categories, int = TRUE
)[-1, ]
)
new_data$normalized[d] <- as.numeric(neuralnet::compute(
nn_model, new_mlts[, nn_features])$net.result
)
}
new_data$pageviews <- (new_data$normalized * normalization_constants$std.dev) + normalization_constants$mean
ggplot(dplyr::filter(r_enwiki, date >= "2016-10-01"),
aes(x = date, y = pageviews)) +
geom_line() +
geom_line(aes(y = pageviews), color = "red",
data = dplyr::filter(new_data, date >= "2017-05-16")) +
theme_minimal() +
labs(x = "Date", y = "Pageviews",
title = "90 day forecast of <https://en.wikipedia.org/wiki/R_(programming_language)> pageviews",
subtitle = "Red is for predictions made with a neural network with 3 layers containing 9, 7, and 5 hidden neurons, respectively")
