[BUG] TFTExplainer.explain() raises "time index is missing the freq attribute" when target and prediction DatetimeIndex resolutions differ

[jenkoj]

Describe the bug

TFTExplainer.explain() raises ValueError: The time index is missing the 'freq' attribute... when the target
TimeSeries uses a DatetimeIndex with resolution other than ns (e.g. datetime64[ms], common after reading
Parquet with pandas ≥ 2.0).

In darts/explainability/tft_explainer.py:224:

times = series.time_index[-icl:].union(pred_series.time_index)                                                 
attention = TimeSeries(values=..., times=times, components=...)

series.time_index is ms, pred_series.time_index (from model.predict) is ns. DatetimeIndex.union()
promotes both to ns and drops freq.

To Reproduce

import numpy as np, pandas as pd                          
from darts import TimeSeries
from darts.models import TFTModel           
from darts.explainability import TFTExplainer

idx = pd.date_range("2026-02-12 09:15", periods=2880, freq="15min").as_unit("ms")                                
series = TimeSeries.from_times_and_values(
    times=idx, values=np.random.rand(2880, 1).astype(np.float32), columns=["x"]                                  
)                                                                                                                
model = TFTModel(input_chunk_length=1, output_chunk_length=1,
                 add_relative_index=True, n_epochs=1).fit(series)                                                
TFTExplainer(model, background_series=series).explain()  # raises                                                

Expected behavior

TFTExplainer.explain() should succeed for any regular-frequency TimeSeries, regardless of the DatetimeIndex
resolution.

System (please complete the following information):

• Python version: 3.14
• darts version: 0.43.0
• numpy version: 2.3.5
• pandas version: 2.3.3

Additional context

Workaround for users: df.index = df.index.as_unit("ns") before building the TimeSeries.

[dennisbader]

Hi @jenkoj, and thanks for raising this issue. I was able to reproduce the bug with your dependency version.

I tried again with latest dependency versions and the issue is resolved.

• MacBook Pro M1 (platform_machine == 'x86_64' and sys_platform == 'darwin')
• Python version: 3.14.3
• darts version: (latest master head)
• numpy version: 2.4.4
• pandas version: 3.0.2

Can you try once with the dependency versions mentioned above? If it fixes the issue, it might be due to some issue on pandas side (or other). In that case we might not fix this on our side, since it has resolved itself.

[jenkoj]

Hi, @dennisbader! I checked on my side as well, issue was indeed resolved by upgrading pandas to 3.0.2.

[dennisbader]

Great! In that case I'll close the issue :)