Could I use deep_reinforcement_learning for UBS Quant Hackathon 2020 ?

Alphathon
#1

Hello administrator:

We are a team coming from china, i am a PHD student studing in ChongQing University. My Major is finance , and my research direction is deep reinforcement learning( abreviation is DeepRL),

Here our team have some questions about the competition rules：

First , in the " Outperform the S&P 500 Index " rule, which says portfolio must be equally weighted, but if we use DeepRL for trading and get a better return, the portfolio may not be equally weighted.

Second, in the "  Machine Learning for Complex Pricing Models" objective , which says it’s a supervised regression problem, but DeepRL does’s belongs to supervised machine learning.

So, Could i use deep reinforcement learing for  UBS Quant Hackathon 2020? If we can use DeepRL for this compition, that’s to say we can break the rule shows above?

#2

Hi Huanggang,


All submissions have to abide by the rules and portfolio constraints to be considered for the next round. In " Outperform the S&P 500 Index ", portfolio must be equally weighted.


As for “Machine Learning for Complex Pricing Models”, there is no restriction on the type of model to be “supervised regression”.


Hope this clarifies, thank you.

#3

Hello -


Just to clarify Reiyun’s point. At each rebalancing, you must put an equal weight on each stock you select. You are free to use DeepRL to make your stock selection.


For the second problem, the objective is to find the right ‘price’ given the set of inputs. This is thought to be ‘at the end’ a supervised regression problem as for the input data you get a ‘price’. If you manage to achieve to a price using DeepRL, even though your code will use DeepRL we will consider your model to be overall a good result to the problem. Thus we do not think it will breach the rules.


Please note that you can only use data what is provided on the platform and for the competition, be aware of ‘pre-training’ your model with other data, that may-be seen as a breach of the rules. Widely used pre-trained models that are accessible via third party packages should be acceptable.


Happy coding ! Lionel.