kpi_without_optimum vs original kpi about lightweight_mmm HOT 11 CLOSED

Hello @virithavanama !
kpi_without_optimum runs a prediction based on historical average media data. If you would like to compare against a different baseline you can always do so.

Can you further explain what you would expect it to match against?

from lightweight_mmm.

If the total historical conversions is 48000, in the 'plot_pre_post_budget_allocation_comparison' plot preoptimized predicted target(kpi_without_optimum) is showing as 16000

from lightweight_mmm.

That is possible if your total historic data is eg. 30 weeks and you are optimizing for 10 weeks. Could that be your case? (values are relative not absolute in my example)

from lightweight_mmm.

No, we are using complete data for optimizing as well

from lightweight_mmm.

Can you show the inputs you are passing to the function?

from lightweight_mmm.

n_time_periods = 52
prices = jnp.ones(mmm.n_media_channels)
budget = jnp.sum(jnp.dot(prices, ch_mean))* n_time_periods

solution, kpi_without_optim, previous_budget_allocation = optimize_media.find_optimal_budgets(
n_time_periods=n_time_periods,
media_mix_model=mmm,
#extra_features=extra_features_test,
budget=budget,
prices=prices,
media_scaler=media_scaler,
target_scaler=target_scaler,
bounds_lower_pct=0.75,
bounds_upper_pct=jnp.array([1.5,1.5,1.5,1.5,1.5,1.44,1.5,1.5,1.5]),
seed=SEED)

from lightweight_mmm.

Thanks! Let me investigate this one and get back to you.

from lightweight_mmm.

I am not able to recreate this great discrepancies with mock data.

What is the output when passing all the extra features as well to the optimisation? I see you did not pass them. To match the output you are expecting you should in this case pass the training extra features.

from lightweight_mmm.

What if you want to predict the future and then we will not have the extra features right?

from lightweight_mmm.

Depends what those extra features are, for some (holidays, promotions, ...) you might know. For others no, you will have to estimate them or set them to zero.

However if you extra features have a high positive impact and you want to be optimising for the future without them, output will not match historic absolute values. But I would not focus too much on the absolute value of the optimisation, in this case you know you are missing certain part of the value due to those missing extra features.

from lightweight_mmm.

Discussion on #64 might be helpful as it is related to a very similar topic.

from lightweight_mmm.