learning_rate = 0.001
batch_size = 16
display_step = 100
max_epochs = 10
symbols = ['AAPL', 'AMZN'] #, AAPL, 'GOOG', 'GOOGL', 'FB', 'AMZN']
n_stocks = len(symbols)
n_hidden1 = 128
n_hidden2 = 128
n_steps_encoder = 20 # time steps, length of time window
n_output = n_stocks
T = 30
# training data
dset = SP500('data/sandp500/individual_stocks_5yr',
symbols=symbols,
start_date='2013-01-01',
end_date='2013-07-31',
T=T,
step=1)
train_loader = DataLoader(
dset,
batch_size=batch_size,
shuffle=False,
num_workers=4,
pin_memory=True # CUDA only
)
x, y = train_loader.dataset[0]
print(x.shape)
# Network Parameters
model = LSTM(hidden_size=128,
hidden_size2=300,
num_securities=n_stocks,
def generate_pictures(train_loader, n_features, use_scaler, symbols,
train_split, n_epoch):
predictions = np.zeros((len(train_loader.dataset.chunks), n_features))
ground_tr = np.zeros((len(train_loader.dataset.chunks), n_features))
batch_size_pred = 4
# Create test data set
dtest = SP500("../data/merge.csv",
T=T,
step=1,
scr=use_scaler,
train_split=0)
test_loader = DataLoader(
dtest,
batch_size=batch_size_pred,
shuffle=False,
num_workers=4,
pin_memory=True # CUDA only
)
# Create list of n_stocks lists for storing predictions and GT
predictions = [[] for i in repeat(None, len(symbols))]
gts = [[] for i in repeat(None, len(symbols))]
k = 0
# Predictions
for batch_idx, (data, target) in enumerate(test_loader):
data = Variable(data.permute(1, 0, 2)).contiguous()
target = Variable(target)
if use_cuda:
data = data.cuda()
target = target.cuda()
if target.data.size()[0] == batch_size_pred:
output = model(data)
for k in range(batch_size_pred):
s = 0
for stock in symbols:
# print("output.data")
predictions[s].append(output.data[k])
# print(target.data)
gts[s].append(target.data[k])
s += 1
k += 1
if len(symbols) == 1:
if (use_scaler == True):
pred = dtest.scaler.inverse_transform(
np.array(predictions[0]).reshape((len(predictions[0]), 1)))
gt = dtest.scaler.inverse_transform(
np.array(gts[0]).reshape(len(gts[0]), 1))
else:
pred = np.array(predictions[0]).reshape((len(predictions[0]), 1))
gt = np.array(gts[0]).reshape(len(gts[0]), 1)
if len(symbols) >= 2:
if (use_scaler == True):
p = np.array(predictions)
pred = dtest.scaler.inverse_transform(
np.array(predictions).transpose())
gt = dtest.scaler.inverse_transform(np.array(gts).transpose())
else:
pred = np.array(predictions).transpose()
gt = np.array(gts).transpose()
# Plots for all stocks in list of symbols
# x = np.array(range(pred.shape[0]))
# x = [np.datetime64(start_date) + np.timedelta64(x, 'D') for x in range(0, pred.shape[0])]
# x = np.array(x)
# months = MonthLocator(range(1, 10), bymonthday=1, interval=1)
# monthsFmt = DateFormatter("%b '%y")
s = 0
print("image {} gt[:,0].shape = " + str(gt[:, 0].shape))
for stock in symbols:
# fig, ax = plt.subplots(figsize=(8, 6), dpi=100)
plt.ioff()
plt.figure()
plt.plot(range(gt[:, s].shape[0]), gt[:, s], label="True")
plt.plot(range(0, int(pred[:, s].shape[0] * (1 - train_split))),
pred[:, s][0:int(pred[:, s].shape[0] * (1 - train_split))],
label='Predicted - Train')
plt.plot(range(int(pred[:, s].shape[0] * (1 - train_split)),
pred[:, s].shape[0]),
pred[:, s][int(pred[:, s].shape[0] * (1 - train_split)):],
label='Predicted - Train')
plt.title('LSTM_Epoch_{}'.format(n_epoch))
plt.legend(loc='upper left')
plt.savefig("../fig/" + stock + str(n_epoch) + '.png')
n_hidden1 = 128
n_hidden2 = 128
n_steps_encoder = 20 # time steps, length of time window
n_output = n_stocks
T = 10
n_step_data = 10
start_date = '2013-01-01'
end_date = '2013-12-31'
fn_base = "lstm_nstocks_" + str(n_stocks) + "_epochs_" + str(max_epochs) + "_T_" + str(T) + "_train_" + \
start_date + "_" + end_date
# training data
dset = SP500('data/sandp500/individual_stocks_5yr',
symbols=symbols,
start_date=start_date,
end_date=end_date,
T=T,
step=1)
train_loader = DataLoader(
dset,
batch_size=batch_size,
shuffle=False,
num_workers=4,
pin_memory=True # CUDA only
)
# Network Definition + Optimizer + Scheduler
model = LSTM(hidden_size=n_hidden1,
hidden_size2=n_hidden2,
num_securities=n_stocks,
dropout=0.2,
display_step = 50
max_epochs = 100
n_features = 116 #num of features
n_hidden1 = 500
n_hidden2 = 500
T = 10
start_date = '1999-11-10'
use_scaler = False
symbols = ['600000.SH']
fn_base = "lstm_nstocks_" + str(n_features) + "_epochs_" + str(
max_epochs) + "_T_" + str(T) + "_train"
# training data
dset = SP500("../../data/merge.csv",
T=T,
step=1,
scr=use_scaler,
train_split=0.3)
train_loader = DataLoader(
dset,
batch_size=batch_size,
shuffle=True,
num_workers=4,
pin_memory=True # CUDA only
)
# Network Definition + Optimizer + Scheduler
model = LSTM(hidden_size=n_hidden1,
hidden_size2=n_hidden2,
num_securities=n_features,
dropout=0.2,