Multiple Series

TimeGPT provides a robust solution for multi-series forecasting, which involves analyzing multiple data series concurrently, rather than a single one. The tool can be fine-tuned using a broad collection of series, enabling you to tailor the model to suit your specific needs or tasks. Note that the forecasts are still univariate. This means that although TimeGPT is a global model, it won’t consider the inter-feature relationships within the target series. However, TimeGPT does support the use of exogenous variables such as categorical variables (e.g., category, brand), numerical variables (e.g., temperature, prices), or even special holidays. Let’s see this in action.

As always, we start off by intializing an instance of NixtlaClient.

import pandas as pd
from nixtla import NixtlaClient

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

Load the data

The following dataset contains prices of different electricity markets in Europe. Mutliple series are automatically detected in TimeGPT using the unique_id column. This column contains labels for each series. If there are multiple unique values in that column, then it knows it is handling a multi-series scneario. In this particular case, the unique_id column contains the value BE, DE, FR, JPM, and NP.

df = pd.read_csv('https://raw.githubusercontent.com/Nixtla/transfer-learning-time-series/main/datasets/electricity-short.csv')
df.head()

nixtla_client.plot(df)

Forecasting

To forecast all series at once, we simply pass the dataframe to the df argument. TimeGPt will automatically forecast all series.

timegpt_fcst_multiseries_df = nixtla_client.forecast(df=df, h=24, level=[80, 90])
timegpt_fcst_multiseries_df.head()

INFO:nixtlats.nixtla_client:Validating inputs...
INFO:nixtlats.nixtla_client:Preprocessing dataframes...
INFO:nixtlats.nixtla_client:Inferred freq: H
INFO:nixtlats.nixtla_client:Restricting input...
INFO:nixtlats.nixtla_client:Calling Forecast Endpoint...

nixtla_client.plot(df, timegpt_fcst_multiseries_df, max_insample_length=365, level=[80, 90])

From the figure above, we can see that the model effectively generated predictions for each unique series in the dataset.

Historical forecast

You can also compute prediction intervals for historical forecasts adding the add_history=True. To specify the confidence interval, we use the level argument. Here, we pass the list [80, 90]. This will compute a 80% and 90% confidence interval.

timegpt_fcst_multiseries_with_history_df = nixtla_client.forecast(df=df, h=24, level=[80, 90], add_history=True)
timegpt_fcst_multiseries_with_history_df.head()

INFO:nixtlats.nixtla_client:Validating inputs...
INFO:nixtlats.nixtla_client:Preprocessing dataframes...
INFO:nixtlats.nixtla_client:Inferred freq: H
INFO:nixtlats.nixtla_client:Calling Forecast Endpoint...
INFO:nixtlats.nixtla_client:Calling Historical Forecast Endpoint...

nixtla_client.plot(
    df, 
    timegpt_fcst_multiseries_with_history_df.groupby('unique_id').tail(365 + 24), 
    max_insample_length=365, 
    level=[80, 90],
)

In the figure above, we now see the historical predictions made by TimeGPT for each series, along with the 80% and 90% confidence intervals.