Categorical variables

Categorical variables are external factors that can influence a forecast. These variables take on one of a limited, fixed number of possible values, and induce a grouping of your observations.

For example, if you’re forecasting daily product demand for a retailer, you could benefit from an event variable that may tell you what kind of event takes place on a given day, for example ‘None’, ‘Sporting’, or ‘Cultural’.

To incorporate categorical variables in TimeGPT, you’ll need to pair each point in your time series data with the corresponding external data.

1. Import packages

First, we install and import the required packages and initialize the Nixtla client.

import pandas as pd
import os

from nixtla import NixtlaClient
from datasetsforecast.m5 import M5

nixtla_client = NixtlaClient(
    # defaults to os.environ.get("NIXTLA_API_KEY")
    api_key = 'my_api_key_provided_by_nixtla'   
)

👍

Use an Azure AI endpoint

To use an Azure AI endpoint, remember to set also the base_url argument:

nixtla_client = NixtlaClient(base_url="you azure ai endpoint", api_key="your api_key")

2. Load M5 data

Let’s see an example on predicting sales of products of the M5 dataset. The M5 dataset contains daily product demand (sales) for 10 retail stores in the US.

First, we load the data using datasetsforecast. This returns:

• Y_df, containing the sales (y column), for each unique product (unique_id column) at every timestamp (ds column).
• X_df, containing additional relevant information for each unique product (unique_id column) at every timestamp (ds column).

Y_df, X_df, _ = M5.load(directory=os.getcwd())
Y_df['ds'] = pd.to_datetime(Y_df['ds'])
X_df['ds'] = pd.to_datetime(X_df['ds'])
Y_df.head(10)

For this example, we will only keep the additional relevant information from the column event_type_1. This column is a categorical variable that indicates whether an important event that might affect the sales of the product takes place at a certain date.

X_df = X_df[['unique_id', 'ds', 'event_type_1']]

X_df.head(10)

As you can see, on February 6th 2011, there is a Sporting event.

3. Forecasting product demand using categorical variables

We will forecast the demand for a single product only. We choose a high selling food product identified by FOODS_3_090_CA_3.

product = 'FOODS_3_090_CA_3'
Y_df_product = Y_df.query('unique_id == @product')
X_df_product = X_df.query('unique_id == @product')

We merge our two dataframes to create the dataset to be used in TimeGPT.

df = Y_df_product.merge(X_df_product)

df.head(10)

In order to use categorical variables with TimeGPT, it is necessary to numerically encode the variables. We will use one-hot encoding in this tutorial.

We can one-hot encode the event_type_1 column by using pandas built-in get_dummies functionality. After one-hot encoding the event_type_1 variable, we can add it to the dataframe and remove the original column.

event_type_1_ohe = pd.get_dummies(df['event_type_1'], dtype=int)
df = pd.concat([df, event_type_1_ohe], axis=1)
df = df.drop(columns = 'event_type_1')

df.tail(10)

As you can see, we have now added 5 columns, each with a binary indicator (1 or 0) whether there is a Cultural, National, Religious, Sporting or no (nan) event on that particular day. For example, on June 19th 2016, there is a Sporting event.

Let’s turn to our forecasting task. We will forecast the first 7 days of February 2016. This includes 7 February 2016 - the date on which Super Bowl 50 was held. Such large, national events typically impact retail product sales.

To use the encoded categorical variables in TimeGPT, we have to add them as future values. Therefore, we create a future values dataframe, that contains the unique_id, the timestamp ds, and the encoded categorical variables.

Of course, we drop the target column as this is normally not available - this is the quantity that we seek to forecast!

future_ex_vars_df = df.drop(columns = ['y'])
future_ex_vars_df = future_ex_vars_df.query("ds >= '2016-02-01' & ds <= '2016-02-07'")

future_ex_vars_df.head(10)

Next, we limit our input dataframe to all but the 7 forecast days:

df_train = df.query("ds < '2016-02-01'")

df_train.tail(10)

Let’s call the forecast method, first without the categorical variables.

timegpt_fcst_without_cat_vars_df = nixtla_client.forecast(df=df_train, h=7, level=[80, 90])
timegpt_fcst_without_cat_vars_df.head()

INFO:nixtla.nixtla_client:Validating inputs...
INFO:nixtla.nixtla_client:Preprocessing dataframes...
INFO:nixtla.nixtla_client:Inferred freq: D
INFO:nixtla.nixtla_client:Restricting input...
INFO:nixtla.nixtla_client:Calling Forecast Endpoint...

📘

Available models in Azure AI

If you are using an Azure AI endpoint, please be sure to set model="azureai":

nixtla_client.forecast(..., model="azureai")

For the public API, we support two models: timegpt-1 and timegpt-1-long-horizon.

By default, timegpt-1 is used. Please see this tutorial on how and when to use timegpt-1-long-horizon.

We plot the forecast and the last 28 days before the forecast period:

nixtla_client.plot(
    df[['unique_id', 'ds', 'y']].query("ds <= '2016-02-07'"), 
    timegpt_fcst_without_cat_vars_df, 
    max_insample_length=28, 
)

TimeGPT already provides a reasonable forecast, but it seems to somewhat underforecast the peak on the 6th of February 2016 - the day before the Super Bowl.

Let’s call the forecast method again, now with the categorical variables.

timegpt_fcst_with_cat_vars_df = nixtla_client.forecast(df=df_train, X_df=future_ex_vars_df, h=7, level=[80, 90])
timegpt_fcst_with_cat_vars_df.head()

INFO:nixtla.nixtla_client:Validating inputs...
INFO:nixtla.nixtla_client:Preprocessing dataframes...
INFO:nixtla.nixtla_client:Inferred freq: D
INFO:nixtla.nixtla_client:Using the following exogenous variables: Cultural, National, Religious, Sporting, nan
INFO:nixtla.nixtla_client:Calling Forecast Endpoint...

📘

Available models in Azure AI

If you are using an Azure AI endpoint, please be sure to set model="azureai":

nixtla_client.forecast(..., model="azureai")

For the public API, we support two models: timegpt-1 and timegpt-1-long-horizon.

By default, timegpt-1 is used. Please see this tutorial on how and when to use timegpt-1-long-horizon.

We plot the forecast and the last 28 days before the forecast period:

nixtla_client.plot(
    df[['unique_id', 'ds', 'y']].query("ds <= '2016-02-07'"), 
    timegpt_fcst_with_cat_vars_df, 
    max_insample_length=28, 
)

We can visually verify that the forecast is closer to the actual observed value, which is the result of including the categorical variable in our forecast.

Let’s verify this conclusion by computing the Mean Absolute Error on the forecasts we created.

from utilsforecast.losses import mae

# Create target dataframe
df_target = df[['unique_id', 'ds', 'y']].query("ds >= '2016-02-01' & ds <= '2016-02-07'")

# Rename forecast columns
timegpt_fcst_without_cat_vars_df = timegpt_fcst_without_cat_vars_df.rename(columns={'TimeGPT': 'TimeGPT-without-cat-vars'})
timegpt_fcst_with_cat_vars_df = timegpt_fcst_with_cat_vars_df.rename(columns={'TimeGPT': 'TimeGPT-with-cat-vars'})

# Merge forecasts with target dataframe
df_target = df_target.merge(timegpt_fcst_without_cat_vars_df[['unique_id', 'ds', 'TimeGPT-without-cat-vars']])
df_target = df_target.merge(timegpt_fcst_with_cat_vars_df[['unique_id', 'ds', 'TimeGPT-with-cat-vars']])

# Compute errors
mean_absolute_errors = mae(df_target, ['TimeGPT-without-cat-vars', 'TimeGPT-with-cat-vars'])

mean_absolute_errors

Indeed, we find that the error when using TimeGPT with the categorical variable is approx. 20% lower than when using TimeGPT without the categorical variables, indicating better performance when we include the categorical variable.