ValueError: All-NaN slice encountered when generating Scenario Planner proto with quarterly breakdown on weekly MMM

[MIR-S360]

Hi everyone,

I am running into an issue when generating the Scenario Planner dashboard proto using a quarterly breakdown on a weekly MMM model. The model itself runs without issues, but create_mmm_ui_data_proto fails for certain time slices with the following error:

Setup

• Weekly data (week_start, Mondays)
• Using QuarterlyDateRangeGenerator
• BudgetOptimizationSpec with relative constraints (0.9–1.1) across all paid channels
• use_optimal_frequency = True
• start_date = None
• end_date = None

Code

spreadsheet_name = "Meridian Scenario Planner"
optimization_name = "Base scenario"
include_non_paid_channels = True

start_date = None
end_date = None 

yearly = False
quarterly = True
monthly = False

min_spend_shift_ratio = 0.9
max_spend_shift_ratio = 1.1

use_optimal_frequency = True

# time breakdown
time_breakdown_generators = []

if yearly:
    time_breakdown_generators.append(
        date_range_bucketing.YearlyDateRangeGenerator
    )

if quarterly:
    time_breakdown_generators.append(
        date_range_bucketing.QuarterlyDateRangeGenerator
    )

if monthly:
    time_breakdown_generators.append(
        date_range_bucketing.MonthlyDateRangeGenerator
    )

# channel constraints
channel_constraints = []

for channel in mmm.input_data.get_all_paid_channels():
    channel_constraints.append(
        budget_optimization_processor.ChannelConstraintRel(
            channel_name=channel,
            spend_constraint_lower=min_spend_shift_ratio,
            spend_constraint_upper=max_spend_shift_ratio,
        )
    )

# optimization spec
grid_name_prefix = "-".join(optimization_name.lower().split(" "))

budgetOptSpec = budget_optimization_processor.BudgetOptimizationSpec(
    start_date=start_date,
    end_date=end_date,
    optimization_name=optimization_name,
    grid_name=grid_name_prefix,
    constraints=channel_constraints,
    use_optimal_frequency=use_optimal_frequency,
    max_frequency=max_frequency,
)

# build proto
print("Creating MMM UI proto.")

summary_spec = marketing_processor.MediaSummarySpec(
    include_non_paid_channels=include_non_paid_channels
)

mmm_proto = mmm_ui_gen.create_mmm_ui_data_proto(
    mmm=mmm,
    specs=[
        model_fit_processor.ModelFitSpec(),
        marketing_processor.MarketingAnalysisSpec(
            media_summary_spec=summary_spec,
        ),
        budgetOptSpec,
    ],
    time_breakdown_generators=time_breakdown_generators,
)

print("MMM UI proto created.")

Has anyone experienced this when using quarterly breakdown on weekly data?

[gousesb]

Hello @MIR-S360 ,

Thank you for contacting us!

To help investigate the ValueError you're seeing in the Scenario Planner, could you please confirm if any of your paid channels have zero total spend or missing (NaN) values for an entire calendar quarter? This error usually happens when a quarterly breakdown aligns with a period of time that contains no valid data.

Additionally, could you confirm which version of the Meridian library and setup you are currently using? This will help us determine if you are facing a known alignment issue with quarterly date ranges. In the meantime, you can likely fix this by manually setting a start_date and end_date in your BudgetOptimizationSpec that include only full quarters of your data.

Feel free to reach out for any further queries or feedback regarding Meridian.

Google Meridian Support Team

[sam-barnes-verisure]

I think many advertisers will have channels that are seasonal, or utilised in specific ways that may mean they're not always active.

But they are still part of the mix.

Do you know of any way to effectively allow Meridian to run these analysis even with zero spend periods?

They, of course, have zero impressions, reach, or frequency either, so during that quarterly period calculation would it be appropriate to just skip over those channels during the optimisation of that specific period?

Or would this likely cause issues with the modelling process?

[gousesb]

Hello @sam-barnes-verisure ,

Thank you for contacting us!

To help clarify your question, could you confirm if you are currently running into a specific error where certain segments of your training data contain zero-spend periods? Meridian is generally expected to handle historical zero-spend periods normally, so it would be helpful to know if you are experiencing a direct issue there.

Alternatively, if your question is about manipulating or overriding historical media flighting patterns during the optimization process itself, the recommended approach is to use the new_data object. You can find detailed instructions on how to use new_data to override historical flighting patterns in the Meridian documentation.

Feel free to reach out for any further queries or feedback regarding Meridian.

Google Meridian Support Team

[sam-barnes-verisure]

Hi @gousesb

Thanks for your reply.

Yes I'm currently running into the same issue as the original post contributor.

If I run some analysis over say the last year, looking at quarters, but there are 2 channels that had no spend in a specific quarter for whatever commercial reason, then I cannot run budget optimisation successfully.

I've triple checked the dataframe when it's loaded from the CSV and see no NaN values - just zeros where there is no spend, impressions, reach or frequency (depending on the channel of course).

My belief is that Meridian would simply account for them by saying that this channel has zero impacts within this period (unless there is some carry over based on Adstock from the previous period I suppose).

Interestingly, if I run

optimization_results = budget_optimizer.optimize(
    ...
    use_optimal_frequency=False
)

i.e. with use_optimal_frequency=False

Then I get get individual optimisations to run.

If I set use_optimal_frequency=True then it will not run at all and I get the All NaN Slice error which seems to track back to here:

File ~/jupyter/.venv/lib/python3.12/site-packages/meridian/analysis/analyzer.py:3907, in Analyzer.optimal_freq(self, new_data, max_frequency, freq_grid, use_posterior, use_kpi, selected_geos, selected_times, confidence_level)
   3895   metric_grid_temp = self.roi(
   3896       new_data=new_roi_data,
   3897       use_posterior=use_posterior,
   (...)   3901       use_kpi=use_kpi,
   3902   )[..., -self.model_context.n_rf_channels :]
   3903   metric_grid[i, :] = get_central_tendency_and_ci(
   3904       metric_grid_temp, confidence_level, include_median=True
   3905   )
-> 3907 optimal_freq_idx = np.nanargmax(metric_grid[:, :, 0], axis=0)
   3908 rf_channel_values = (
   3909     self.model_context.input_data.rf_channel.values
   3910     if self.model_context.input_data.rf_channel is not None
   3911     else []
   3912 )
   3914 optimal_frequency = [freq_grid[i] for i in optimal_freq_idx]

It would be interesting to see if there's anything within the optimal frequency calculations or processing that causes this issue.

[gousesb]

Hi @sam-barnes-verisure ,

Thank you for bringing this to our attention and for your detailed feedback.

We have confirmed the issue you reported, and a fix is currently in development. We expect this to be resolved in a future release. You can track previous changes and unreleased feature updates in our Change Log.

In the meantime, continuing to run your analysis with use_optimal_frequency=False is the best way to bypass the crash.

Feel free to reach out for any further queries or feedback regarding Meridian.

Google Meridian Support Team