Hi all,
I read carefully the thread Machine Learning on Quantopian threads and I had a look at the algorithms as well.
I have a couple of questions that are both methological and practical.
A possible testing strategy is to alternate train and test phases for a long period of time (e.g. from 1991 to 2010), in order to have the model experiencing different market conditions. So we can train a model for a fixed time window (say 100 days for instance) then build a portfolio, hold it for a period (e.g. one week) , then re-train the model for a shifted period of time and so on, in order to cover all the wide time span 1991-2010.
1. I believe this is something that can be achieved with Quantopian on the backtesting environment, right? Or we may experience perfomance problems? After all we'll be training the classifier with few data, so the process should take some time but I shouldn't experience out-of-memory problems
2. Here's the question (maybe a naive one, so forgive me): let's imagine I'm happy with the performance of the model I've trained with the approach of point 1, how can I move the same model into production (with live data or in paper trading mode)?
Thanks a lot !
Giovanni
- First note that Quantopian's data source only goes back to Jan 2, 2002. What you are proposing is certainly possible, though your model's parameters may be somewhat volatile over such a long period of time. The advantage of this is that it is self-adjusting, but the disadvantage of this is that the model will have to be implemented into the algorithm itself (which may cause memory problems depending on the complexity of the model).
An alternative method is to train/test the model in Research with a shorter period of time (e.g. last 2 years) but larger time window (full 2 years instead of 100 days) and use only the outputs in the algorithm itself. The advantage of this is that you have much more data to work with, which can lead to better model performance, but the model's parameters remain static as time passes. So there is a trade-off here.
- Following the approach you proposed, you'd end up with a set of different models for different time periods. If possible, implementing the actual training/testing into the algorithm would be ideal (i.e. doing computations within before_trading_start() ).
The material on this website is provided for informational purposes only and does not constitute an offer to sell, a solicitation to buy, or a recommendation or endorsement for any security or strategy, nor does it constitute an offer to provide investment advisory services by Quantopian.
In addition, the material offers no opinion with respect to the suitability of any security or specific investment. No information contained herein should be regarded as a suggestion to engage in or refrain from any investment-related course of action as none of Quantopian nor any of its affiliates is undertaking to provide investment advice, act as an adviser to any plan or entity subject to the Employee Retirement Income Security Act of 1974, as amended, individual retirement account or individual retirement annuity, or give advice in a fiduciary capacity with respect to the materials presented herein. If you are an individual retirement or other investor, contact your financial advisor or other fiduciary unrelated to Quantopian about whether any given investment idea, strategy, product or service described herein may be appropriate for your circumstances. All investments involve risk, including loss of principal. Quantopian makes no guarantees as to the accuracy or completeness of the views expressed in the website. The views are subject to change, and may have become unreliable for various reasons, including changes in market conditions or economic circumstances.