Trading Strategy backtester, hyperparameter optimizer, and hidden markov model tester for crypto and stocks. Soon to add sentiment analysis.

• Python version 3.10.9 has been tested and works, other versions may also work. I used an Anaconda environment to manage the packages needed.

• Clone this repository in the directory of your choice.
• In the project directory with python and pip installed, enter pip install -r requirements.txt in terminal.

• First set up your configuration in Configuration/config.py. This includes making a .env file to store your api key to connect to the finnhub API.
• All strategies are kept in the strategies folder.
• Create a strategy file and follow the template for imports.
• Import any modules you may need for technical analysis. You can also create your own technical analysis functions. For more information, visit the documentation for ta.
• Follow the template to create the strategy function.
  • Step 1: Set the columns of any data you want to keep track of throughout the course of backtesting.
  • Step 2: Update your data (i.e. add on to price history and/or technical indicators data).
  • Step 3: Use your data to decide which orders to place on this tick.
• Now you can run 1. python Backtester.py download (if trading a cryptocurrency, otherwise skip to 2), then 2. streamlit run Backtester.py test. Make sure you are in the project directory in terminal, then enter python Backtester.py. On a localhost using Streamlit, it will generate a graph of position over profit/loss over close price, as well as log messages for the strategy in the logs folder and save the results to a csv.

• Add a graph function to Graphs.py
• This function must take in a pandas dataframe and return a plotly object that can be plotted after the PNL on streamlit using st.plotly_chart(your_figure)
• It is a good idea to, for whichever graph functions you plan to be using, assert that the data exists, as does the RSI example (the pandas dataframe that is passed in consists of all the data obtained from whichever strategy was run)
• At the end of the Graphs.py file, add to the graphs list whichever graphs you want and are applicable to the strategy to be backtested.
• Run Backtester.py with the argument test, as follows: streamlit run Backtester.py test. If you are testing with a cryptocurrency, this assumes you have downloaded the neccessary data for the given CRYPTO_INTERVAL and time range with python Backtester.py download. Otherwise, for stocks, finnhub allows for a decent amount of data to be queried on the spot (for the day interval or greater, a very large range can be queried (i.e. 2000+ days), and for an interval less than a day it is about a month maximum).

• this uses a module called optuna.
• First you must have a strategy that utilizes its params argument. For example, in Step 3 of creating a strategy (the rules to place orders on a given tick), you place a market buy order if params['rsi_buy_threshold'] < 30
• Next you must give a default value for these parameters in Configuration/Config.py under STRATEGY_HYPERPARAMETERS, and give a range for the parameters under STRATEGY_HYPERPARAMETER_RANGES.
• You can also specify the functions by which the parameters will be optimized by editing which functions are in the list optimization_functions found in OptimizeFunctions.py. For example, if using byProfit only, after optimizing hyperparameters, the trial with the highest profit will be logged. You can make your own hyper opt functions in HyperOpt/OptimizeFunctions.py
• You must also specify the number of trials under HYPER_OPT_TRIALS (more trials, better results and longer wait, vice versa) and the symbol, of the symbols to be traded, that will be optimized (SYMBOL_TO_OPTIMIZE).
• Then, from the project directory, 1. cd HyperOpt 2. python HyperparamOptimizer.py (assuming that, if using crypto, you have already downloaded data using python Backtester.py download)
• Logs will be saved to Logs/HyperOpt

• hmmlearn is used to implement Gaussian Hidden Markov Models. The point of this is to predict which state the market is in at any given timestep. This can help for things such as regime detection.
• If you want to include Hidden Markov Models in your strategy, there is a two step process: Train a model, then test the model.
• The first thing you must do to train the model is to make sure the configuration is set up properly
  • In Configuration/Config.py, set up the time range and interval you want to train the model on (DELTA_TIME, TO_DATE_LESS, and either INTERVAL (for stocks) or CRYPTO_INTERVAL (for crypto)). Note: Make sure to download data with python Backtester.py download in main project directory if using crypto.
  • Configure the NUMBER_OF_HIDDEN_STATES, COVARIANCE_TYPE, and NUMBER_OF_TRAINING_ITERATIONS that will be used to train the model.
  • Configure which functions will be used to train the model in HMMTraining/TrainingMethods.py, and add the ones you want to the training_methods list at the bottom of the file. You can follow the example of returns in creating your own training method. Note that when testing, strategy data will be used (i.e. the data stored by the strategy), and when training a model, just price data from api will be used.
• cd into HMMTraining and run python GenModel.py. This will generate a .pkl file of the model in HMMTraining/Models.
• Now that you have the model saved, find the name of the file in HMMTraining/Models and set MODEL_TO_USE to the name of that file without the .pkl extension
• Now you can run the strategy in the main project directory with streamlit run Backtester.py test. It is recommended to make use of the closeColoredByState function by adding it to the graphs list in Graphs.py to see the resulting hidden states predicted by the trained model. This is to see if the model is working as indended and which states mean what (i.e. if NUMBER_OF_HIDDEN_STATES is 2, what do states 0 and 1 represent?).

• You must create Configuration/.env and store your api key for finnhub there as described in Configuration/Config.py.
• DELTA_TIME is the range of time in days, while TO_DATE_LESS is how many days since today that the range will end. For example, with DELTA_TIME = 30 and TO_DATE_LESS = 5 the data that will be tested is 30 days long, ending 5 days ago.
• INTERVAL is for stocks, CRYPTO_INTERVAL is for cryptocurrencies.
• TAKER_FEE is for market orders, MAKER_FEE is for limit orders that don't immediately fill. These are in decimal percents (i.e. 0.01 = 1%)

• Add an optional numerical (or maybe even text-based) sentiment component to the State

ecco257

This project is licensed under the GNU General Public License v3.0 - see the LICENSE file for details