Scott Wylie Insight DS Demo

Multi-touch attribution with Interstate Analytics

Transcript

1. “Helping  marketers  make  be1er  decisions…”   Sco1  Wylie   Insight  Data  Science  
2. The  Problem:   How  should  a  business  spend  its  limited  online  adverCsing  budget?   simple  example:  just  2  “channels”:  Google  and  Facebook.   Which  allocaCon  maximizes  profit?   Goog   FB   Goog   FB   Goog   FB  
3. “Gold  Standard”  Approach:  Trial  and  Error   profit  =  $15,000   profit  =  $12,000   profit  =  $8,000   Drawbacks:   •  difficult  to  execute   •  expensive   •  ignores  trove  of  data   Goog   FB   Goog   FB   Goog   FB  
4. “Visit  paths”  provide  insight  into  what  works,                                       prior  to  experimentaCon   profit  =  $15,000   Goog   FB  
5. “Visit  paths”  provide  insight  into  what  works,                                       prior  to  experimentaCon   profit  =  $15,000   $4,000   $3,000   $500   $2,000   .  .  .   $15,000   many  other  paths   g g g g g
6. My  deliverable:     “A  model  that  assigns  a  dollar  value  to  clicks  leading  to  sales.”     INPUT:  PostgreSQL  DB  of  visit  paths  paired  with  revenue  generated:   OUTPUT:  PosiCon-­‐dependent  regression  model  for  the  value  of  each  click  in   a  visit  path.  
7. Case  study  for  a  certain  clothing  company   The  Size  of  Things…   #  Sales   85,000   #  DisCnct  Paths   26,000   Avg.  Path  Length   6.4   #  DisCnct  Channels   900   #  Channels  Included   11  (96%  of   all  clicks)  
8. Case  study  for  a  certain  clothing  company   Upper-­‐Funnel   Mid-­‐Funnel   Lower-­‐Funnel   The  Size  of  Things…   #  Sales   85,000   #  DisCnct  Paths   26,000   Avg.  Path  Length   6.4   #  DisCnct  Channels   900   #  Channels  Included   11  (96%  of   all  clicks)  
9. Case  study  for  a  certain  clothing  company   Upper-­‐Funnel   Mid-­‐Funnel   Lower-­‐Funnel   MarkeCng  Insights:     •  Google  is  a  good  upper-­‐funnel   but  FB  is  not.     •  Neither  channel  tends  to  be   last-­‐click  (poor  lower-­‐funnels).   •  Upper-­‐funnels  have  more   impact  than  lower  ones.       The  Size  of  Things…   #  Sales   85,000   #  DisCnct  Paths   26,000   Avg.  Path  Length   6.4   #  DisCnct  Channels   900   #  Channels  Included   11  (96%  of   all  clicks)    $  Value    (regression  coefficient)   g
10. Methods   High  Dimensional  Path  Data:  
11. Methods   High  Dimensional  Path  Data:   PosiCon-­‐dependent  features   First  click:     categorical  (~10  levels)     last  click:     categorical  (~10  levels)   frequency  across  path:     (~10  scalar  features)     Finally,  regress  value  ($)  against  posiCon-­‐dependent  features  …  
12. Sco1  Wylie,  PhD  in  Physics   Team  Sports  (and  High  5’s)   Me   Academic  Work:  EvoluCon  and  Biophysics     @p @t = @ 2 @x 2 [ x (1 x ) p ] @ @x [ x (1 x ) p ]