Data Requirements
This section explains the data requirements for TimeGPT
.
TimeGPT
accepts pandas
and polars
dataframes in long
format
with the following necessary columns:
ds
(timestamp): timestamp in formatYYYY-MM-DD
orYYYY-MM-DD HH:MM:SS
.y
(numeric): The target variable to forecast.
(Optionally, you can also pass a DataFrame without the ds
column as
long as it has DatetimeIndex)
TimeGPT
also works with distributed dataframes like dask
, spark
and ray
.
You can also include exogenous features in the DataFrame as additional columns. For more information, follow this tutorial.
Below is an example of a valid input dataframe for TimeGPT
.
import pandas as pd
df = pd.read_csv('https://raw.githubusercontent.com/Nixtla/transfer-learning-time-series/main/datasets/air_passengers.csv')
df.head()
timestamp | value | |
---|---|---|
0 | 1949-01-01 | 112 |
1 | 1949-02-01 | 118 |
2 | 1949-03-01 | 132 |
3 | 1949-04-01 | 129 |
4 | 1949-05-01 | 121 |
Note that in this example, the ds
column is named timestamp
and the
y
column is named value
. You can either:
-
Rename the columns to
ds
andy
, respectively, or -
Keep the current column names and specify them when using any method from the
NixtlaClient
class with thetime_col
andtarget_col
arguments.
For example, when using the forecast
method from the NixtlaClient
class, you must instantiate the class and then specify the columns names
as follows.
from nixtla import NixtlaClient
nixtla_client = NixtlaClient(
api_key = 'my_api_key_provided_by_nixtla'
)
fcst = nixtla_client.forecast(df=df, h=12, time_col='timestamp', target_col='value')
fcst.head()
INFO:nixtla.nixtla_client:Validating inputs...
INFO:nixtla.nixtla_client:Preprocessing dataframes...
INFO:nixtla.nixtla_client:Inferred freq: MS
INFO:nixtla.nixtla_client:Calling Forecast Endpoint...
timestamp | TimeGPT | |
---|---|---|
0 | 1961-01-01 | 437.837921 |
1 | 1961-02-01 | 426.062714 |
2 | 1961-03-01 | 463.116547 |
3 | 1961-04-01 | 478.244507 |
4 | 1961-05-01 | 505.646484 |
In this example, the NixtlaClient
is infereing the frequency, but you
can explicitly specify it with the freq
argument.
To learn more about how to instantiate the NixtlaClient
class, refer
to the TimeGPT
Quickstart
Multiple Series
If you’re working with multiple time series, make sure that each series
has a unique identifier. You can name this column unique_id
or specify
its name using the id_col
argument when calling any method from the
NixtlaClient
class. This column should be a string, integer, or
category.
In this example, we have five series representing hourly electricity
prices in five different markets. The columns already have the default
names, so it’s unnecessary to specify the id_col
, time_col
, or
target_col
arguments. If your columns have different names, specify
these arguments as required.
df = pd.read_csv('https://raw.githubusercontent.com/Nixtla/transfer-learning-time-series/main/datasets/electricity-short.csv')
df.head()
unique_id | ds | y | |
---|---|---|---|
0 | BE | 2016-10-22 00:00:00 | 70.00 |
1 | BE | 2016-10-22 01:00:00 | 37.10 |
2 | BE | 2016-10-22 02:00:00 | 37.10 |
3 | BE | 2016-10-22 03:00:00 | 44.75 |
4 | BE | 2016-10-22 04:00:00 | 37.10 |
fcst = nixtla_client.forecast(df=df, h=24) # use id_col, time_col and target_col here if needed.
fcst.head()
INFO:nixtla.nixtla_client:Validating inputs...
INFO:nixtla.nixtla_client:Preprocessing dataframes...
INFO:nixtla.nixtla_client:Inferred freq: H
INFO:nixtla.nixtla_client:Calling Forecast Endpoint...
unique_id | ds | TimeGPT | |
---|---|---|---|
0 | BE | 2016-12-31 00:00:00 | 45.190453 |
1 | BE | 2016-12-31 01:00:00 | 43.244446 |
2 | BE | 2016-12-31 02:00:00 | 41.958389 |
3 | BE | 2016-12-31 03:00:00 | 39.796486 |
4 | BE | 2016-12-31 04:00:00 | 39.204536 |
When working with a large number of time series, consider using a
distributed computing
framework to
handle the data efficiently. TimeGPT
supports frameworks such as
Spark,
Dask, and
Ray.
Exogenous Variables
TimeGPT
also accepts exogenous variables. You can add exogenous
variables to your dataframe by including additional columns after the
y
column.
df = pd.read_csv('https://raw.githubusercontent.com/Nixtla/transfer-learning-time-series/main/datasets/electricity-short-with-ex-vars.csv')
df.head()
unique_id | ds | y | Exogenous1 | Exogenous2 | day_0 | day_1 | day_2 | day_3 | day_4 | day_5 | day_6 | |
---|---|---|---|---|---|---|---|---|---|---|---|---|
0 | BE | 2016-10-22 00:00:00 | 70.00 | 49593.0 | 57253.0 | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | 1.0 | 0.0 |
1 | BE | 2016-10-22 01:00:00 | 37.10 | 46073.0 | 51887.0 | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | 1.0 | 0.0 |
2 | BE | 2016-10-22 02:00:00 | 37.10 | 44927.0 | 51896.0 | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | 1.0 | 0.0 |
3 | BE | 2016-10-22 03:00:00 | 44.75 | 44483.0 | 48428.0 | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | 1.0 | 0.0 |
4 | BE | 2016-10-22 04:00:00 | 37.10 | 44338.0 | 46721.0 | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | 1.0 | 0.0 |
When using exogenous variables, you also need to provide its future values.
future_ex_vars_df = pd.read_csv('https://raw.githubusercontent.com/Nixtla/transfer-learning-time-series/main/datasets/electricity-short-future-ex-vars.csv')
future_ex_vars_df.head()
unique_id | ds | Exogenous1 | Exogenous2 | day_0 | day_1 | day_2 | day_3 | day_4 | day_5 | day_6 | |
---|---|---|---|---|---|---|---|---|---|---|---|
0 | BE | 2016-12-31 00:00:00 | 64108.0 | 70318.0 | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | 1.0 | 0.0 |
1 | BE | 2016-12-31 01:00:00 | 62492.0 | 67898.0 | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | 1.0 | 0.0 |
2 | BE | 2016-12-31 02:00:00 | 61571.0 | 68379.0 | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | 1.0 | 0.0 |
3 | BE | 2016-12-31 03:00:00 | 60381.0 | 64972.0 | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | 1.0 | 0.0 |
4 | BE | 2016-12-31 04:00:00 | 60298.0 | 62900.0 | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | 1.0 | 0.0 |
fcst = nixtla_client.forecast(df=df, X_df=future_ex_vars_df, h=24)
fcst.head()
INFO:nixtla.nixtla_client:Validating inputs...
INFO:nixtla.nixtla_client:Preprocessing dataframes...
INFO:nixtla.nixtla_client:Inferred freq: H
INFO:nixtla.nixtla_client:Using the following exogenous variables: Exogenous1, Exogenous2, day_0, day_1, day_2, day_3, day_4, day_5, day_6
INFO:nixtla.nixtla_client:Calling Forecast Endpoint...
unique_id | ds | TimeGPT | |
---|---|---|---|
0 | BE | 2016-12-31 00:00:00 | 74.540773 |
1 | BE | 2016-12-31 01:00:00 | 43.344289 |
2 | BE | 2016-12-31 02:00:00 | 44.429220 |
3 | BE | 2016-12-31 03:00:00 | 38.094395 |
4 | BE | 2016-12-31 04:00:00 | 37.389141 |
To learn more about how to use exogenous variables with TimeGPT
,
consult the Exogenous
Variables
tutorial.
Important Considerations
When using TimeGPT
, the data cannot contain missing values. This means
that for every series, there should be no gaps in the timestamps and no
missing values in the target variable.
For more, please refer to the tutorial on Dealing with Missing Values in TimeGPT.
Minimum Data Requirements (for AzureAI)
TimeGPT
currently supports any amount of data for generating point
forecasts. That is, the minimum size per series to expect results from
this call nixtla_client.forecast(df=df, h=h, freq=freq)
is one,
regardless of the frequency.
For Azure AI, when using the arguments level
, finetune_steps
, X_df
(exogenous variables), or add_history
, the API requires a minimum
number of data points depending on the frequency. Here are the minimum
sizes for each frequency:
Frequency | Minimum Size |
---|---|
Hourly and subhourly (e.g., “H”, “min”, “15T”) | 1008 |
Daily (“D”) | 300 |
Weekly (e.g., “W-MON”,…, “W-SUN”) | 64 |
Monthly and other frequencies (e.g., “M”, “MS”, “Y”) | 48 |
For cross-validation, you need to consider these numbers as well as the
forecast horizon (h
), the number of windows (n_windows
), and the gap
between windows (step_size
). Thus, the minimum number of observations
per series in this case would be determined by the following
relationship:
Minimum number described previously + h + step_size + (n_windows - 1)