Forecasting with GluonTS

[jumpingfella]

Hi,

Did anybody work with GluonTS?
https://github.com/awslabs/gluon-ts

I created forecasting model with CNN using Keras, in Python, feeding it with few technical indicators from talib. Unfortunately accuracy is pretty low and results are not that encouraging so far, so thinking about switching to something else.

 

[tommcginnis]

If you are trying to forecast market activities with a learning set over the past 2-3-5 years, you're going to have big issues:
• the market moves as Big Agenda movers [reluctantly] move it.
• the Big Agenda movers used to take in fundamental data, prognosticate their own 'wisdom' ["add alpha'] to it, and then executed with as little fanfare as they could manage.
• in the past 18-24 months, the vast majority of market movement has been triggered by non-market events, even to the point of competing with (or entirely swamping) earnings season for moving the indexes.

Long story short? Throwing money, time, or Keras (or Gluons) at the problem won't solve it.
The problem is less in the model than in the environment: modeling depends on stable data -- the fin/markets environment has not only changed (from what it was), but it remains in flux.

One potential solution? Go out ~20 years, such that the financial crisis/subsequent recovery represent only half of your data. Your model will have to be much looser; your return expectations will have to be much lower; your risk/position management will have to be much tighter. BUT, you will likely have a model/trading RPA set that is much more robust (profit-wise and risk-wise) to the errant Tweet.