-
Notifications
You must be signed in to change notification settings - Fork 5
/
portfolio_opt_experiment.py
183 lines (153 loc) · 8.37 KB
/
portfolio_opt_experiment.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
import matplotlib
matplotlib.use('Agg')
from torch.utils.tensorboard import SummaryWriter
from src.portfolio_opt_methods import *
from src.portfolio_opt_data import PortDataset
######################################
# Hyperparameters #
######################################
START_EPOCH = 0
NUM_EPOCHS = 10000
HISTORY_LEN = 120
FUTURE_LEN = 120
NUM_FEATURE = 32
DEVICE = torch.device('cuda:0')
LR = 1e-3
LR_STEPS = [1500]
RF = 0.03 # risk-free return to compute the Sharpe ratio
K = 20 # for the cardinality (topK) constraint
######################################
# training #
######################################
def train_test_portfolio(model, train_set, mse_weight=1, sharpe_weight=1, opt_method='predict-and-opt', test_set=None, test_items='all'):
assert mse_weight > 0 or sharpe_weight > 0
working_lr_steps = []
for step in LR_STEPS:
if step - START_EPOCH < 0:
continue
working_lr_steps.append(step - START_EPOCH)
optimizer = torch.optim.Adam(model.parameters(), lr=LR)
lr_scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer, working_lr_steps)
writer = SummaryWriter()
for epoch in range(START_EPOCH, NUM_EPOCHS):
train_num = len(train_set)
epoch_mse = 0
epoch_sharpe = 0
for iter_idx, train_data in enumerate(np.random.permutation(train_set)):
# prepare data
history = torch.tensor(train_data['history'].values).to(device=DEVICE, dtype=torch.double)
future = torch.tensor(train_data['future'].values).to(device=DEVICE, dtype=torch.double)
# forward pass
if sharpe_weight > 0:
pred_seq, weight = model(history, FUTURE_LEN, opt_method, RF, K,
gumbel_sample_num=10 if opt_method == 'jpo-old' else 1000)
else:
pred_seq = model(history, FUTURE_LEN)
# compute loss
loss = 0
mse = torch.sum((pred_seq - future) ** 2) / FUTURE_LEN
writer.add_scalar('train_mse', mse.detach(), iter_idx + epoch * train_num)
epoch_mse += mse.detach() / train_num
if mse_weight > 0:
loss += mse_weight * mse
if sharpe_weight > 0:
mu, cov = compute_measures(future)
sharpe = compute_sharpe_ratio(mu, cov, weight, RF)
writer.add_scalars(
'train_sharpe',
{'mean': sharpe.mean().detach(), 'min': sharpe.min().detach(), 'max': sharpe.max().detach()},
iter_idx + epoch * train_num
)
loss += - sharpe_weight * sharpe.mean()
epoch_sharpe += sharpe.mean().detach() / train_num
loss.backward()
optimizer.step()
optimizer.zero_grad()
lr_scheduler.step()
writer.add_scalar('epoch_mse', epoch_mse, (epoch + 1) * train_num)
if sharpe_weight > 0:
writer.add_scalar('epoch_sharpe', epoch_sharpe, (epoch + 1) * train_num)
print(f'e={epoch}/{NUM_EPOCHS}, loss={mse_weight * epoch_mse - sharpe_weight * epoch_sharpe:.6f}')
if epoch % 1 == 0:
torch.save(model.state_dict(), f'output/portfolio_lstm_epoch{epoch}_{opt_method}.pt')
if epoch % 1 == 0 and test_set is not None:
test_result = test_portfolio(model, test_set, test_items, verbose=True)
writer.add_scalars('test',
{k: sum(v) / len(v) for k, v in test_result.items()},
(epoch + 1) * train_num)
######################################
# testing #
######################################
def test_portfolio(model, test_set, test_items='all', verbose=True):
with torch.set_grad_enabled(False):
if test_items == 'all':
test_items = ['mse', 'predict-and-opt', 'history-opt', 'predict-then-opt']
return_dict = {k + '-return': [] for k in set(test_items)-set(['mse'])}
return_dict.update({k + '-risk' : [] for k in set(test_items)-set(['mse'])})
if 'mse' in test_items:
return_dict.update({'mse' : []})
for iter_idx, test_data in enumerate(test_set):
history = torch.tensor(test_data['history'].values).to(device=DEVICE, dtype=torch.double)
future = torch.tensor(test_data['future'].values).to(device=DEVICE, dtype=torch.double)
mu, cov = compute_measures(future)
test_print = [f'test id={iter_idx}, date={test_data["real_date"]}']
# simple price prediction
if 'mse' in test_items:
pred_seq = model(history, FUTURE_LEN)
mse = torch.sum((pred_seq - future) ** 2) / FUTURE_LEN
test_print.append(f'pred_mse={mse:.4f}')
return_dict['mse'].append(mse)
# cardinality constrained predict-and-optimize
if 'predict-and-opt' in test_items:
_, jpo_weight = model(history, FUTURE_LEN, 'predict-and-opt', RF, K, gumbel_sample_num=1000, gumbel_noise_fact=0.1, return_best_weight=True)
sharpe, risk, return_ = compute_sharpe_ratio(mu, cov, jpo_weight, RF, return_details=True)
test_print.append(f'predict-and-opt: sharpe={sharpe:.4f}, return={return_:.4f}, risk={risk:.4f}')
return_dict['predict-and-opt-return'].append(return_)
return_dict['predict-and-opt-risk'].append(risk)
# find and return the best portfolio in history
if 'history-opt' in test_items:
_, history_weight = model(history, FUTURE_LEN, 'history-opt', RF, K)
sharpe, risk, return_ = compute_sharpe_ratio(mu, cov, history_weight, RF, return_details=True)
test_print.append(f'history-opt: sharpe={sharpe:.4f}, return={return_:.4f}, risk={risk:.4f}')
return_dict['history-opt-return'].append(return_.detach())
return_dict['history-opt-risk'].append(risk.detach())
# cardinality constrained predict-then-optimize
if 'predict-then-opt' in test_items:
_, pto_weight = model(history, FUTURE_LEN, 'predict-then-opt', RF, K)
sharpe, risk, return_ = compute_sharpe_ratio(mu, cov, pto_weight, RF, return_details=True)
test_print.append(f'predict-then-opt: sharpe={sharpe:.4f}, return={return_:.4f}, risk={risk:.4f}')
return_dict['predict-then-opt-return'].append(return_)
return_dict['predict-then-opt-risk'].append(risk)
if verbose:
print(', '.join(test_print))
print('Evaluation complete.')
for k, v in return_dict.items():
print(f'{k}: {sum(v) / len(v):.4f}')
if 'risk' in k: # print the sharpe ratio
r_k = '-'.join(k.split('-')[:-1])
r_v = return_dict[r_k + '-return']
print(r_k)
print(r_v)
print(f'{r_k + "-sharpe"}: {(sum(r_v) / len(r_v) - RF) / (sum(v) / len(v)):.4f}')
return return_dict
if __name__ == '__main__':
import argparse
parser = argparse.ArgumentParser(description='CardNN experiment protocol (predictive portfolio optimization).')
parser.add_argument("--train", help="enable training", action="store_true")
parser.add_argument('--method', help="the name of method "
"(from \'predict-and-opt\', \'predict-then-opt\', \'history-opt\')",
default='predict-and-opt', type=str)
parser.add_argument('--start_epoch', help="the number of starting epoch", default=0, type=int)
args = parser.parse_args()
dataset = PortDataset('snp500', HISTORY_LEN, FUTURE_LEN, train_test_split=0.75)
model = LSTMModel(dataset.num_assets, NUM_FEATURE, 1).to(device=DEVICE)
model.double()
if args.start_epoch > 0:
pretrained_path = f'output/portfolio_lstm_epoch{args.start_epoch}_{args.method}.pt'
print(f'Loading model weights from {pretrained_path}...')
model.load_state_dict(torch.load(pretrained_path))
if args.train:
train_test_portfolio(model, dataset.train_set, mse_weight=0, sharpe_weight=1, opt_method=args.method,
test_set=dataset.test_set, test_items=['mse', args.method])
else:
test_portfolio(model, dataset.test_set, test_items=['mse', args.method], verbose=True)