Similarities among models

• Crime model generates a prediction about where to police just as flu model generates a prediction of where to staff hospitals
• Political polls weren’t quite accurate just like flu prediction
  • doesn’t consider changes in keywords that people are using - model based on different data than being generated
  • don’t consider particular effects of this election that’s unique
• potentially relating two factors that are related one year but not the next
  • need some human awareness of how predictions are being made
  • need more testing
• data is very noisy
  • don’t overfit your data
• sales model is predictive - data makes sense with response
  • google - data = everything, may or may not make sense
• when we use the census we have to account for how we measure race
  • google doesn’t account for changes in human behavior