Reading 8: Representative bias and base-rate neglect

I am on p.56, Vol 2 of curriculum 2014.

From my understanding, representative bias means that people tend to overweigh new information because they think that the new information is representative of the whole population. Sample-size neglect seems consistent with this definition. However, I don’t understand why base-rate neglect (the base rate or probability of the categorization is not adequately considered) is a type of representative bias.

CFAI uses Company ABC as an example of base-rate neglect. But in that example, what is the ‘new’ information? It seems that this example is more consistent with the definition in Wikipedia:

Thanks for your help!

Representativeness bias is taking new information and putting into “buckets” because the information is similar or “representative” to what you have experienced in the past

Base-Rate Neglect is not adequately considering whether the logic for putting new information into “buckets” is correct. In the ABC case the FMP has “bucketed” Company ABC into a growth stock without adequately considering if it truly is a growth stock. The “new” information is simply the characteristics of Company ABC.

Sample-Size Neglect is taking a small sample size of new information and assuming that this is representative of the population as a whole.

… Or at least that’s my understanding.

Thank you Marathon_runner! I think your explanation of representative bias is much better than that in the curriculum (Perhaps because I am not a native speaker of English?).

Can you also help me understand base-rate neglect in mathematical terms? According to, base-rate neglect is wrongly equating P(A|B) = P(B|A) and neglecting P(A)/P(B) in the Bayes formula. Applying to Company ABC example, FMPs neglect P(growth) and are trying to estimate P(success|growth). Does that mean FMPs wrongly think that P(success|growth) = P(growth|sucess)?

I think it would go like this. The FMP would think like this:

P(ABC actually is growth stock | ABC looks like a growth stock) = 100% (if it looks like a growth stock then it IS a growth stock)… when in fact by Bayes:

P(ABC actually is growth stock | ABC looks like a growth stock) =

P(ABC looks like a growth stock | ABC actually is growth stock) x P(ABC actually is growth stock)/(ABC looks like a growth stock)

= 100% x (something less than 100%) / (100%) = something less than 100%

… but I could be wrong – I’m not an expert.

Thanks a lot! That makes more sense to me!

If applying Bayesian lanauge to Question 2, Example 2, p.57,

Verte is evaluating P(recover | -ve news).

P(recover | -ve news)

= P(-ve news | recover) P(recover) / P(-ve news)

= P(-ve news | recover) P(recover) (since P(-ve news) = 1 )

P(recover) is high but P(-ve news | recover) is low. And Verte neglects the base rate P(recover) and so wrongly concludes that P(recover | -ve news) is low.

Am I correct?

you should read Daniel Kahnemen’s book and paper. Sheds a lot of light on these topics

Is it worth? It seems CFAI does not recommend doing this. Can you share your experience?