why net_worth charts are different between posted article and github notebook file? about tensortrade

[codeninja]

One of the primary differences is in the episodes that the neural network agent was able to train on. The example in the source is set to train for 10 episodes so that it can finish within a demo period. The agent you see represented in the final image was tuned over many thousand episodes. This of course takes time that a demo would not have (weeks or more depending on how you train it). The framework has also undergone somewhat of a surge of development recently as people have joined the platform. We are still working out the kinks as we race towards Alpha. The 2.0rc1 that is on the master branch now is substantially different than the 0.16 version that existed 2 weeks ago. Thanks, Dallas --------------------------------------- *Professional leader of highly performant engineering teams since 1997* *512.923.4426 | Github <https://github.com/codeninja>| Twitter <https://twitter.com/dallaspool>| Linkedin <https://linkedin.com/in/dallaspool>*

…

On Fri, Nov 1, 2019 at 10:56 PM Thomas Mao ***@***.***> wrote: hi, I appreciate the TensorTrade_Tutorial.ipynb file got fixed, and I may run it myself. However, I have a question about the content. The networth chart, what we really care about as the final result, is different between the one in the Medium article (title: "Trade and Invest Smarter -- The Reinforcement Learning Way) and the one (actually two) in TensorTrade_Tutorial.ipynb file. The networth chart in Medium article, rise slowly and smoothly. Refer to MediumArticle.png And the two networth charts in TensorTrade_Tutorial.ipynb, just fall all the way (1st one) or rise shortly then fall. Refer to github-chart1.png and github-chart2 So what went wrong? What is different? It is not easy for us to choose a trading strategy to all the way lose, unless the tradebot completely trade OPPOSITE TO THE OPTIMIZED STRATEGY. Do you mind to take a look? [image: github-chart2] <https://user-images.githubusercontent.com/6857537/68065653-51f54c80-fd67-11e9-855a-132b96aa1ad8.png> [image: github-chart1] <https://user-images.githubusercontent.com/6857537/68065654-51f54c80-fd67-11e9-8811-95bef77d3840.png> [image: MediumArticle] <https://user-images.githubusercontent.com/6857537/68065655-51f54c80-fd67-11e9-8af6-a7139684c03d.png> .png — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#86>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAADUQROOPGOBBZ4BZFCQZDQRT3ATANCNFSM4JICP4DQ> .

[maochaokuo]

Hi Dallas:

I just updated everything, from github tensortrade, including the TensorTrade_Tutorial.ipynb file. I changed steps to episodes(=50), but still get similar result, as attached.

[maochaokuo]

Daily result just come out. It is a bit better, but eventually still
.....

[codeninja]

50 episodes is not nearly enough training for a production system. Train for 1000 episodes and your results should improve to production performance levels. As with any framework, the outcome will be dependent on how you implement your network, your rewards, and your strategy. Also, you must test your network on data which it has never seen before to ensure it's performance is not overfit to the training environment. (FBMExchange will do this automatically every iteration.) Thanks, Dallas --------------------------------------- *Professional leader of highly performant engineering teams since 1997* *512.923.4426 | Github <https://github.com/codeninja>| Twitter <https://twitter.com/dallaspool>| Linkedin <https://linkedin.com/in/dallaspool>*

…

On Sat, Nov 2, 2019 at 10:28 AM Thomas Mao ***@***.***> wrote: Daily result just come out. It is a bit better, but eventually still ..... [image: github-chart2episodes50] <https://user-images.githubusercontent.com/6857537/68073213-79c8cc80-fdc8-11e9-909d-c6d49f9a73db.png> — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#86>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAADUQQFIP3ZGBNRTPPCYVLQRWMC7ANCNFSM4JICP4DQ> .

[maochaokuo]

I get it. However, the difference between the article and Jupiter notebook file is misleading and confusing. Anyway, I look forward to the framework to become available in the near future. Thank you for your hard work and contribution.