---
title: "Get started"
output: rmarkdown::html_vignette
vignette: >
%\VignetteIndexEntry{Get started}
%\VignetteEngine{knitr::rmarkdown}
%\VignetteEncoding{UTF-8}
---
```{r, include = FALSE}
knitr::opts_chunk$set(collapse = TRUE, comment = "#>", fig.width = 7, fig.height = 4)
```
A model that fits the past well can still forecast badly. `foresightr` chooses
models the other way round: it replays the past, and at each point forecasts
with what was known then.
```{r}
library(foresightr)
```
## The data
Monthly ICMS, the main tax revenue of the Brazilian state of Piauí, from the
public fiscal reports (Siconfi/STN). The first observation is March 2017.
```{r}
path <- system.file("extdata", "piaui_revenue.csv", package = "foresightr")
data <- read.csv(path, comment.char = "#")
# BRL million, monthly from March 2017
icms <- ts(data$icms / 1e6, start = c(2017, 3), frequency = 12)
```
## Choose a model by what would have worked
`backtest()` refits every candidate at each of the last 36 months, using only
the data before it, and forecasts up to 12 months ahead. The candidates are
ranked by their mean absolute percentage error, the average of the best two is
added as one more candidate, and the best of all forecasts from the whole
series.
```{r}
bt <- backtest(icms)
bt
```
The forecast of the chosen candidate, with intervals that are quantiles of
the errors it made in the backtest, horizon by horizon:
```{r}
head(bt$forecast)
autoplot(bt)
```
The error of each candidate by horizon:
```{r}
head(bt$candidates[[bt$best]]$accuracy)
```
## Totals
The total of the next six months has its own interval, measured on totals in
the backtest. Adding up the monthly limits would give a wider and wrong
interval, because the errors of successive months partly cancel.
```{r}
total_forecast(bt, 6)
```
## One model on its own
```{r}
fit <- fit_model(model_auto_arima(), log(icms))
fit$details$order
fit$details$seasonal_order
exp(predict(fit, 6))
```
## Your own candidates
Any list of models works, renamed as you like; ensembles learn their weights
again at every origin of the backtest.
```{r}
mine <- list(
with_name(model_seasonal_naive(), "same_month"),
model_log(model_airline()),
model_ensemble(candidates_default(), weighting = "stacked")
)
backtest(icms, mine, origins = 24, horizon = 6)$ranking[, c("name", "score")]
```
## Decomposition
```{r}
d <- decompose_stl(log(icms), seasonal_window = 13)
d
autoplot(d)
```