MMM.optimal_allocation_dict not giving correct values

[emil-matti117]

Hey everyone! I'm using the budget allocation function MMM.allocate_budget_to_maximize_response and when trying to print out the mmm.optimal_allocation_dict , it is giving me an equal budget across all of my channels.
My initial thought is because I did not assign budget bounds to any channel this was the reason, but based on this example in the documentation, that shouldn't be the case. Any thoughts here?

For reference, the Allocation column is the optimal_allocation_dict but it is not matching the blue bars in the mmm.plot_budget_allocation function (which the source code points to optimal_allocation_dict.

[wd60622]

@cetagostini would know best

[cetagostini]

Hello @emil-matti117

Can you share the code you are using? A few lines would be enough. I want to see how did you get the dataset and calculate the response.

[cetagostini]

Hey @AlfredoJF thats totally correct.

@emil-matti117 what parameters have you been using? Can you share? Whats the budget you are using? Are you able to access to minimize object? Observe the results there should help to understand whats actually happening during the optimization.

@tim-mcwilliams did you tried increasing iterations as well?

[tim-mcwilliams]

@cetagostini, I have not as of yet. Where would we increase the iterations at? Not seeing an arg in the allocate_budget_to_maximize_response for that. If that has to be done in the **minimize_kwargs, right now that isn't working for us. It's yelling at us about a TypeError.

[emil-matti117]

response1 = model_load.allocate_budget_to_maximize_response(
            budget=27_464_236, 
            time_granularity='monthly',
            num_periods=1,
            budget_bounds=None
        )

response2 = model_load.allocate_budget_to_maximize_response(
            budget=20_000_000, 
            time_granularity='monthly',
            num_periods=3,
            budget_bounds=None,
            noise_level =.02
        )

response3 = model_load.allocate_budget_to_maximize_response(
            budget=27_464_236, 
            time_granularity='quarterly',
            num_periods=3,
            budget_bounds=budget_bounds 
        )

@cetagostini these are examples of iterations we've been using. We are also unsure how to access the minimize_kwargs object to pass that to the function, do you mind showing us how we'd go about that as well?

Thanks!

[cetagostini]

Are you changing granularity for the same model? Even when model was train with the same granularity? @emil-matti117

On the other hand, are you trying something like the following? @tim-mcwilliams

results = mmm.allocate_budget_to_maximize_response(
    budget=100000,
    time_granularity="weekly",
    num_periods=4,
    budget_bounds=budget_bounds,
    minimize_kwargs={
        "method": "SLSQP",
        "options": {
            "maxiter": 3000,
            "disp": True
        }
    }
)

[tim-mcwilliams]

@cetagostini, Yea we are changing the granularity for the same model. To make sure I am fully understanding here, if the model was trained on weekly data can we only use the weekly time granularity for the allocator??

Yea, we tried that exact block and are getting the following error - TypeError: MMM.allocate_budget_to_maximize_response() got an unexpected keyword argument 'options'

[emil-matti117]

@cetagostini yes, we've tried every time granularity and am still getting the same results. Also, the above minimize_kwargs argument is giving me the following error:

TypeError: MMM.allocate_budget_to_maximize_response() got an unexpected keyword argument 'minimize_kwargs'