/Cryptocurrency-Price-Prediction

Primary LanguagePython

Cryptocurrency-Price-Prediction

Description

  • Implementation of cryptocurrency price prediction model based on CRNN(Convolutional Recurrent Neural Network).

Requirements

python>=3.7
torch
sklearn
matplotlib
tqdm

Usage

Train

  1. Edit config.json like below. Note that you need to set mode to train.
{
    "mode" : "train",
    "data_dir" : "../datasets",
    "output_dir" : "../resources",
    "weight_dir" : "",
    "batch_size" : 4,
    "epochs" : 30,
    "look_back" : 30,
    "hidden_dim" : 128,
    "num_layers" : 4
}
  1. Type the command below.
cd path_to_this_repository/src
python3 main.py

Test

  1. Edit config.json like below. Note that you need to set mode to test and set weight_dir appropriately. There is a pre-trained best model ../resources/weights_best.pt for hidden_dim = 128 and num_layers = 4.
{
    "mode" : "test",
    "data_dir" : "../datasets",
    "output_dir" : "../resources",
    "weight_dir" : "WEIGHT_DIR",
    "batch_size" : 4,
    "epochs" : 30,
    "look_back" : 30,
    "hidden_dim" : 128,
    "num_layers" : 4
}
  1. Type the command below.
cd path_to_this_repository/src
python3 main.py

Predict

  1. Edit config.json like below. Note that you need to set mode to predict and set weight_dir appropriately. There is a pre-trained best model ../resources/weights_best.pt for hidden_dim = 128 and num_layers = 4.
{
    "mode" : "predict",
    "data_dir" : "../datasets",
    "output_dir" : "../resources",
    "weight_dir" : "WEIGHT_DIR",
    "batch_size" : 4,
    "epochs" : 30,
    "look_back" : 30,
    "hidden_dim" : 128,
    "num_layers" : 4
}
  1. Type the command below.
cd path_to_this_repository/src
python3 main.py

Model Architecture

  • 3 conv1d layers -> num_layers-layerd LSTM -> 2 fully connected layers with ReLU
  • 3 conv1d layers
    • kernel_size = 3, padding = 1 to keep same size
    • in_channels = 1, hidden_dim, hidden_dim, respectively
    • out_channels = hidden_dim, hidden_dim, 1, respectively
    • ReLU activation function between all layers
  • num_layers-layerd LSTM
    • input_size = 1
    • hidden_size = hiddem_dim
    • dropout_rate = 0.3
  • 2 fully connected layers with ReLU
    • in_features = hidden_dim, 32, respectively
    • out_features = 32, 1, respectively
    • ReLU activation function between all layers

Experimental Results

Train

Predict

  • All models are trained using ethereum price dataset.
ethereum ethereum
(different look_back)		 bitcoin dogecoin
ethereum_to_ethereum ethereum_to_ethereum_different_lookback ethereum_to_bitcoin ethereum_to_dogecoin

Discussion

  • I first tried to train the model on the entire dataset, but it was very slow to train and the result was bad. So I trained only on ethereum and predicted the price of other cryptocurrencies.
  • The results were better than I thought. Trained model predicted the price of ethereum as well as other cryptocurrencies like bitcoin and dogecoin properly. You can see the result graph above. It also worked well with different look_backs.
  • It was possible to predict other cryptocurrencies not only because the model's performance is good, but also the prices of cryptocurrencies tend to follow each other.
  • It would be helpful to devise a general-purpose model architecture that can be learned for the entire dataset.

References