Criticisms of MLE
Maximum Likelihood Estimator (MLE)
It is pretty intuitive that you want the data that you observed to be the most likely in a proposed model. However, there’s no theoretical backing for this. Nothing says that the data you observed is the most likely realization. If you flip a (fair) coin 10 times and get all heads, this is still a valid realization, but the MLE would be miles off.
References
- Conversations with Tyler Brough
- @rice2007
- @lancaster2004