Data Sources

[aron-alarik]

Hi I have a different data source than what is available here, are the data sources customizable or can we define our own data source, if so which module ? Also on the environment and technical indicators, there are several custom indicators i have built, can we add them as well ? The central idea is to add to the environment settings? Thanks !! Hope to use this in a live production environment soon !!

[zhumingpassional]

This version does not support the function. We will update the code to support it. Thanks for your suggestion. Can you list the technical indicators or provide the code to calculate them?

[aron-alarik]

this diagram is a bit patchy, but does give the central Idea, I can provide the code but its more than one long notebook with (connectors to IBrokers and Plygon and offline data streamer to Redis), a complete library if combined (total of 3000+ derived technical indicators), meanwhile will the other version of FinRL solve this purpose? Thanks !!

[Athe-kunal]

Hi @aron-alarik. In this tutorial https://github.com/AI4Finance-Foundation/FinRL-Meta/blob/master/Demo_FinRL_Meta_Integrate_Trends_data_to_DOW_Jones.ipynb, we have a custom user-defined features integration. Here I appended the data frame of user-defined features and did a left join based on date and ticker symbol. This may help you to integrate custom state space into your model. Do try and let me know if you have some other ideas to do this in a better way.

[aron-alarik]

Thanks ! I am looking at it, will try and contribute as much possible to this project, really excited about using FinRL in our core production environment. Also I was looking at the data connectors and checking if they can be customized. The central idea being real time streaming data.

[cryptocoinserver]

Awesome @aron-alarik - this would also fix #45 . I think it's better like you said to make it more modular by accepting the features via a dataframe.

BTW what reward function do you use? You might want to check out the two other issues / suggestions I added: #26 and #27

[aron-alarik]

Absolutely if it can be made more modular with a dataframe !! that would be awesome !!

Also if instead of processing data/Indicators/Env-vars directly via vendor APIs like Yahoo, make it more generic by just accepting a stream of data in JSON format from a forwarded port (like a universal subscriber or specifically from Redis/ZMQ/ Amazon etc). That ways its up to the user to choose their indicator library, vendor, data format- live streaming or historical data and the compression level. Dont think the agents care too much about where the data is coming from !!! Almost anyone can spin up a publisher in a few mins, so that will not be an issue.

I think #26 points are Bang ON !! # 1, 2 and 3 clearly cover the reward function we use, we have a slightly modified version of DSR but mostly in line with this. So not much to add to if this is available. Looking forward to calculating and optimizing the rewards using a GAN. Starting work on this after Christmas.

On the issue of point #27 , again, excellent points, both are extremely important feature sets, we call them temporal zones and Vix alternatives, I have some code for them, we use an extended temporal zone, one additional function is to add temporal zones that are specific to intraday, daily, weekly regimes for a certain index. I will share that scanner here in some time, once I am done with a working model from FinRL, please do share your views/code on this.

Your ideas are excellent, I will share what I have, as and when I am done with this RL project, and maybe we can all do the same. It will be really great to see these ideas take some shape in FinRL. I think the real challenge is here is to let the user expand his scope of the state space and source data and not make it too specific.

Awesome @aron-alarik - this would also fix #45 . I think it's better like you said to make it more modular by accepting the features via a dataframe.

BTW what reward function do you use? You might want to check out the two other issues / suggestions I added: #26 and #27

[aron-alarik]

also forgot to mention: in a high frequency environment, TA-Lib would be working with polars or terality and not Pandas, simply due to the huge pre- processing involved, so that might be an additional consideration.

[zhumingpassional]

Your suggestions are valuable.

We are developing another version, dict, where key is indicator (e.g., "close price"), value is 1-dimension numpy.array, since it is more efficient than dataframe.

I think JSON is more general than other formats. It's a good idea. Other codes (e.g., java) can process them efficiently.

"Dont think the agents care too much about where the data is coming from" It's good. In our defined ticker_list, it's easy to identify which data source it is from. While if users define their own ticker_list, we may have to search tickers over all markets. We have to store the tickers in different time. If some several tickers are away from the markets in 2010, they are not in the newest dataset.

[aron-alarik]

Thanks !! You guys are great !!

[Athe-kunal]

Hi @aron-alarik " Looking forward to calculating and optimizing the rewards using a GAN" Can you please provide more information on this?

[aron-alarik]

Hi @aron-alarik " Looking forward to calculating and optimizing the rewards using a GAN" Can you please provide more information on this?

Hey !! Yes, I am running some experiments on this over the next few days. If the 'fake agent' adds to the model performance, I will post the results here. In fact starting work on this right now.

[Athe-kunal]

Oh alright! This seems interesting. Looking forward to the results, also please post in the slack channel.

[aron-alarik]

Sure thing! Will do !