# Machine learning for non optimal behaviour

I was working on the pricing of complex bermudean swaption when I noticed that the exercise is often (very) subobptimal. It seems that the clients are more sensitive to past growth or drop in rates than to their value at the moment.

I am looking for a way to modelise the suboptimal behaviour and I tought about Machine learning. But I can't find any reference on suboptimal options exercice.

Do you have any broader exemple of Machine learning applied to the replication of human non optimal behaviour ?

edit: Well, I have a little background in ML (Finished Andrew Ng course on Coursera and currently going trough ESLII at a great pace). I know there is a lot of applications (see here for tons of exemple). I have played a bit with some basic algorithm and my data. I have some interesting results but also things to investigate. My question was more about quantitative finance.

-

Your question is too broad, but I there is plenty of examples of uses of machine learning to mimic human behaviour. For instance deep learning has been used 25 years ago to read checks in banks, or support vector machines 15 years ago to implement artificial vision, or bayesian networks to mimic expert diagnosis.

I guess it would not be that hard to use machine learning in your case if you can implement supervised learning. It mean you should have a database of human decision $D$ associated with the product $P$ and the market context $C$. Then your goal will be to emulate the function: $$D=F(P,C)+\varepsilon.$$ Of course (as usual in machine learning), you will need to focus on pre-processing to be sure the characteristics of the product and the market context are mixed in $F$ a convenient way.

Then you will have to choose a class of model ; it is difficult to help you at such an early stage since a description of the variables and the number of data are needed. Nevertheless I gave details on Artificial Neural Networks on quant.stackexchange.

Edit:

1. you should look after credit scoring. It models the way a consumer will be risky for a credit.
2. Neural networks are usually good to estimate a probability. It comes from the fact that if you train them on a database with 0 or 1 as outputs (for you, will be the observation of an exercise or not), they end up with a real between 0 and 1, i.e. the probability of an exercice.
-
Thanks for your answer, maybe a bit broad. (see my edit). I will look at your last link a bit deeper tomorrow as DNN may be sufficient for what I wanted to do. Would you say that DNN can account for (non-linear ?) lags in the market ? Or is this 'just' a NN on a moving windows ? –  lmorin Aug 12 at 22:15
Basically: the output take positive values some time after there is a drop on client rates. The lag between the drop and thechange in output depends on the size of the drop. –  lmorin Aug 12 at 22:25

I'm not sure that machine learning would lead to any practical solutions here. Do you really have enough data for that kind of techniques?

I would suggest a different approach: assume that the exercise is optimal, but just based on a different cost function than the expected pay-off. If you can find a function that replicates well enough the past exercise decisions, maybe you can use it to predict the future ones.

-
> If you can find a function that replicates well enough the past exercise decisions... Well that is the point of machine learning... –  lmorin Aug 14 at 10:40