Inspired by Gekko
Aim of this project is to create a bot which can create a portfolio from a given list of stocks, optimizes the portfolio by efficiently dividing the assets and then does automatic trading to maximize the profit.
Most tutorials on Internet(and my previous project on same topic) use only historial stock price to predict future prices. However in this project, important features are extracted from the data and are used in conjuction with historical prices. Due to this, it's performance is significantly better than other projects available on Internet.
Almost everything can be controlled by changing the parameters in config.py. No need to change anything in core code.
No tight coupling, proper encapsulations and abstractions ensure that code is flexible and easy to extend.
git clone https://github.com/Prakash2403/Chase.git
cd Chase
pip3 install -r requirements.txt
python3 main.py
keras
tensorflow
quandl
pandas
numpy
matplotlib
To get data using quandl API, you need to have an API key. You can find your API key on your Quandl account settings page. Once you have that, go to secrets.txt in project root directory and save your key there. Finally, secrets.txt should look like {'quandl_key': 'dffergtrth421321'}.
MAKE SURE THAT quandl_key and your API key ARE ENCLOSED WITHIN SINGLE QUOTE.
Since almost everything is controlled by parameters defined in config.py, you don't need to be a programming and/or Machine Learning wizard to predict future stock prices. If you know how to create lists and dictionaries in python, you can easily predict the future.
Currently, data is extracted using quandl API. It provides following information about the given stock.
1. Date
2. Open
3. High
4. Low
5. Close
6. Adjusted values of all above features
7. Volume
Chase uses Adjusted Close(Adj_Close) and Volume as direct historical features to be used for forecasting. If you want to use any other feature(s), change the elements of REL_DATA_COLUMNS. For example, if you want to use Open and High, set REL_DATA_COLUMNS=['Open', 'High'].
Apart from this, Chase contains implementations to extract the following features from the data
1. Simple Moving Average. -> sma
2. Exponential Moving Average. -> ema
3. Minimum Bollinger Bound -> min_bollinger_bound
4. Maximum Bollinger Bound. -> max_bollinger_bound
5. Daily Returns. -> daily_returns
Above features are defined in compute_stock_features.py in utils package. In order to use them, change the elements of REL_PREDEFINED_FEATURES list defined in config.py. For example, if you want to use Simple Moving Average and Daily Returns, set REL_PREDEFINED_FEATURES = ['sma', 'daily_returns'].
Summarizing above points, if you want to use Adjusted Close, Volume, Simple Moving Average and Daily Returns for forecasting, set REL_DATA_COLUMNS=['Adj_Close', 'Volume'] and set REL_PREDEFINED_FEATURES = ['sma', 'daily_returns'].
Ultimately, if you want to implement a feature of your own, you can do that too. Although, you will be needing a little knowledge of programming for that. Let's say that you want to add Rolling standard deviation as a feature. To do this, create a python file(say my_features.py) preferably in project root directory and define a method(say get_rolling_std). After you have done this, open config.py and modify EXTRA_FEATURES dictionary as follows: set the key to the name of feature and value to the function pointer. In this case, EXTRA_FEATURES = {'roll_std': get_rolling_std}. You can have multiple extra features like EXTRA_FEATURES = {'roll_std': get_rolling_std, 'another_feature': feature_2_pointer}. Only restriction you'll have to follow here is your function should accept only two parameters, first one being the dataframe containing all the data and second one being the column(s) which are required to calculate the feature. For above example, get rolling std will look like
def get_rolling_std(data: pd.DataFrame, feature)
return data[feature].rolling(5).std()
Documentation of all parameters will be added soon.
Chase offers you a complete playground to test your intuitions regarding hyperparameter settings. I presume you are familier with hyperparameters in keras. Not a single parameter is hardcoded in the core codebase. Every parameter can be altered by changing it's value in config.py.
Moreover, adding new network architecture is as easy as pie. Just inherit the Network base class, define the abstract methods and you are good to go.
Well, it took me some time to design the UML for this project. I tried to stick to S.O.L.I.D. principles as much as possible. I hope it may help you in case you face any design issues. Although I am not an expert in this area, so if you think that there is scope for improvement, feel free to open an issue.
Future price predictor and portfolio optimizer are ready independently. For predicting future price, following features have been used
1. Adjusted Close (Directly available from dataset)
2. Volume (Directly available from dataset)
3. Simple moving average (Computed explicitly)
4. Daily Returns (Computed explicitly)
1. Integrate portfolio optimizer in current project.
2. Add more flexibility. (make config.py more simple).
3. Write the documentation.
4. Write testcases.
Actual stock prices are shown in black color. Training fit is shown in blue and testing fit is shown in red.
The Walt Disney Company
Boeing Corporations