Posts: 13 since Jun 2009
Thanks Given: 7
Thanks Received: 28
|
I was able to develop several models using Microsoft's ML.Net library by training on tabular setup data and then inference these models within NinjaTrader 8. Because NT8 is built on an older .Net framework, I could not access the models directly, so I had to save the models in ONYX format for inferencing within NT8 and figure out the requisite libraries to reference. I developed separate models for long and short setups and was able to get the AutoML feature working to select the best ML methods and hyperparameters to train on.
Accessing the models for go/no-go setup recommendations was fast as I was able to backtest, optimize, and walkforward my setup threshold levels and position management parameters. Also, keep in mind, I was only using machine learning methods and not deep learning techniques, but I believe ML.Net is evolving to be able to retrain and inference DL models in TensorFlow. I do like the idea of Azure ML Studio and MLOps to manage model development workflow and deployment, but have not tried it to date.
Prior to working with ML.net, I had developed similar predictive models using Microsoft's CNTK deep learning library, but only simple multi-layer ANNs. When I transitioned to ML.NET it was my opinion LSTMs where falling out of favor for futures price prediction in favor of Convolutionary Networks (CNN) and Reinforcement Learning (RL) methods, but your mileage may vary.
Whitmark
|