Instructions to train a Python trading bot with Kraken

Aim of the project

This project presents how the Kraken trading bot I previously developed was trained and how the optimal set of parameters was selected.

Disclaimer

This is an educational project, not a real investment strategy or advice, and I take no responsibility for any financial losses or misuse of this code.

Dependencies

Requirements

• Python 3.6 or Higher
• Kraken's Python API client krakenex (set-up described below)
• Script algorithm.py
• Script evaluate_algorithm.py
• The cryptocurrency price history data stored in a parent folder

First set-up

1. Install virtualenv if it is not already installed

pip install virtualenv

2. Create virtual environment in your folder of interest

virtualenv krakenex

3. Activate the virtual environment

source krakenex/bin/activate

4. Install the required libraries in the virtual environment using the requirements.txt file

pip install -r requirements.txt

Reactivate the virtual environment each time you need to use the trading bot

source krakenex/bin/activate

General project structure

Input data

The cryptocurrency price history data used to simulate trading was stored in the parent folder and was made of the complete trade history available on Kraken. Please refer to my crypto history extraction project if you want to extract the history data on your own.

Script organization

Algorithm benchmarking script

The scripts algorithm4.py and algorithm5.py were provided as examples of training codes, but in practice, several other scripts were used with different time frames and parameter sets, to test several different algorithms and algorithm versions, as detailed in Code_Specifications.txt. The present instructions will only detail the use of the script algorithm5.py.

Utility package

The modules utilities.py and query_public_info.py from the package tradingbot are used by the scripts algorithm4.py and algorithm5.py to provide different utility functions.

Benchmark result evaluation script

The script evaluate_algorithm.py processes the output files of the algorithm benchmark and selects the optimal set of parameters for profit generation.

Output data

• pipeline_parameters.txt: File that contains the different parameter combinations for the algorithm benchmark. Generated by algorithm5.py
• $CRYPTO_records.txt: File that contains the algorithm trading decision history for the cryptocurrency $CRYPTO. Generated by algorithm5.py
• algorithm5_ranked_results.txt: File that contains the algorithm profit results for each parameter combination. Generated by evaluate_algorithm.py
• algorithm5_benefit.png: Graph that shows the average gains / losses / net benefit per day for the top 10 parameter sets. Generated by evaluate_algorithm.py
• algorithm5_total.png: Graph that shows the total gains / losses / net benefit for the top 10 parameter sets. Generated by evaluate_algorithm.py
• $CRYPTO_balance_evolution.png: Evolution of the base currency balance over time based on the trading results for the cryptocurrency $CRYPTO. Generated by evaluate_algorithm.py

Settings and requirements

Fiat base currency selection

You will need to specify the fiat base currency as the value of the variable base_currency in the script algorithm5.py which is set to ZEUR by default which stands for the Euro.

List of cryptocurrencies

You can define the list of cryptocurrencies you want the trading bot to speculate on with the variable currencies in the script algorithm5.py. The pairs should also be specified with the variable all_pairs.

Minimal balance

The available base currency balance per cryptocurrency is defined by the variable funds.

Algorithm parameters

The following parameters were assigned various values, and all combinations of these values were tested by simulating trades:

• drop_rate: Price drop percentage value over a drop_time_frame that will trigger a buy order
• drop_time_frame: History time frame on which the price drop is assessed, e.g. a drop_rate of 2% will trigger the buy condition only if the price dropped by at least 2% on this specified period of time
• end_tendency: A price dropping tendency will be considered over after the price raised of this specific percentage. E.g. if the price drops for a while and then increases of 2%, the dropping tendency will be considered as finished. The buy condition will be triggered only after the dropping tendency finishes
• min_benef: The trading bot will only sell a cryptocurrency if it made at least this percentage of benefit
• history_regression: History time frame on which the price trend and direction is assessed

Launch the cryptocurrency trading bot with all the parameter combinations on the cryptocurrency price history

A simple Python command will launch the benchmark:

python3 algorithm5.py  # To test the algorithm 5

The benchmark script tests several parameter combinations over a large history. Therefore, I recommend running it in a detached terminal state. You can use the terminal multiplexer Screen for this purpose.

Generate metrics and plots to rank the different sets of parameters

The following command uses the output of the previous benchmark to generate graphs and statistics, thereby enabling to select the most optimal parameter values for the selected algorithm:

python3 evaluate_algorithm.py

Results

Algorithm selection

The overall best strategy is to select the algorithm with the highest total net benefit per day. In the case of this project, the algorithm was developed during a bull run, and trained on a short bull run period in order to optimize the short term profits. The algorithm is therefore designed to run only during a bull run period, as it focuses on high fluctuations. If you want more details regarding my algorithm development history and strategy, you can examine the file Code_Specifications.txt.

Parameter selection

Based on the graph algorithm5_benefit.png the most optimal parameter sets can be selected:

Based on the graphs per crypto, as shown in the below example, algorithms that tend to trade more regularly can be selected: The purpose of selecting an algorithm that trades more regularly is to adapt the strategy to the short duration of a bull run.

Project Timeline

• Start Date: September 2020
• Completion Date: December 2020
• Maintenance status: Inactive