Machine Learning for Complex Pricing Models Neural Network Questiosn

Alphathon
#1


Under Selection Criteria of Machine Learning for Complex Pricing Models, it says "it is a pre-requisite that models have a high explanatory power, i.e. that predictions can be traced. Participants need to keep this in mind if they decide to use neural networks for instance: they have to design a property to trace how a prediction is made. Conversely, this property already exists for tree-based models."


Could you elaborate on how to design a property to trace how a prediction is made if we are using neural networks?


#2

Hello - I will try to answer the question the best i can but there is a ‘grey’ area as explanatory power is not black or white


Explanatory will rank better on :

  • Models with less parameters compared to models with a lot of (hyper or not) parameters.
  • Models which you can understand the logic / decision variable (i.e. like trees) will be preferred to models which you can’t understand.
  • Models which have smooth sensitivities on the inputs will be preferred compared to a model which have jumps of results with a small change of variable.
  • Models which have clear comments / explanations will be preferred compared to models which the code is unstructured and not abstracted (abstraction is clearly very important).
  • Models which training / retraining is fully explained and transparent.
  • Models which uses less CPU / RAM will be preferred.


There may be other criteria but you understand the spirit I hope ! Probably best summarized by this quote : "“Everything should be made as simple as possible, but no simpler.” from we all know is from !


This applies to neural networks, be aware that Keras on tensorflow is provided but only on CPU with limited RAM !


I hope that helps ! Happy coding, Lionel.