Examples:
- [Capturing Financial markets to apply Deep Reinforcement Learning]
- [A Deep Reinforcement Learning Framework for the Financial Portfolio Management Problem]
- [Deep Reinforcement Learning for Automated Stock Trading]
- Output/hidden-layer embeddings of a financial news sentiment analysis pre-trained model. [FinBERT]. Justification: such a model is able to extract valuable info from unstructured data such as financial news. Could provide important insights for trade decisions.
- Technical Indactors, price history (OHLCV), financial statements (all with timestamps). Pass through a Transform network (3 layers?).
- Portfolio State. 2D Matrix with 3 columns (price, number of shares, cash) and N rows (where N is the number of different/pre-selected assets in the portfolio)
The AI trader can perform 1 of 3 actions for each asset in the portfolio: buy/sell/hold
Implement AI for predicting stock market / turning a profit on auto.
Investors look at many factors that influence markets in order to decide to invest/trade/buy/hold/sell an asset.
portfolio Factors can include (but no limited to):
- financial statements
- financial news
- psychology of the masses regarding the company
- competition
- stock price and history
- macro economic factors (i.e: interest rates, currency, country, industry, trade agreements, subsidies, local, national and foreign governments policies)
- style factors: momentum, volatility, value and quality etc...
Most of these require knowledge in the field which is acquired by many hours of reading and digesting public information.
The state-of-the-art NLP model (Google: GPT-3) is now able to generate text which looks generated by a human. This means that it could also be able to "digest" or interpret text information.
Perhaps an AI system to predict the stock market / inform trade decisions could be developed since AI models now have the ability to "digest" text information similarly to a human (i.e: similarly to an investor?). Such a system could comprise of multiple models (1 for each factors?) which evaluates the various factors that influence investor decisions.
One of these factors could be the sentiment/psychology of the investors regarding the stock/asset. An atempt at representing this could be the sentiment analysis of financial news and other media categories i.e. social[Twitter, FB etc...], blogs/vlogs, publications, articles. Topics related to the stock/asset price, market, competition etc. The volume of the available content is massive and choosing which data to mine is a form of human/manual feature engineering. Expanding this manually via human creativity and igenuity could be possible. Choosing which data to feed to the model is the craft. (Could a generative model learn which data to mine? Choosing which article to read and which content to consume is something an investor has to do.)
An intial draft of such a model could be as follows:
Features/Inputs:
- financial statements embeddings (?-dimensions)
- sentiment analysis embeddings
- stock price history embedding
- other?
The model takes as input a combination (concatenation?) of the listed features and outputs buy/sell decision of a given time-length (10-years?, 5-year?, 1-year?, 6months?, 1 week?, 1day?). How long would it take to process all available stock data? New data/content gets created every minute. The time between start of model training and final model prediction could be a practical challenge to reduce. Perhaps 2 models would be needed? 1st is trained on all data mined before time T (historical) and 2nd is trained on all data mined after time T.
-
[This article] shows one method for sentiment analysis of financial news. Improvement would be to replace the sentiment analysis model with a more state-of-the-art model such as [FinBERT].
-
This article shows how to web-scrape financial data -> https://medium.com/datadriveninvestor/use-python-to-value-a-stock-automatically-3b520422ab6
GPT-3 paper: https://arxiv.org/pdf/2005.14165.pdf
GPT-3 is a Transformer model. A Transformer model would likely be the first choice for text analysis and stock price prediction as it is the state-of-the-art sequence-sequence models
Another transformer model could be used to encode/embed stock price data.
Use this to simulate the stock market for the AI training -> https://github.com/dchrostowski/investopedia_simulator_api Google/Find other APIs for Investopedia
Website that appears to have a successful implementation -> https://iknowfirst.com/
gamification of web platform?: could use human/expert input as imitation learning? imitate warren buffet? collection of "experiences" for experience replay which are human-driven such as the historical investing performance of known succesful traders/investors?
could extend this openAI gym environment -> https://awesomeopensource.com/project/AminHP/gym-anytrading