elicithub / conferences Goto Github PK

Hello -
I attended your presentation at ODSC. It was really interesting! I am wondering if you can explain a bit more about the loss calculation?

For example, the loss associated with picking the new coupon at day 1 is (in words): "What is the cost we incur if we made a bad choice and the old coupon was better?". We need to have an estimate of how many less responders we would see over some period of time multiplied by their margin. You are picking 180 day as you expect that you'd have the chosen coupon in place after making a choice and that you see 1000 visitors a day on your website and with the old coupon. each missed out on buyer would generate on average $100. Is all of this correct?

I think we want to use the probability that old is actually better to weight the above loss (as some draws show old better than new, while others show new better than old).

Then I get lost...... Why set some to zero where old is worse than new? Why divide by 4000 (total number of posteriors)? Why use 500 for the days when looking at the cost of waiting 1 more day)?