vignettes/long-horizon.Rmd

---
title: "Long-Horizon Forecasting"
output: 
  rmarkdown::html_vignette:
    toc: true 
    toc_depth: 2
vignette: >
  %\VignetteIndexEntry{Long-Horizon Forecasting}
  %\VignetteEngine{knitr::rmarkdown}
  %\VignetteEncoding{UTF-8}
---
```{r setup, include=FALSE}
library(httptest2)
.mockPaths("../tests/mocks")
start_vignette(dir = "../tests/mocks")

original_options <- options("NIXTLA_API_KEY"="dummy_api_key", digits=7)

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>", 
  fig.width = 7, 
  fig.height = 4
)
```

```{r}
library(nixtlar)
```

## 1. Long-horizon forecasting 

In some cases, it is necessary to forecast long horizons. Here "long horizons" refer to predictions that exceed more than two seasonal periods. For example, this would mean forecasting more than 48 hours ahead for hourly data or more than 7 days for daily data. The specific definition of "long horizon" varies depending on the data frequency.

There is a specialized `TimeGPT` model designed for long-horizon forecasting, which is trained to predict far into the future, where the uncertainty increases as the forecast extends further. Here we will explain how to use the long horizon model of `TimeGPT` via `nixtlar`. 

This vignette assumes you have already set up your API key. If you haven't done this, please read the [Get Started](https://nixtla.github.io/nixtlar/articles/get-started.html) vignette first. 

## 2. Load data 

For this vignette, we'll use the electricity consumption dataset that is included in `nixtlar`, which contains the hourly prices of five different electricity markets. 

```{r}
df <- nixtlar::electricity
head(df)
```

For every `unique_id`, we'll try to predict the last 96 hours. Hence, we will first separate the data into training and test sets. 

```{r}
test <- df |> 
  dplyr::group_by(unique_id) |> 
  dplyr::slice_tail(n = 96) |> 
  dplyr::ungroup() 

train <- df[df$ds %in% setdiff(df$ds, test$ds), ]
```

## 3. Forecast with a long-horizon 

To use the long-horizon model of `TimeGPT`, set the `model` argument to `timegpt-1-long-horizon`. 

```{r}
fcst_long_horizon <- nixtlar::nixtla_client_forecast(train, h=96, model="timegpt-1-long-horizon")
head(fcst_long_horizon)
```

## 4. Plot the long-horizon forecast

`nixtlar` includes a function to plot the historical data and any output from `nixtlar::nixtla_client_forecast`, `nixtlar::nixtla_client_historic`, `nixtlar::nixtla_client_detect_anomalies` and `nixtlar::nixtla_client_cross_validation`. If you have long series, you can use `max_insample_length` to only plot the last N historical values (the forecast will always be plotted in full). 

```{r}
nixtlar::nixtla_client_plot(train, fcst_long_horizon, max_insample_length = 200)
```

## 5. Evaluate the long-horizon model 

To evaluate the long-horizon forecast, we will generate the same forecast with the default model of `TimeGPT`, which is `timegpt-1`, and then we will compute and compare the [Mean Absolute Error](https://en.wikipedia.org/wiki/Mean_absolute_error) (MAE) of the two models. 

```{r}
fcst <- nixtlar::nixtla_client_forecast(train, h=96)
head(fcst)
```

We will rename the `TimeGPT` long-horizon model to merge it with the default `TimeGPT` model. After that, we will combine them with the actual values from the test set and compute the MAE. Note that in the output of the `nixtla_client_forecast` function, the `ds` column contains dates. This is because `nixtla_client_plot` uses the dates for plotting. However, to merge the actual values, we will convert the dates to characters. 

```{r}
names(fcst_long_horizon)[which(names(fcst_long_horizon) == "TimeGPT")] <- "TimeGPT-long-horizon"

res <- merge(fcst, fcst_long_horizon) # merge TimeGPT and TimeGPT-long-horizon
res$ds <- as.character(res$ds)

res <- merge(test, res) # merge with actual values
head(res)
```

```{r}
print(paste0("MAE TimeGPT: ", mean(abs(res$y-res$TimeGPT))))
print(paste0("MAE TimeGPT long-horizon: ", mean(abs(res$y-res$`TimeGPT-long-horizon`))))
```

As we can see, the long-horizon version of `TimeGPT` produced a model with a lower MAE than the default `TimeGPT` model. 