Extending mlr3 to time series forecasting.
This package is in an early stage of development and should be considered experimental. If you are interested in experimenting with it, we welcome your feedback!
Install the development version from GitHub:
# install.packages("pak")
pak::pak("mlr-org/mlr3forecast")The goal of mlr3forecast is to extend mlr3 to time series forecasting. This is achieved by introducing new classes and methods for forecasting tasks, learners, and resamplers. For now the forecasting task and learner is restricted to time series regression tasks, but might be extended to classification tasks in the future.
We have two goals, one to support traditional forecasting learners and the other to support machine learning forecasting, i.e. using regression learners and applying them to forecasting tasks. The design of the latter is still in flux and may change.
Currently, we support native forecasting learners from the forecast package. In the future, we plan to support more forecasting learners.
library(mlr3forecast)
task = tsk("airpassengers")
task
#> <TaskFcst:airpassengers> (144 x 1): Monthly Airline Passenger Numbers 1949-1960
#> * Target: passengers
#> * Properties: ordered
#> * Order by: date
#> * Frequency: monthly
# or plot the task
autoplot(task)
learner = lrn("fcst.auto_arima")$train(task)
prediction = learner$predict(task, 140:144)
prediction$score(msr("regr.rmse"))
#> regr.rmse
#> 13.85493
# generate new data to forecast unseen data
newdata = generate_newdata(task, 12L)
newdata
#> date passengers
#> 1: 1961-01-01 NA
#> 2: 1961-02-01 NA
#> 3: 1961-03-01 NA
#> 4: 1961-04-01 NA
#> 5: 1961-05-01 NA
#> 6: 1961-06-01 NA
#> 7: 1961-07-01 NA
#> 8: 1961-08-01 NA
#> 9: 1961-09-01 NA
#> 10: 1961-10-01 NA
#> 11: 1961-11-01 NA
#> 12: 1961-12-01 NA
learner$predict_newdata(newdata, task)
#> <PredictionRegr> for 12 observations:
#> row_ids truth response
#> 1 NA 445.6349
#> 2 NA 420.3950
#> 3 NA 449.1983
#> --- --- ---
#> 10 NA 494.1266
#> 11 NA 423.3327
#> 12 NA 465.5075
# works with quantile response
learner = lrn("fcst.auto_arima",
predict_type = "quantiles",
quantiles = c(0.1, 0.15, 0.5, 0.85, 0.9),
quantile_response = 0.5
)$train(task)
learner$predict_newdata(newdata, task)
#> <PredictionRegr> for 12 observations:
#> row_ids truth q0.1 q0.15 q0.5 q0.85 q0.9 response
#> 1 NA 430.8903 433.7105 445.6349 457.5593 460.3794 445.6349
#> 2 NA 403.0907 406.4004 420.3950 434.3895 437.6993 420.3950
#> 3 NA 429.7726 433.4880 449.1983 464.9085 468.6240 449.1983
#> --- --- --- --- --- --- --- ---
#> 10 NA 469.8624 474.5033 494.1266 513.7498 518.3908 494.1266
#> 11 NA 398.8381 403.5231 423.3327 443.1422 447.8272 423.3327
#> 12 NA 440.8228 445.5442 465.5075 485.4709 490.1922 465.5075library(mlr3learners)
task = tsk("airpassengers")
learner = lrn("regr.ranger")
flrn = ForecastLearner$new(learner, lags = 1:12)$train(task)
newdata = generate_newdata(task, 3L)
prediction = flrn$predict_newdata(newdata, task)
prediction
#> <PredictionRegr> for 3 observations:
#> row_ids truth response
#> 1 NA 436.4729
#> 2 NA 437.5867
#> 3 NA 455.3640
prediction = flrn$predict(task, 142:144)
prediction
#> <PredictionRegr> for 3 observations:
#> row_ids truth response
#> 1 461 455.8082
#> 2 390 410.3209
#> 3 432 433.2186
prediction$score(msr("regr.rmse"))
#> regr.rmse
#> 12.12956
flrn = ForecastLearner$new(learner, lags = 1:12)
resampling = rsmp("forecast_holdout", ratio = 0.9)
rr = resample(task, flrn, resampling)
rr$aggregate(msr("regr.rmse"))
#> regr.rmse
#> 48.11858
resampling = rsmp("forecast_cv")
rr = resample(task, flrn, resampling)
rr$aggregate(msr("regr.rmse"))
#> regr.rmse
#> 26.44163Or with some feature engineering using mlr3pipelines:
library(mlr3pipelines)
graph = ppl("convert_types", "Date", "POSIXct") %>>%
po("datefeatures",
param_vals = list(
week_of_year = FALSE,
day_of_year = FALSE,
day_of_month = FALSE,
day_of_week = FALSE,
is_day = FALSE,
hour = FALSE,
minute = FALSE,
second = FALSE
)
)
task = tsk("airpassengers")
task$set_col_roles("date", add = "feature")
flrn = ForecastLearner$new(lrn("regr.ranger"), lags = 1:12)
glrn = as_learner(graph %>>% flrn)$train(task)
prediction = glrn$predict(task, 142:144)
prediction$score(msr("regr.rmse"))
#> regr.rmse
#> 13.68579library(mlr3learners)
library(mlr3pipelines)
task = tsk("electricity")
task$set_col_roles("date", add = "feature")
graph = ppl("convert_types", "Date", "POSIXct") %>>%
po("datefeatures",
param_vals = list(
year = FALSE,
is_day = FALSE,
hour = FALSE,
minute = FALSE,
second = FALSE
)
)
flrn = ForecastLearner$new(lrn("regr.ranger"), 1:3)
glrn = as_learner(graph %>>% flrn)$train(task)
max_date = task$data()[.N, date]
newdata = data.frame(
date = max_date + 1:14,
demand = rep(NA_real_, 14L),
temperature = 26,
holiday = c(TRUE, rep(FALSE, 13L))
)
prediction = glrn$predict_newdata(newdata, task)
prediction
#> <PredictionRegr> for 14 observations:
#> row_ids truth response
#> 1 NA 187660.9
#> 2 NA 197831.7
#> 3 NA 188001.2
#> --- --- ---
#> 12 NA 221199.3
#> 13 NA 224450.2
#> 14 NA 225969.8library(mlr3learners)
library(mlr3pipelines)
library(tsibble)
task = tsibbledata::aus_livestock |>
as.data.table() |>
setnames(tolower) |>
_[, month := as.Date(month)] |>
_[, .(count = sum(count)), by = .(state, month)] |>
setorder(state, month) |>
as_task_fcst(
id = "aus_livestock",
target = "count",
order = "month",
key = "state",
freq = "monthly"
)
task$set_col_roles("month", add = "feature")
graph = ppl("convert_types", "Date", "POSIXct") %>>%
po("datefeatures",
param_vals = list(
week_of_year = FALSE,
day_of_week = FALSE,
day_of_month = FALSE,
day_of_year = FALSE,
is_day = FALSE,
hour = FALSE,
minute = FALSE,
second = FALSE
)
)
task = graph$train(task)[[1L]]
flrn = ForecastLearner$new(lrn("regr.ranger"), 1:3)$train(task)
prediction = flrn$predict(task, 4460:4464)
prediction$score(msr("regr.rmse"))
#> regr.rmse
#> 24904.97
flrn = ForecastLearner$new(lrn("regr.ranger"), 1:3)
resampling = rsmp("forecast_holdout", ratio = 0.9)
rr = resample(task, flrn, resampling)
rr$aggregate(msr("regr.rmse"))
#> regr.rmse
#> 91315.86In machine learning forecasting the difference between forecasting a time series and longitudinal data is often refered to local and global forecasting.
# TODO: find better task example, since the effect is minor here
graph = ppl("convert_types", "Date", "POSIXct") %>>%
po("datefeatures",
param_vals = list(
week_of_year = FALSE,
day_of_week = FALSE,
day_of_month = FALSE,
day_of_year = FALSE,
is_day = FALSE,
hour = FALSE,
minute = FALSE,
second = FALSE
)
)
# local forecasting
task = tsibbledata::aus_livestock |>
as.data.table() |>
setnames(tolower) |>
_[, month := as.Date(month)] |>
_[state == "Western Australia", .(count = sum(count)), by = .(month)] |>
setorder(month) |>
as_task_fcst(id = "aus_livestock", target = "count", order = "month")
task = graph$train(task)[[1L]]
flrn = ForecastLearner$new(lrn("regr.ranger"), 1L)$train(task)
tab = task$backend$data(
rows = task$row_ids,
cols = c(task$backend$primary_key, "month.year")
)
setnames(tab, c("row_id", "year"))
row_ids = tab[year >= 2015, row_id]
prediction = flrn$predict(task, row_ids)
prediction$score(msr("regr.rmse"))
# global forecasting
task = tsibbledata::aus_livestock |>
as.data.table() |>
setnames(tolower) |>
_[, month := as.Date(month)] |>
_[, .(count = sum(count)), by = .(state, month)] |>
setorder(state, month) |>
as_task_fcst(id = "aus_livestock", target = "count", order = "month", key = "state")
task = graph$train(task)[[1L]]
task$col_roles$key = "state"
flrn = ForecastLearner$new(lrn("regr.ranger"), 1L)$train(task)
tab = task$backend$data(
rows = task$row_ids,
cols = c(task$backend$primary_key, "month.year", "state")
)
setnames(tab, c("row_id", "year", "state"))
row_ids = tab[year >= 2015 & state == "Western Australia", row_id]
prediction = flrn$predict(task, row_ids)
prediction$score(msr("regr.rmse"))library(mlr3learners)
library(mlr3pipelines)
task = tsk("airpassengers")
pop = po("fcst.lag", lags = 1:12)
new_task = pop$train(list(task))[[1L]]
new_task$data()
task = tsk("airpassengers")
graph = po("fcst.lag", lags = 1:12) %>>%
ppl("convert_types", "Date", "POSIXct") %>>%
po("datefeatures",
param_vals = list(
week_of_year = FALSE,
day_of_week = FALSE,
day_of_month = FALSE,
day_of_year = FALSE,
is_day = FALSE,
hour = FALSE,
minute = FALSE,
second = FALSE
)
)
flrn = ForecastLearner2$new(lrn("regr.ranger"))
glrn = as_learner(graph %>>% flrn)$train(task)
prediction = glrn$predict(task, 142:144)
prediction$score(msr("regr.rmse"))
newdata = generate_newdata(task, 12L)
glrn$predict_newdata(newdata, task)Some common target transformations in forecasting are:
- differencing (WIP)
- log transformation, see example below
- power transformations such as Box-Cox and Yeo-Johnson currently only supported as feature transformation and not target
- scaling/normalization, available see here
trafo = po("targetmutate",
param_vals = list(
trafo = function(x) log(x),
inverter = function(x) list(response = exp(x$response))
)
)
graph = po("fcst.lag", lags = 1:12) %>>%
ppl("convert_types", "Date", "POSIXct") %>>%
po("datefeatures",
param_vals = list(
week_of_year = FALSE,
day_of_week = FALSE,
day_of_month = FALSE,
day_of_year = FALSE,
is_day = FALSE,
hour = FALSE,
minute = FALSE,
second = FALSE
)
)
task = tsk("airpassengers")
flrn = ForecastLearner2$new(lrn("regr.ranger"))
glrn = as_learner(graph %>>% flrn)
pipeline = ppl("targettrafo", graph = glrn, trafo_pipeop = trafo)
glrn = as_learner(pipeline)$train(task)
prediction = glrn$predict(task, 142:144)
prediction$score(msr("regr.rmse"))graph = po("fcst.lag", lags = 1:12) %>>%
ppl("convert_types", "Date", "POSIXct") %>>%
po("datefeatures",
param_vals = list(
week_of_year = FALSE,
day_of_week = FALSE,
day_of_month = FALSE,
day_of_year = FALSE,
is_day = FALSE,
hour = FALSE,
minute = FALSE,
second = FALSE
)
)
task = tsk("airpassengers")
flrn = ForecastLearner2$new(lrn("regr.ranger"))
glrn = as_learner(graph %>>% flrn)
trafo = po("fcst.targetdiff", lags = 12L)
pipeline = ppl("targettrafo", graph = glrn, trafo_pipeop = trafo)
glrn = as_learner(pipeline)$train(task)
prediction = glrn$predict(task, 142:144)
prediction$score(msr("regr.rmse"))