Using sktime to forecast series with datetime index

[yarnabrina]

Let's consider the following setup:

import numpy
import pandas

numpy.random.seed(seed=0)

past_data = pandas.DataFrame(
    {
        "time_identifier": pandas.to_datetime(
            [
                "2024-01-01",
                "2024-01-02",
                "2024-01-03",
                "2024-01-04",
                "2024-01-05",
                "2024-01-06",
                "2024-01-07",
                "2024-01-08",
            ],
            format="%Y-%m-%d",
        ),
        "series_data": numpy.random.random(size=8),
    }
)
past_data = past_data.set_index(["time_identifier"])

As far as I can tell, this is in one of the sktime compatible format. And, training a model works fine.

import sktime.forecasting.naive

model = sktime.forecasting.naive.NaiveForecaster()
model.fit(past_data)

But I am unable to use this for prediction neither with integer array, nor with timedelta index.

import sktime.forecasting.base

numeric_horizon = sktime.forecasting.base.ForecastingHorizon(values=[1, 2])
model.predict(fh=numeric_horizon)

future_dates = pandas.to_datetime(["2024-01-09", "2024-01-10"], format="%Y-%m-%d")
past_end_date = past_data.index.max()

timedelta_horizon = sktime.forecasting.base.ForecastingHorizon(values=future_dates - past_end_date)
model.predict(fh=timedelta_horizon)

Both attempts fail with the following error, and I'm not sure what value is None.

TypeError: unsupported operand type(s) for *=: 'int' and 'NoneType'

So, I want to know how to use the (successfully) fitted model to predict in such cases.

---

NB: If I try a very similar case with multiple series in a long format, this works fine. Attaching that example below. So, I'm certain that I'm missing something very small but important stuff somewhere, and will really appreciate some guidance.

Complete Long Format Example

import numpy
import pandas

numpy.random.seed(seed=0)

past_dates = pandas.to_datetime(
            [
                "2024-01-01",
                "2024-01-02",
                "2024-01-03",
                "2024-01-04",
                "2024-01-05",
                "2024-01-06",
                "2024-01-07",
                "2024-01-08",
            ],
            format="%Y-%m-%d",
        )

past_x_data = pandas.DataFrame(
    {
        "series_identifier": "x",
        "time_identifier": past_dates,
        "series_data": numpy.random.random(size=8),
    }
)
past_y_data = pandas.DataFrame(
    {
        "series_identifier": "y",
        "time_identifier": past_dates,
        "series_data": numpy.random.random(size=8),
    }
)

past_data = pandas.concat([past_x_data, past_y_data], ignore_index=True)
past_data = past_data.set_index(["series_identifier", "time_identifier"])

import sktime.forecasting.naive

model = sktime.forecasting.naive.NaiveForecaster()
model.fit(past_data)

import sktime.forecasting.base

numeric_horizon = sktime.forecasting.base.ForecastingHorizon(values=[1, 2])
model.predict(fh=numeric_horizon)

future_dates = pandas.to_datetime(["2024-01-09", "2024-01-10"], format="%Y-%m-%d")
past_end_date = past_data.index.levels[1].max()

timedelta_horizon = sktime.forecasting.base.ForecastingHorizon(values=future_dates - past_end_date)
model.predict(fh=timedelta_horizon)

>>> numeric_horizon = sktime.forecasting.base.ForecastingHorizon(values=[1, 2])
>>> model.predict(fh=numeric_horizon)
                                   series_data
series_identifier time_identifier             
x                 2024-01-09          0.891773
                  2024-01-10          0.891773
y                 2024-01-09          0.087129
                  2024-01-10          0.087129
>>> 
>>> timedelta_horizon = sktime.forecasting.base.ForecastingHorizon(values=future_dates - past_end_date)
>>> model.predict(fh=timedelta_horizon)
                                   series_data
series_identifier time_identifier             
x                 2024-01-09          0.891773
                  2024-01-10          0.891773
y                 2024-01-09          0.087129
                  2024-01-10          0.087129

[fkiraly]

quick guess (will look into this in more detail): something goes wrong with freq - the multiplication issue seems like it is trying to translate the integer offset into a multiple of the period, but cannot infer the period because freq ends up being None rather than whatever would be the equivalent of "one day". The key to understanding the bug possibly lies in inspecting what type of index it is converted to, and what the freq attribute is.

[yarnabrina]

Can confirm that's the case. For failing single series, frequency is None and for working multiple series, it's D.

# source
## sktim/utils/datetime.py - L181 - `_shift` function

# single
## DatetimeIndex(['2024-01-08'], dtype='datetime64[ns]', name='time_identifier', freq=None)

# multiple
## DatetimeIndex(['2024-01-08'], dtype='datetime64[ns]', name='time_identifier', freq='D')

So, how do you recommend to predict in this case? And, what's causing this difference?

[yarnabrina]

Interestingly, even wide form data works fine. So, the issue is somehow restricted to single series somehow. The following works as well.

Two Series with Datetime Index

import numpy
import pandas

numpy.random.seed(seed=0)

past_data = pandas.DataFrame(
    {
        "time_identifier": pandas.to_datetime(
            [
                "2024-01-01",
                "2024-01-02",
                "2024-01-03",
                "2024-01-04",
                "2024-01-05",
                "2024-01-06",
                "2024-01-07",
                "2024-01-08",
            ],
            format="%Y-%m-%d",
        ),
        "series_1_data": numpy.random.random(size=8),
        "series_2_data": numpy.random.random(size=8),
    }
)
past_data = past_data.set_index(["time_identifier"])

import sktime.forecasting.naive

model = sktime.forecasting.naive.NaiveForecaster()
model.fit(past_data)

import sktime.forecasting.base

numeric_horizon = sktime.forecasting.base.ForecastingHorizon(values=[1, 2])
model.predict(fh=numeric_horizon)

future_dates = pandas.to_datetime(["2024-01-09", "2024-01-10"], format="%Y-%m-%d")
past_end_date = past_data.index.max()

timedelta_horizon = sktime.forecasting.base.ForecastingHorizon(values=future_dates - past_end_date)
model.predict(fh=timedelta_horizon)

[fkiraly]

Confirmed on my env, I opened a bug issue here: #4405