Seahorse Strategies

Table of Contents

• Contributors
• Overview
• Main Scripts
• Repository Structure
• Software Installation
• Usage
• Functions
• Package Dependencies

---

Contributors:

Name	Simon Chiu	Gilbert Lei	Fan Wu	Linyang Yu
User Id	cheukman1207	gilbertlei	fwu03	lyyu0413

Overview

Seahorse Strategies Inc. developed a system, the Oscillator, that tracks stock price movements and generates buy and sell signals when certain conditions are met. Statistical analysis shows that these signals are not always reliable and there is room for improvements. This repo's program is designed for evaluating each buy/sell signal generated by this system and identify good ones on which the company may choose to trade. The model developed in this program is a supervised learning model based on the LightGBM framework. It can efficiently evaluate a huge number of signals and score each signals. A signal with a score of 90 or 100 has an expected return 2.5 - 5 times higher than that of a randomly selected signal.

Main Scripts

Train.py

This function performs model training that applies the LightGBM Regressor to the data in the Train_data folder. It then saves the model for prediction later.

Predict.py

This function performs prediction on the returns of the signals in the Prediction folder and provides a score for each signal to indicate if it is likely to be a gain or a loss. It takes the choice of a saved model generated by the Train function as an argument.

Repository Structure:

DSCI_591_capstone-Seahorse

• Functions & Unit Testing: folder contains all the required functions and their unit testing
  • 01_data_import.py
  • 02_data_organize.py
  • 03_smooth_generator.py
  • 04_derivative.py
  • 05_volatility.py
  • 06_ratio.py
• Models: folder to store trained models
• Prediction: folder to store prediction results
• Test_data: folder that contains all the data to be evaluated
• Train_data: folder that contains all the data for model training
• Training_report: folder to store reports of cross validation results
• Others: folder to store all the other works for this project
  • Documentations
  • Images
  • Scripts
• CODE_OF_CONDUCT.md
• Predict.py
• Train.py
• README.md

The following screenshot shows the workflow of the system:

Software Installation

Software required to run the system:

• Git
• Python 3

Git

The command line version of Git will be used here.

[Windows]

1. Click here to download the windows version git. After the download is completed, run the installer and select the following options:
   • On the Select Components page, check “On the Desktop” under “Additional icons”
   • On the Choosing the default editor used by Git page, select “Use the Nano editor by default” from the drop-down menu
   • For all other pages, use the default options
2. After installation, a Git Bash icon will be on the Desktop

[macOS]

1. Open Terminal and type the following command:

xcode-select --install

2. Check the Git version by:

git --version

Python 3

The system uses the Python 3, not Python 2. Anaconda is an easy-to-install distribution of Python and includes many popular libraries.

1. Please download the Python 3 for Windows or macOS
2. After the installation is completed, follow the instruction to run the installer
3. To check the Python version, you can type the following command in the terminal:

python --version

4. If you were successful, you will see something like this:

Python 3.7.1

For more details, please see MDS's document on install guide

Usage

Step 1. Clone Repository from Github

Click the green Clone or Download button, and copy the URL to the clipboard.

Open the Unix Shell instance (e.g. Terminal or Git for Windows), and navigate to your target directory by typing:

cd (path to the directory)

Type git clone and paste the URL:

git clone https://github.com/UBC-MDS/DSCI_591_capstone-Seahorse.git

Press enter to clone the repository to your local machine.

For more details, please see GitHub's document on cloning

Step 2. Load Train and Test Data

Load train data into Train_data folder, and test data into Test_data folder. Please ensure the following:

• Train data must include 124 columns; 41 for oscillator values, 41 for stock prices, 41 for macd values, and 1 for investment return
• Test data must include 123 columns; 41 for oscillator values, 41 for stock prices and 41 for macd values
• File names must contain Buy or Sell to indicate whether it is a set of buy or sell signals
• Files must be in .txt format

Step 3. Run Train.py Script

Open Unix Shell instance and navigate to the cloned repository by:

cd (the path to the repo)

and then enter:

python Train.py

The process may take a few minutes to run. After it completes, four files will be generated with the current date and time:

1. A training report in the Training_report folder (Train_Report_yyyy_mm_dd_hhmm.pdf)
2. A trained model in the Model folder (Model_yyyy_mm_dd_hhmm.sav)
3. A rubric for prediction purpose in the Training_report/rubric folder (Train_Rubric_yyyy_mm_dd_hhmm.csv)
4. A screenshot for the training report in the Training_report/img folder (Cross_Validation_Result_Boxplot_yyyy_mm_dd_hhmm.png)

Step 4. Training Report Review

The training report is in the Training_report folder. It contains two cross validation tables (mean and standard deviation) and a boxplot.

The two tables contains parameters (mean and standard deviation) to describe the signals with scores in a certain group, averaging over all the cross validation run.

Column Name	Meaning
score_rd	Signals grouped by the score assigned by the model
StkUpRate	Ratio of signals with positive returns in the group (0:0% - 1:100%)
Stk Mvt%	Average return of the signals in the group
BuySig%	Ratio of signals that is a buy signal
% of All Trades	Ratio of total signal in the group (0:none - 1:all signals in this group)

Boxplot: The boxplot shows the average returns of each score group, each point represent the average return from one cross-validation run. A wider box means the signals in that group is relatively risker, and vice versa.

Once you find a statisfied report, the corresponding model can be used for prediction.

Step 5. Prediction

Open the Unix Shell instance, navigate to the cloned repository, and enter:

python Prediction.py

You will see the following screen:

Enter the date and time of the selected model (e.g. 2019_06_22_1123)

The prediction result will be generated under the Prediction folder (Prediction_yyyy_mm_dd_hhmm.csv) with a timestamp follows by the selected model. The last column of the file, Score_rd, is the predicted resulted for the test data.

Step 6. Clean Files (Optional)

Open the Unix Shell instance, navigate to the cloned repository, and enter:

python Clean.py

This operation will clean all files that are generated by both Train.py and Prediction.py.

Functions

We developed six functions that are called by the train.py and prediction.py programs.

1. Data Import: Load (train/test) data into the system
2. Data Organize: Break the data into multiple groups such as oscillator(osc), stock price(stk), macd
3. Smooth Generator: Calculate the smoothness of a curve
4. Derivative: Calculate the derivative of a curve (either absolute or relative change)
5. Volatility: Calculate the volatility of a curve
6. Ratio: Calculate the ratios between the values of two curves

Package Dependencies

Software	Version
Git	2.17.2
Python	3.7.1

Package	Version	Package	Version
numpy	1.16.2	matplotlib	3.0.3
pandas	0.24.2	seaborn	0.9.0
scipy	1.2.1	reportlab	3.5.23
lightgbm	2.2.3	pickle	1.0.2
sklearn	0.20.3	datetime	4.3