This model actually demonstrates how we can deal with environment in the financial markets. It is usually quite difficult to model environment in financial markets. OpenAI Gym provides rich environments about video games, classic control tasks, robotics simulations but except financial markets.
However, there is a problem in llSourcell method, which always gets a profit of 0 when evaluating model. Then xtr33me tried to force the first iteration a buy, which often generates one trade only (one buy/sell signal). Still, it can’t solve the problem fundamentally. I suppose that the agent(multilayer perception network) falls into local minima/local optima during training process. In fact, training a neural network with an optimization algorithm is hard, not least because the error surface is non-convex and contains local minima, flat spots, and is highly multidimensional.
Alternatively, I apply Actor-Critic method, which combines value-based and policy-based methods. The results indicate that the problem mentioned above is improved a lot.
To train the model, just enter the command: python train_A2C.py.
To evaluate the model, enter the command: python evaluate_buy. I already include the trained model in the code, or you can modify the program to change model.
Besides, I include winning odds as the other indicator. Trading is very “personal” style. Everybody has different risk appetite. You can also add other indicators such as Sharpe ratio, Sterling ratio, etc. Finally, I must reiterate that building a model for trading is very time-consuming. It might NOT need very advanced theory or techniques, but it absolutely requires strict confirmation and validation.
""" ####################################### """
This is the code for this video on Youtube by Siraj Raval. The author of this code is edwardhdlu . It's implementation of Q-learning applied to (short-term) stock trading. The model uses n-day windows of closing prices to determine if the best action to take at a given time is to buy, sell or sit.
As a result of the short-term state representation, the model is not very good at making decisions over long-term trends, but is quite good at predicting peaks and troughs.
Some examples of results on test sets:
!^GSPC 2015 S&P 500, 2015. Profit of $431.04.
Alibaba Group Holding Ltd, 2015. Loss of $351.59.
Apple, Inc, 2016. Profit of $162.73.
Google, Inc, August 2017. Profit of $19.37.
To train the model, download a training and test csv files from Yahoo! Finance into data/
mkdir models
python train.py ^GSPC 10 1000
Then when training finishes (minimum 200 episodes for results):
python evaluate.py ^GSPC_2011 model_ep1000
- Had to modify sigmoid slightly to prevent overflow occuring in Math.exp
- Added a boolean for first iteration which forces a buy so there is something in agent.inventory (May look into finding the best statistical time to buy in a future imp based on the current price of entry...for now this has gotten things working when evaluating)
Deep Q-Learning with Keras and Gym - Q-learning overview and Agent skeleton code