top of page

Model

The GAF-CNN model is capable of detecting Buy and Sell signals with over 70% average  accuracy. Its precisoin for the Buy singal class reaches 76%.

architecture.png

Overview

After GAF encoding, a 2D matrix is generated from each 1D time series data. The input data shape for the CNN classifier is (10, 10, F), where F is the number of features.  The CNN model consists of two convolutional layers following by maxpool layer and two fully connected layers. 

Morning Star/ Evening Star Detector

Model Parameters

  • Framework: Tensorflow

  • Model: GAF-CNN

  • Epoch: 300

  • Batch Size: 800

  • Dateset Size: 3600 (56% train, 14% validate, 30% test)

  • Optimizer: adam

  • Learning Rate Start: 0.001

Features Channels:

  • OHLC (open, high, low and close prices)

Label:

  • Morning star

  • Evening star

  • No pattern​

Evaluation:

  • Overall Accuracy: 90.65%

  • Precision for Morning Star: 90.2%

  • Precision for Evening Star: 93.18%

confusion_matrix_MEN_blues.png
confusion_matrix_BSN_blues.png

Buy/Sell Signal Detector

Model Parameters

  • Framework: Tensorflow

  • Model: GAF-CNN

  • Epoch: 100

  • Batch Size: 500

  • Dateset Size: 3600 (56% train, 14% validate, 30% test)

  • Optimizer: adam

  • Learning Rate Start: 0.001

Features Channels:

  • OHLC (open, high, low and close prices), volume

  • Reddit active users, reddit posts per hour

  • posts,total page views

Label:

  • Buy signal (positive morning star)

  • Sell signal (positive evening star)

  • Others ( ​negative morning star, negative evening star, no pattern)

Evaluation:

  • Overall Accuracy: 70.83%

  • Precision for Buy Signal: 76.11%

  • Precision for Sell Signal: 70.86%

Buy/Sell Signal Detector

Model Parameters

  • Framework: Tensorflow

  • Model: GAF-CNN

  • Epoch: 300

  • Batch Size: 500

  • Dateset Size: 2793 (56% train, 14% validate, 30% test)

  • Optimizer: adam

  • Learning Rate Start: 0.001

Features Channels:

  • OHLC (open, high, low and close prices), volume

  • News rank score, news text sentiment, news title sentiment

Label:

  • Buy signal (positive morning star)

  • Sell signal (positive evening star)

  • Others ( ​negative morning star, negative evening star, no pattern)

Evaluation:

  • Overall Accuracy: 68.97%

  • Precision for Buy Signal: 71.07%

  • Precision for Sell Signal: 75.58%

confusion_matrix_BSN_news_blues.png
bottom of page