bug: Prophet model loading error

[Jigit-ship-it]

Initial Checklist

• I confirm that I'm using the latest version of TimeCopilot.
• I confirm that I've reviewed my issue at https://github.com/AzulGarza/timecopilot/issues before opening this one.

What happened?

<<ValueError: Dataframe has less than 2 non-NaN rows.>> occurs even though the data has a lot of rows and correct form.

Could you please advise or fix it?

Relevant log output

Environment

Latest version

[spolisar]

I recreated this with TimeForecaster.detect_anomalies()using the Excel file you provided in the discord server, is that where you encountered this issue?

This is how I recreated the issue.

df = pd.read_excel('Sample.xlsx')

tcf = TimeCopilotForecaster(
    models=[Prophet()],
)

tcf.detect_anomalies(
    df=df
)

With TimeCopilot as it is right now, you can avoid this by adjusting the horizon h argument or n_windows argument. An h of 2 or n_windows of 36 seem to work.

Internally, the data is getting split into multiple sections and run through the model. With arguments being inferred, the first split ends up with a single row of data resulting in Prophet throwing the less than 2 NaN rows error. Below is the section where the splitting occurs.

https://github.com/AzulGarza/timecopilot/blob/b85ac48abd0b5deffaa1f336719acd3d12eb019d/timecopilot/models/utils/forecaster.py#L245-L265

@AzulGarza What do you think would be the best way of handling issues like this internally?
Forcing a larger h when Prophet would throw this error?
Forcing a step_size greater than 1 for the backtest_split call for Prophet?
Bumping down the maximum for n_windows?

[Jigit-ship-it]

I recreated this with TimeForecaster.detect_anomalies()using the Excel file you provided in the discord server, is that where you encountered this issue?

This is how I recreated the issue.

df = pd.read_excel('Sample.xlsx')

tcf = TimeCopilotForecaster(
models=[Prophet()],
)

tcf.detect_anomalies(
df=df
)
With TimeCopilot as it is right now, you can avoid this by adjusting the horizon h argument or n_windows argument. An h of 2 or n_windows of 36 seem to work.

Internally, the data is getting split into multiple sections and run through the model. With arguments being inferred, the first split ends up with a single row of data resulting in Prophet throwing the less than 2 NaN rows error. Below is the section where the splitting occurs.

timecopilot/timecopilot/models/utils/forecaster.py

Lines 245 to 265 in b85ac48

splits = backtest_splits(
df,
n_windows=n_windows,
h=h,
id_col="unique_id",
time_col="ds",
freq=pd.tseries.frequencies.to_offset(freq),
step_size=h if step_size is None else step_size,
)
for _, (cutoffs, train, valid) in tqdm(enumerate(splits)):
if len(valid.columns) > 3:
raise NotImplementedError(
"Cross validation with exogenous variables is not yet supported."
)
y_pred = self.forecast(
df=train,
h=h,
freq=freq,
level=level,
quantiles=quantiles,
)
@AzulGarza What do you think would be the best way of handling issues like this internally? Forcing a larger h when Prophet would throw this error? Forcing a step_size greater than 1 for the backtest_split call for Prophet? Bumping down the maximum for n_windows?

Hi. I just used the “hello world” sample code.

[spolisar]

This section from the README/quickstart, but with your data?

tc = TimeCopilot(
    llm="openai:gpt-4o",
    retries=3,
)

result = tc.forecast(df=df, freq="MS")

If it does not look like this, share code to recreate the issue if possible.

[Jigit-ship-it]

Right. But I use locally with Ollama.

[spolisar]

With what llm?

[Jigit-ship-it]

Gpt-oss 20b

[spolisar]

I don't think that exact example is the best fit for the data that I think you're using.

You should either leave the freq argument blank (which lets TimeCopilot infer frequency) or use the appropriate value for freq. I think that would be 'W' for your case.

result = tc.forecast(df=df)

or

result = tc.forecast(df=df, freq="W")

If I run it as you described, I get this error indicating that there is not enough data.

ValueError: All series are too short for the cross validation settings, at least 25 samples are required.
Please reduce `n_windows` or `h`.

You did share a Sample.xlsx file in discord, right? Because that's what I've been using to recreate this.

[Jigit-ship-it]

Right. But it is wired if I can’t set “h”. I just want to get a value of a concrete day

[AzulGarza]

thanks @Jigit-ship-it! we are taking a look at it and we'll release a fix soon

[Jigit-ship-it]

Hi! Any news?