Related Posts
What r y’all doing for Naija independence ?? 🇳🇬
Anonymous User
Conversation Starter
Bowl Leader
Bowl Leader
More Posts
Enthusiast
Conversation Starter
Level 10.1 how much hike and vpay?
How was you memorial day weekend?
Additional Posts in Data and Analytics in Advertising & Marketing
Best resource to learn SQL?
New to Fishbowl?
Download the Fishbowl app to
unlock all discussions on Fishbowl.
unlock all discussions on Fishbowl.
sounds like a MMM project, I would suggesst figure out the saturation points for your media channels where more budget would no longer correlate to more conversions. Planners wanna get an idea of how much they should spend in each channels since they most likely have a fixed total budget and need to figure out spend distrubution, so they'd appreciate knowing the min/max spend for each channels.
Ty! See comment below
Bowl Leader
Given you already have your model... In addition to any inflection points V1 suggested, throw on some Bayesian stats. Get reasonable prior distributions on the inputs and run simulations to provide utility curve(s) with error bounds. Then simplify the results to show different risk-reward categories.
Your bayesian model should be giving you predictions with uncertainty. The only way to make more accurate would be to train on more data, reparameterize, or find more informative priors. If the prediction is conservative I would start there rather than trying to combine models.
Depending on where you work, you might have some cool resources you can work with. In my experience agency proprietary tools are kinda meh but if you crack them open you might find something really cool.
Mentor
<3 this bowl
Bowl Leader
Are you training the model on the entire data set or do you have a holdout group to validate the model ?
I validate the model use time aware cross validation. But then retain on all data when I’m ready to create the actual scenarios
Our MTA vendor also provides a scenario planning optimization tool like this but I wish they’d give us more information as to how the model was built. I feel like asking these great questions above would result in runaround half-answers 🙄 how can I trust the model if no one seems to have the ability to explain it?
lollll i knew it, those people will never give you any details on their models
Bowl Leader
Google “incrementality” - in short - to find the true impact of one channel, you need to turn the channel off somehow [geo, pulsing, etc] and awesome what happens.
MTA vendors are notoriously black-box.
Bowl Leader
Build your own models.