# Forecast future values

With the **Forecast future values** task, you can predict the future values in a
time-series of numerical observations. For example, you can predict future sales
given the records of past ones.

Read the [Forecasting tutorial](tutorial_forecast) to see an example of
application of the forecasting task.

The forecast task expects values to be sorted by date and sampled at regular
intervals, for example, each day, each week, or each month. The **first column**
must contain dates. The next columns must contain one or more series of values.
The **first row** must contain the name of the series. All the series must have
the same number of values/rows. The *Forecast future values* task will predict
values in any row that has a date but doesn't have values.

Following is an example of a sheet compatible with the forecasting task. The
forecasting task will predict the values of series *A*, *B* and *C* for the
dates *2022-04-05* and *2022-04-06*.

Date       | A   | B   | C
---------- | --- | --- | ---
2022-04-01 | 112 | 125 | 150
2022-04-02 | 118 | 149 | 178
2022-04-03 | 132 | 170 | 163
2022-04-04 | 129 | 170 | 172
2022-04-05 |     |     |
2022-04-06 |     |     |

## Forecasting multiple series

By default, series (columns) are forecasted independently. Alternatively, it is
possible to forecast them together. Predicting series of a similar nature
together likely yields more accurate forecasts.

To forecast series together, provide one or more categorical values for each of
the series. Series with similar categorical values will be forecasted together
using a
[hierarchical reconciliation algorithm](https://otexts.com/fpp3/hierarchical.html).

Note: The reconciliation algorithm used is a
[top-down using forecast proportions](https://otexts.com/fpp3/single-level.html#forecast-proportions).

The next example shows the sales of 5 different products. Under the product
name, the type, and the category of each product are indicated in the first two
rows.

Date       | Lemonade | Orange Juice | White Wine | Analog watch | Digital watch
---------- | -------- | ------------ | ---------- | ------------ | -------------
Category   | drink    | drink        | drink      | wearable     | wearable
Type       | soda     | soda         | wine       | watch        | watch
2022-04-01 | 112      | 125          | 243        | 150          | 211
2022-04-02 | 118      | 149          | 264        | 178          | 180
2022-04-03 | 132      | 170          | 272        | 163          | 201
2022-04-04 | 129      | 170          | 237        | 172          | 204
2022-04-05 |          |              |            |              |
2022-04-06 |          |              |            |              |

Here, *Lemonade* and *Orange Juice* are both *soda* and *drink*. Both products
are serving the same purpose (those are called *substitution products*) and
probably have similar sales trends (e.g., increase sales in summer). *White
Wine* is also a *drink*, but it is not a soda.

## How are values forecasted

The algorithm used is a tuned model with cleaning, holiday, seasonality and
trend structure. Look at the *Advanced* option section for the *Forecast future
values* task. Details of the algorithms are described in
[Forecasting: Principles and Practice](https://otexts.com/fpp3/index.html) by
Hyndman et al.