5. Experts can really suck
Wide range of domains:
● College admissions
● Parole officers
● Hiring decisions (unstructured interviews)
● Medical professionals
● etc
Linear Models rule
6. Experts can really suck Linear Models rule
Why?
pourquoi?
No feedback loops
7. Linear model ~
weight x variable + weight x variable + … + weight x variable = value
0.4 x 0.5 + -0.6 x 0.7 + … + 0.1 x 1.3 = 0.8
Variables
Outcomes
Good statistics Linear model
9. How lazy can we get?
● Bootstrap your model
○ A linear model off just one expert
● Random linear model
○ Random weights but non-random sign (+/-)
● Ignore most variables
○ Simply add a few normalized scores
Experts hate him!
Pay attention for one weird trick:
10. Simple strategy for better results:
1. Estimate / make prediction
2. Consider the opposite / assume you are wrong
3. Make second estimate / prediction
4. Average the two
11. Don’t think about probabilities, think about frequencies
1. Patient thinks she has a rare disease (1 in 1000)
2. Test is 99% accurate (1% chance it is wrong each administration)
3. Test is positive
4. What is the probability the patient has disease?
hint: only 18% of Harvard Medical School faculty & staff got an equivalent problem right
hint: average diagnosis was 28 times too high