

In [1]:

    
%matplotlib inline
import pandas as pd
import numpy as np
import tensorflow as tf
tf.reset_default_graph()
from tqdm import tqdm
import matplotlib.pyplot as plt
import seaborn as sns
sns.set()

Actions: Buy & Sell
Utility for each day determined by Open & Close.

Loading Data and getting utility values



In [2]:

    
def find_utils(row):
    if row['Adj Close'] - row['Open'] < 0: return -1
    else: return 1



In [3]:

    
df = pd.read_csv('data/GOOGL.csv')
df['Utility'] = df.apply(find_utils, axis=1)

data = df.drop(['Utility', 'Date'], 1).values
target = df['Utility'].values

a = plt.plot(df['Open'])
b = plt.plot(df['Adj Close'])
plt.title('GOOGL')
plt.legend((a[0], b[0]), ('Open', 'Adj Close'), loc=1, bbox_to_anchor=(1.35, 1))
plt.show()
plt.close()
df.head()

Formatting Data for LSTM Network



In [5]:

    
features=[]
for i in data:
    temp_list = []
    for j in i:
        temp_list.append([j])
    features.append(np.array(temp_list))

labels = []
for i in target:
    if i == -1: labels.append([1, -1])
    else: labels.append([-1, 1])

LSTM

To keep the notebook short lstm network is implemented in ./lstm.py



In [5]:

    
from lstm import LSTM



In [7]:

    
epoch = 35000
num_cells=100

features_test = features[3000:]
labels_test = features[3000:]
features = features[:3000]
labels = labels[:3000]
len(features_test)

model = LSTM(num_hidden=num_cells)
predictions, train_error, test_error = model.run(features, labels, epoch=epoch, test_features=features_test,
                                                test_labels=labels_test)
plt.title('GOOGL - Mean Error Per Batch')
plt.plot(train_error)
plt.show()
plt.close()
print('Error on Testing Datset:', test_error)
model.reset()









    



/Users/jarvis/anaconda/lib/python3.6/site-packages/tensorflow/python/ops/gradients_impl.py:93: UserWarning: Converting sparse IndexedSlices to a dense Tensor of unknown shape. This may consume a large amount of memory.
  "Converting sparse IndexedSlices to a dense Tensor of unknown shape. "
Epoch: 100%|██████████| 35000/35000 [1:59:55<00:00,  4.88it/s]  






    












    



Error on Testing Datset: 0.189



In [ ]:

Berkshire Hathaway



In [6]:

    
df = pd.read_csv('data/BRK-A.csv')
df['Utility'] = df.apply(find_utils, axis=1)

data = df.drop(['Utility', 'Date'], 1).values
target = df['Utility'].values

a = plt.plot(df['Open'])
b = plt.plot(df['Adj Close'])
plt.title('BRK-A')
plt.legend((a[0], b[0]), ('Open', 'Adj Close'), loc=1, bbox_to_anchor=(1.35, 1))
plt.show()
plt.close()

features=[]
for i in data:
    temp_list = []
    for j in i:
        temp_list.append([j])
    features.append(np.array(temp_list))

labels = []
for i in target:
    if i == -1: labels.append([1, 0])
    else: labels.append([0, 1])

df.head()



In [7]:

    
epoch = 35000
num_cells=100

features_test = features[8500:]
labels_test = features[8500:]
features = features[:8500]
labels = labels[:8500]
len(features_test)

model = LSTM(num_hidden=num_cells)
predictions, train_error, test_error = model.run(features, labels, epoch=epoch, test_features=features_test,
                                                test_labels=labels_test)
plt.title('BRK-A - Mean Error Per Batch')
plt.plot(train_error)
plt.show()
plt.close()
print('Error on Testing Datset:', test_error)
model.reset()









    



/Users/jarvis/anaconda/lib/python3.6/site-packages/tensorflow/python/ops/gradients_impl.py:93: UserWarning: Converting sparse IndexedSlices to a dense Tensor of unknown shape. This may consume a large amount of memory.
  "Converting sparse IndexedSlices to a dense Tensor of unknown shape. "
Epoch: 100%|██████████| 35000/35000 [5:45:22<00:00,  1.71it/s]  






    












    



Error on Testing Datset: 0.526

Johnson and Johnson



In [3]:

    
df = pd.read_csv('data/JNJ.csv')
df['Utility'] = df.apply(find_utils, axis=1)

data = df.drop(['Utility', 'Date'], 1).values
target = df['Utility'].values

a = plt.plot(df['Open'])
b = plt.plot(df['Adj Close'])
plt.title('JNJ')
plt.legend((a[0], b[0]), ('Open', 'Adj Close'), loc=1, bbox_to_anchor=(1.35, 1))
plt.show()
plt.close()

features=[]
for i in data:
    temp_list = []
    for j in i:
        temp_list.append([j])
    features.append(np.array(temp_list))

labels = []
for i in target:
    if i == -1: labels.append([1, 0])
    else: labels.append([0, 1])

df.head()



In [16]:

    
epoch = 35000
num_cells=25

features_test = features[10000:]
labels_test = features[10000:]
features = features[:10000]
labels = labels[:10000]
len(features_test)

model = LSTM(num_hidden=num_cells)
predictions, train_error, test_error = model.run(features, labels, epoch=epoch, test_features=features_test,
                                                test_labels=labels_test)
plt.title('JNJ - Mean Error Per Batch')
plt.plot(train_error)
plt.show()
plt.close()
print('Error on Testing Datset:', test_error)









    



/Users/jarvis/anaconda3/lib/python3.5/site-packages/tensorflow/python/ops/gradients_impl.py:91: UserWarning: Converting sparse IndexedSlices to a dense Tensor of unknown shape. This may consume a large amount of memory.
  "Converting sparse IndexedSlices to a dense Tensor of unknown shape. "
Epoch: 100%|██████████| 35000/35000 [3:35:40<00:00,  2.95it/s]  






    












    



Error on Testing Datset: 0.509



In [8]:

    
model.reset()

Increase number of cells to see if that yeilds better results



In [9]:

    
epoch = 40000
num_cells=50

model = LSTM(num_hidden=num_cells)
predictions, train_error, test_error = model.run(features, labels, epoch=epoch, test_features=features_test,
                                                test_labels=labels_test)
plt.title('JNJ - Mean Error Per Batch')
plt.plot(train_error)
plt.show()
plt.close()
print('Error on Testing Datset:', test_error)









    



/Users/jarvis/anaconda3/lib/python3.5/site-packages/tensorflow/python/ops/gradients_impl.py:91: UserWarning: Converting sparse IndexedSlices to a dense Tensor of unknown shape. This may consume a large amount of memory.
  "Converting sparse IndexedSlices to a dense Tensor of unknown shape. "
Epoch: 100%|██████████| 40000/40000 [7:09:09<00:00,  1.89it/s]  






    












    



Error on Testing Datset: 0.186



In [ ]:

	Date	Open	High	Low	Close	Adj Close	Volume	Utility
0	2004-08-19	50.050049	52.082081	48.028027	100.339996	50.220219	44659000	1
1	2004-08-20	50.555557	54.594593	50.300301	108.310005	54.209209	22834300	1
2	2004-08-23	55.430431	56.796795	54.579578	109.400002	54.754753	18256100	-1
3	2004-08-24	55.675674	55.855854	51.836838	104.870003	52.487488	15247300	-1
4	2004-08-25	52.532532	54.054054	51.991993	106.000008	53.053055	9188600	1

	Date	Open	High	Low	Close	Adj Close	Volume	Utility
0	1980-03-17	290.0	310.0	290.0	290.0	290.0	10000	1
1	1980-03-18	290.0	290.0	290.0	290.0	290.0	0	1
2	1980-03-19	290.0	310.0	290.0	290.0	290.0	20000	1
3	1980-03-20	290.0	290.0	290.0	290.0	290.0	0	1
4	1980-03-21	290.0	290.0	290.0	290.0	290.0	0	1

	Date	Open	High	Low	Close	Adj Close	Volume	Utility
0	1970-01-02	1.250000	1.253472	1.218750	0.453297	1.218750	1195200	-1
1	1970-01-05	1.218750	1.222222	1.187500	0.442320	1.189236	964800	-1
2	1970-01-06	1.189236	1.211806	1.184028	0.450714	1.211806	1036800	1
3	1970-01-07	1.211806	1.215278	1.203125	0.448777	1.206597	331200	-1
4	1970-01-08	1.206597	1.225694	1.206597	0.452006	1.215278	460800	1