Prior distributions?

Follow

Ezequiel Melitón

• July 02, 2020 20:17

Hi everyone! Does anyone know which prior distributions does Dynamic Yield use for conversion rate and for Average Revenue Per User? 

Does it uses  non-informative priors or subjective priors? And in each case, how do they model them, with what parameters?


Thanks!

0

• 

  Kamal Abdul karim

  • July 08, 2020 12:30
  • Edited

  Hi Ezequiel, 

  for binary metrics like CR: 

  1. Set a minimum P2BB value to declare a winner
  2. Prior distribution - Beta(1,1) 
  3. Gather experiment data per variation, namely # of successes (S) and # of observations (N)
  4. Posterior distribution Beta (S+1,N-S+1)
  5. Sample n observations from each posterior (Monte-Carlo)
  6. Calculate P2BB by comparing samples

  for non-binary metrics like AOV

  1. Set a minimum P2BB value to declare a winner
  2. Prior distribution - Gamma(K,T)
  3. Gather experiment data per variation, namely # of purchases (N) and sum of all purchases values (S)
  4. Posterior distribution of 1/AOV is Gamma(K+N, T/(1+T*S)
  5. Sample n observations from each posterior (Monte-Carlo)
  6. Transform each sampled value to 1/(sampled value)
  7. Calculate P2BB by comparing transformed samples

  Hope that helps, Kamal

  0

  Comment actions Permalink

Please sign in to leave a comment.