Add README

Author: haydenshively

README.md

@@ -1 +1,71 @@
-# Tezos-Prediction
+# Tezos Prediction
+
+This repository explores the use of technical indicators to predict the value
+of Tezos (XTZ) over time. Though technical analysis is often frowned upon in
+comparison with fundamental analysis, recent work
+[[1]](https://arxiv.org/pdf/1901.05237.pdf)
+[[2]](http://www.ieee-jas.org/fileZDHXBEN/journal/article/zdhxbywb/2020/3/PDF/JAS-2019-0392.pdf)
+has demonstrated the effectiveness of CNN's on market prediction tasks - without
+the need for extra-market data sources (e.g. sentiment analysis of headlines).
+
+Since early 2020, I've been collecting data on Bitcoin and Tezos every 6
+minutes. This includes:
+- Rank (based on Market Cap)
+- Market Cap
+- Price
+- 24 Hour Volume
+- Percent Change (last hour)
+- Percent Change (last week)
+
+You can download the dataset and trained models
+[here](https://drive.google.com/drive/folders/1m8Km28rn6RPEPKIJ_gAsp8aV6TJU2Sfq?usp=sharing).
+
+## Architecture
+The models I've released here look only at price over time. Specifically,
+they see 40 time steps and predict the next 5. At 6 minutes per time step,
+that's like knowing the past 4 hours and predicting the next half hour.
+
+The CNN architecture receives Grammian Angular Fields as input. This is
+basically a transformation from the time domain to the frequency domain
+in polar coordinates (at least that's how I think of it).The LSTM, on the
+other hand, see the raw time series.
+
+For comparison purposes, both networks train for 6 epochs with batch sizes
+of 16. Both use the mean squared error as their loss function, and are
+similar in terms of parameter count (277k vs 159k). Obviously, it's
+possible that optimizations could be made to improve the accuracy of each
+model, but I'm really just doing this as an exploratory exercise.
+
+## Findings
+Model | Mean Squared Error | Mean Absolute Error
+--- | :---: | :---: |
+CNN | 0.275 | 0.4458
+LSTM | 1.095 | 0.6432
+
+It's clear that the CNN with GAF images outperforms the LSTM on this dataset
+(at least with these hyperparameters). This result is in line with recent
+research in the field.
+
+## Prerequisites
+To run this code, you will need `numpy`, `scipy`, and `tensorflow.keras`. I
+recommend installing via a package manager like conda, but that's up to you.
+
+## Usage
+Download the dataset from the link above, unzip it, and place the 'dataset'
+folder in the repository's root directory (on the same level as 'models',
+'testing', 'training', and 'transforms').
+
+Once you have the dataset, you can replicate my results by calling `python
+main.py`, which will train (1) a CNN on GAF images and (2) an LSTM on 1D
+sequences. It will then save the model weights to disk and evaluate both
+architectures on the test set.
+
+If you want to dig deeper or run your own experiments, start with the files
+in the 'training' and 'testing' folders. You'll be able to modify all of the
+hyperparameters. If you want to change the architecture itself, look in the
+'models' folder.
+
+## Disclaimer
+I don't recommend using these models in combination with any sort of trading
+bot or trading strategy. This work is for research purposes only. Use at your
+own risk.

main.py

@@ -0,0 +1,15 @@
+from training.train_cnn_timeseries import main as train_cnn_timeseries
+from training.train_lstm import main as train_lstm
+from testing.test_cnn_timeseries import main as test_cnn_timeseries
+from testing.test_lstm import main as test_lstm
+
+
+if __name__ == '__main__':
+    # train_cnn_timeseries('dataset/train')
+    # train_lstm('dataset/train')
+
+    cnn_results = test_cnn_timeseries('models/cnn_timeseries_16_40_5_1.h5', 'dataset/test')
+    lstm_results = test_lstm('models/lstm_16_40_5.h5', 'dataset/test')
+
+    print('CNN Results \t|\t\tmse: {},\t\tmae: {}'.format(*cnn_results[:2]))
+    print('LSTM Results\t|\t\tmse: {},\t\tmae: {}'.format(*lstm_results[:2]))

models/lstm.py

@@ -32,5 +32,5 @@ def build(self):
         super().build()
 
     def compile(self):
-        self.model.compile(optimizer='sgd', loss='mse', metrics=['mae'])
+        self.model.compile(optimizer='sgd', loss='mse', metrics=['mae', 'accuracy'])
         super().compile()

predict.py

@@ -0,0 +1,26 @@
+from tensorflow.keras import models
+
+from training.train_cnn_timeseries import MyData
+
+
+def main(model_dir, data_dir):
+    BATCH_SIZE = 16
+    N_SAMPLES_IN = 40  # divide by 10 to get # hours the sequence covers
+    N_SAMPLES_OUT = 5
+    PROB_DISTRIB = 1
+
+    generator = MyData(
+        data_dir,
+        BATCH_SIZE,
+        N_SAMPLES_IN,
+        N_SAMPLES_OUT,
+        PROB_DISTRIB
+    )
+
+    cnn = models.load_model(model_dir)
+    cnn.summary()
+    return cnn.evaluate(generator, verbose=1)
+
+
+if __name__ == '__main__':
+    print(main('../models/cnn_timeseries_16_40_5_1.h5', '../dataset/test'))

see_results.py

@@ -0,0 +1,24 @@
+from tensorflow.keras import models
+
+from training.train_lstm import MyData
+
+
+def main(model_dir, data_dir):
+    BATCH_SIZE = 16
+    N_SAMPLES_IN = 40  # divide by 10 to get # hours the sequence covers
+    N_SAMPLES_OUT = 5
+
+    generator = MyData(
+        data_dir,
+        BATCH_SIZE,
+        N_SAMPLES_IN,
+        N_SAMPLES_OUT
+    )
+
+    lstm = models.load_model(model_dir)
+    lstm.summary()
+    return lstm.evaluate(generator, verbose=1)
+
+
+if __name__ == '__main__':
+    print(main('../models/lstm_16_40_5.h5', '../dataset/test'))

testing/test_cnn_timeseries.py

@@ -1,6 +1,3 @@
-"""
-TRAIN
-"""
 import os
 
 from scipy import stats
@@ -65,16 +62,14 @@ def __getitem__(self, idx):
         return X, Y
 
 
-if __name__ == '__main__':
-
-    DATA_DIR = '../dataset/train'
+def main(data_dir):
     BATCH_SIZE = 16
     N_SAMPLES_IN = 40  # divide by 10 to get # hours the sequence covers
     N_SAMPLES_OUT = 5
     PROB_DISTRIB = 1
 
     generator = MyData(
-        DATA_DIR,
+        data_dir,
         BATCH_SIZE,
         N_SAMPLES_IN,
         N_SAMPLES_OUT,
@@ -95,6 +90,10 @@ def __getitem__(self, idx):
                   verbose=1,
                   steps_per_epoch=len(generator))
 
-    cnn.model.save('cnn_%d_%d_%d_%d.h5' % (
+    cnn.model.save('models/cnn_timeseries_%d_%d_%d_%d.h5' % (
         BATCH_SIZE, N_SAMPLES_IN, N_SAMPLES_OUT, PROB_DISTRIB
     ))
+
+
+if __name__ == '__main__':
+    main('../dataset/train')

testing/test_lstm.py

@@ -1,6 +1,3 @@
-"""
-TRAIN
-"""
 import os
 
 import numpy as np
@@ -57,15 +54,13 @@ def __getitem__(self, idx):
         return X, Y
 
 
-if __name__ == '__main__':
-
-    DATA_DIR = '../dataset/train'
+def main(data_dir):
     BATCH_SIZE = 16
     N_SAMPLES_IN = 40  # divide by 10 to get # hours the sequence covers
     N_SAMPLES_OUT = 5
 
     generator = MyData(
-        DATA_DIR,
+        data_dir,
         BATCH_SIZE,
         N_SAMPLES_IN,
         N_SAMPLES_OUT
@@ -81,6 +76,10 @@ def __getitem__(self, idx):
                    verbose=1,
                    steps_per_epoch=len(generator))
 
-    lstm.model.save('lstm_%d_%d_%d.h5' % (
+    lstm.model.save('models/lstm_%d_%d_%d.h5' % (
         BATCH_SIZE, N_SAMPLES_IN, N_SAMPLES_OUT
     ))
+
+
+if __name__ == '__main__':
+    main('../dataset/train')