-
Notifications
You must be signed in to change notification settings - Fork 2
/
lstm_bl_rusell.py
213 lines (172 loc) · 7.32 KB
/
lstm_bl_rusell.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
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
import pandas as pd
import yfinance as yf
import config_params
start_date = config_params.start_date
end_date = config_params.end_date
sp500 = yf.download(tickers=["^RUT"], start=start_date, end=end_date, interval="1mo")
bonds = yf.download(tickers=["AGG"], start=start_date, end=end_date, interval="1mo")
bonds.reset_index(inplace=True)
sp500.reset_index(inplace=True)
sp500["ticker"] = "^RUT"
bonds["ticker"] = "AGG"
data = pd.concat([bonds, sp500], axis=0)
data.sort_values(["Date", "ticker"], inplace=True)
test_data = data[data.Date >= "2022"]
train_data = data[data.Date < "2022"]
from ray.air import ScalingConfig
import numpy as np
import pandas as pd
import yfinance as yf
from ray import tune
from ray.tune.schedulers import ASHAScheduler
from ray.tune.search.optuna import OptunaSearch
import config_params
from black_litterman_env import BlackLittermanEnv
from ray.air.integrations.wandb import WandbLoggerCallback
from drllibv2 import DRLlibv2
import config_params
# parser = argparse.ArgumentParser(description="If confidence output")
# parser.add_argument(
# "-if", "--if_confidence", type=bool, help="Whether to output confidence",default=True
# )
# args = parser.parse_args()
# stock_env = BlackLittermanEnv(
# data_df=data,
# )
# print("Reset")
# print(stock_env.reset())
# state,reward,terminal,_,info = stock_env.step([(0.1,0.2,0.05),(0.3,0.4,0.5)])
from ray.tune.registry import register_env
def run_lstm_bl_rusell(if_confidence,test_data=test_data):
if if_confidence=="true":
if_confidence = True
elif if_confidence=="false":
if_confidence=False
env_name = "BlackLitterManEnv-v1"
register_env(
env_name,
lambda config: BlackLittermanEnv(
data_df=train_data, if_confidence=if_confidence
),
)
train_env_instance = BlackLittermanEnv(train_data, if_confidence=if_confidence)
def sample_ppo_params():
return {
"params":{
# "entropy_coeff": tune.choice([1e-5,1e-4]),
# "lr": tune.loguniform(5e-5, 0.0001),
# "sgd_minibatch_size": tune.choice([32, 64, 128, 256]),
# "lambda": tune.choice([0.1, 0.3, 0.5, 0.7, 0.9, 1.0]),
"entropy_coeff": 0.0000001,
"lr": 5e-5,
"sgd_minibatch_size": 64,
"lambda":0.9,
"framework": "torch",
"model": {
"use_lstm": True,
"lstm_cell_size": tune.choice([128, 256, 512])
# 'lstm_cell_size':256
},
"num_envs_per_worker":config_params.num_envs_per_worker
},
"scaling_config": ScalingConfig(
num_workers=config_params.num_workers,
resources_per_worker={"CPU":config_params.worker_cpu,"GPU":config_params.worker_gpu},
use_gpu=True
)
}
model_name = "PPO"
metric = "episode_reward_mean"
mode = "max"
search_alg = OptunaSearch(metric=metric, mode=mode)
scheduler_ = ASHAScheduler(
metric=metric,
mode=mode,
max_t=config_params.training_iterations,
grace_period=config_params.training_iterations//10,
reduction_factor=2,
)
wandb_callback = WandbLoggerCallback(
project=config_params.WANDB_PROJECT,
api_key=config_params.WANDB_API_KEY,
upload_checkpoints=True,
log_config=True,
)
drl_agent = DRLlibv2(
trainable=model_name,
train_env=train_env_instance,
train_env_name="StockTrading_train",
framework="torch",
num_workers=config_params.num_workers,
log_level=config_params.log_level,
run_name="FINRL_TEST_LSTM_RUSELL",
storage_path="FINRL_TEST_LSTM_RUSELL",
params=sample_ppo_params(),
num_samples=config_params.num_samples,
num_gpus=config_params.num_gpus,
num_cpus = config_params.num_cpus,
training_iterations=config_params.training_iterations,
checkpoint_freq=config_params.checkpoint_freq,
# scheduler=scheduler_,
# search_alg=search_alg,
callbacks=[wandb_callback],
)
lstm_res = drl_agent.train_tune_model()
results_df, best_result = drl_agent.infer_results()
# checkpoint = lstm_res.get_best_result().checkpoint
# testing_agent = Algorithm.from_checkpoint(checkpoint)
results_df.to_csv(f"LSTM_{if_confidence}_RUSELL.csv")
ds = []
for i in test_data.groupby("ticker"):
i[1].reset_index(drop=True, inplace=True)
ds.append(i[1])
test_data = pd.concat(ds, axis=0)
test_data.sort_values(["Date"], inplace=True)
test_agent = drl_agent.get_test_agent()
lstm_cell_size = lstm_res.get_best_result().config["model"]["lstm_cell_size"]
init_state = state = [np.zeros([lstm_cell_size], np.float32) for _ in range(2)]
import wandb
wandb.login()
run = wandb.init(project="Test Data")
test_table = wandb.Table(dataframe=test_data)
# Add the table to an Artifact to increase the row
# limit to 200000 and make it easier to reuse
# test_table_artifact = wandb.Artifact("test_data_artifact", type="dataset")
# test_table_artifact.add(test_table, "test_table")
# run.log({f"LSTM_Test_data": test_table})
# run.log_artifact(test_table_artifact)
for runs in range(5):
test_env_instance = BlackLittermanEnv(test_data, if_confidence=if_confidence)
test_env_name = "StockTrading_testenv"
obs = test_env_instance.reset()
done = False
while not done:
action, state, _ = test_agent.compute_single_action(
observation=obs, state=state
)
obs, reward, done, _, _ = test_env_instance.step(action)
log_df = pd.DataFrame(index=test_data.Date.unique()[3:])
if not if_confidence:
log_df["Absolute_ret_1"] = [i[0] for i in test_env_instance.actions]
log_df["Absolute_ret_2"] = [i[1] for i in test_env_instance.actions]
log_df["Relative_ret"] = [i[2] for i in test_env_instance.actions]
else:
log_df["Absolute_ret_1"] = [i[0][0] for i in test_env_instance.actions]
log_df["Confidence_ret_1"] = [i[1][0] for i in test_env_instance.actions]
log_df["Absolute_ret_2"] = [i[0][1] for i in test_env_instance.actions]
log_df["Confidence_ret_2"] = [i[1][1] for i in test_env_instance.actions]
log_df["Relative_ret"] = [i[0][2] for i in test_env_instance.actions]
log_df["Relative_confidence"] = [i[1][2] for i in test_env_instance.actions]
log_df["Weights_1"] = [i[0] for i in test_env_instance.weights_memory]
log_df["Weights_2"] = [i[1] for i in test_env_instance.weights_memory]
log_df["Assets"] = test_env_instance.asset_memory
log_df["Portfolio_return"] = test_env_instance.portfolio_return_memory
log_table = wandb.Table(dataframe=log_df)
# Add the table to an Artifact to increase the row
# limit to 200000 and make it easier to reuse
test_log_artifact = wandb.Artifact("test_log_artifact", type="dataset")
test_log_artifact.add(log_table, "log_table")
# Log the table to visualize with a run...
run.log({f"LSTM_Log_data_{runs}_{if_confidence}_RUSELL": log_table})
# and Log as an Artifact to increase the available row limit!
run.log_artifact(test_log_artifact)