-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathtrain_rl.py
More file actions
163 lines (137 loc) · 8.76 KB
/
train_rl.py
File metadata and controls
163 lines (137 loc) · 8.76 KB
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
from tsrl.environments import generate_candle_features
from tsrl.experiments.market.experiment import MarketExperiment
from pathlib import Path
from plotly import graph_objects as go
import numpy as np
# 1) Let's load data and generate features
# coin_list = ['EURCHF']
# coin_list = ['AUDCAD', 'EURCHF']
# coin_list = ['AUDCAD', 'AUDCHF', 'AUDJPY', 'AUDNZD', 'AUDUSD', 'CADCHF', 'CADJPY', 'CHFJPY', 'EURAUD', 'EURCAD',
# 'EURCHF', 'EURDKK', 'EURGBP', 'EURJPY', 'EURNOK', 'EURNZD', 'EURSEK', 'EURUSD', 'GBPAUD', 'GBPCAD',
# 'GBPCHF', 'GBPJPY', 'GBPNZD', 'GBPUSD', 'NZDCAD', 'NZDCHF', 'NZDJPY', 'NZDUSD', 'SGDJPY', 'USDCAD',
# 'USDCHF', 'USDDKK', 'USDHKD', 'USDJPY', 'USDNOK', 'USDSEK', 'USDSGD']
# PATH_TO_FEATHER = '../data/Forex/'
for exp in range(1, 11):
# let's use crypto for now
coin_list = ['BTCUSDT','ADAUSDT','ATOMUSDT','BTCBUSD','EOSUSDT','ETCUSDT','ETHUSDT','NEOUSDT','TRXUSDT','VETUSDT','WAVESUSDT','XLMUSDT','XMRUSDT','XRPUSDT' ]#,'ethusdt','adausdt','algousdt','atomusdt','avaxusdt','axsusdt','bchusdt','bnbusdt']
#'BCHUSDT','DOTUSDT','FILUSDT','LTCUSDT','OMGUSDT','UNIUSDT',
# PATH_TO_FEATHER = '../data/Crypto/minute_binance'
#'AXSUSDT','SANDUSDT','MANAUSDT','IOTXUSDT','AVAXUSDT',
PATH_TO_FEATHER = "fin_tsrl_sentiment\\Data\\Crypto - Baseline May\\Crypto - Baseline May\\"
#Path('C:\\Users\\MICHAIL DADOPOULOS\\OneDrive - Αριστοτέλειο Πανεπιστήμιο Θεσσαλονίκης\\Documents\\GitHub\\fin_tsrl_sentiment\\Data\\Feathers\\candles/')
index_col = 'date'
# RESAMPLE = '30T'
resample = '1H'
train_end = "2021-03-15"# "2021" # "2021"
data = generate_candle_features(train_end=train_end,
pairs=coin_list,
feather_folder=PATH_TO_FEATHER,
shift_mins=0,
timescale=resample,
feature_config=(
# basic features
dict(name='int_bar_changes', func_name='inter_bar_changes',
columns=['close', 'high', 'low'],
use_pct=True),
# dict(name='int_bar_changes_10', func_name='inter_bar_changes',
# columns=['close', 'high', 'low'], use_pct=True,
# smoothing_window=10),
# dict(func_name='hl_to_pclose'),
# dict(func_name='next_return'), # this is a cheat feature, to test we can learn
))
print(data)
print(list(data['feature_dict'].values())[0].shape[1])
# 2) Let's set parameters for our environment
env_params = dict(
max_episode_steps=40,
commission_punishment=2e-5,
)
# 3) Let's set parameters for our model (LSTM)
net_size = 32
model_params = dict(
combine_policy_value=False,
nb_actions=3,
lstm_size=net_size,
actor_size=[net_size],
critic_size=[net_size],
dropout=0.2
)
# exp_path=Path(f'~/_experiment').expanduser()
# 4) Now we can initialize our experiment
exp_path = Path(f'~/experiment_contractiveloss_lesspairs_last_grads_autograd/grad_weight_plstmsquare_0_3_random_3Mfreq_lr_5e_5/1511024_exp{exp}').expanduser()
exp = MarketExperiment(exp_path=exp_path, use_sentiment=False)
# 5) It's time to train our agent using specific parameters
exp.train(data= data, model_params=model_params, env_params=env_params, train_range=("2000","2021-03-15"),
ppo_clip=0.2,contractive_weight= 0.3 , n_envs=128, n_reuse_value=1,use_amp=False,rew_limit=6.,truncate_bptt=(5,20),
tau=0.95, env_step_init=1.0,
n_epochs=400, validation_interval=10, show_progress=True, weight_decay=0.,
entropy_weight=0.01, recompute_values=False, batch_size=32,
value_horizon=np.inf,
lr=5e-5, lookahead=False,
#advantage_type='exponential',
advantage_type='direct_reward',
gamma=0.99,
n_reuse_policy=3, n_reuse_aux=0, checkpoint_dir=None)
# 6) Let's check the agent's performance
eval_dict = exp.eval(data)
detailed_pnls = exp.detailed_backtest(data, eval_dict, train_end="2021-03-15")
fig = go.FigureWidget()
for pair, pnls in detailed_pnls.items():
train_pnl = pnls['train_pnl'].cumsum()
test_pnl = train_pnl.values[-1] + pnls['test_pnl'].cumsum()
fig.add_scatter(x=train_pnl.index,y=train_pnl, legendgroup=pair, name=pair)
fig.add_scatter(x=test_pnl.index,y=test_pnl, legendgroup=pair, name=pair)
fig.write_html(f'{exp_path}/figure.html')
# 7)Let's check the agent's performance in thce shifted data
from torch.utils.tensorboard import SummaryWriter
summary_writer = SummaryWriter(str(exp_path), max_queue=500)
shift_periods = [-15,-14,-13,-12,-11,-10,-9,-8,-7,-6,-5,-4,-3,-2,-1,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15]
train_pnls = []
test_pnls = []
train_evals = []
test_evals = []
for shift_period in shift_periods:
data_shifted= generate_candle_features(train_end=train_end,
pairs=coin_list,
feather_folder=PATH_TO_FEATHER,
shift_mins=shift_period,
timescale=resample,
feature_config=(
# basic features
dict(name='int_bar_changes', func_name='inter_bar_changes',
columns=['close', 'high', 'low'],
use_pct=True),
# dict(name='int_bar_changes_10', func_name='inter_bar_changes',
# columns=['close', 'high', 'low'], use_pct=True,
# smoothing_window=10),
# dict(func_name='hl_to_pclose'),
# dict(func_name='next_return'), # this is a cheat feature, to test we can learn
))
eval_dict = exp.eval(data_shifted)
detailed_pnls = exp.detailed_backtest(data_shifted, eval_dict, train_end="2021-03-15")
#save figures for shifted too
# fig_shifted = go.FigureWidget()
# for pair, pnls in detailed_pnls.items():
# train_pnl = pnls['train_pnl'].cumsum()
# test_pnl = train_pnl.values[-1] + pnls['test_pnl'].cumsum()
# fig_shifted.add_scatter(x=train_pnl.index,y=train_pnl, legendgroup=pair, name=pair)
# fig_shifted.add_scatter(x=test_pnl.index,y=test_pnl, legendgroup=pair, name=pair)
# fig_shifted.write_html(f'{exp_path}/figure_shifted{shift_period}.html')
cur_performance_shifted = exp.backtest(data_shifted, eval_dict, train_end="2021-03-15")
cur_performance_pnl_shifted = exp.backtest(data_shifted,eval_dict,train_end="2021-03-15",commission=1e-3)
train_pnls.append(cur_performance_pnl_shifted['train_pnl'])
test_pnls.append(cur_performance_pnl_shifted['test_pnl'])
train_evals.append(cur_performance_shifted['train_pnl'])
test_evals.append(cur_performance_pnl_shifted['test_pnl'])
for k, v in cur_performance_shifted.items():
summary_writer.add_scalar(f'eval_shifted_range/{k}', v,shift_period)
for k, v in cur_performance_pnl_shifted.items():
summary_writer.add_scalar(f'pnl_shifted_range/{k}', v, shift_period)
summary_writer.add_scalar('shifted_mean/mean_train_pnl', np.mean(train_pnls), 1000)
summary_writer.add_scalar('shifted_mean/var_train_pnl', np.var(train_pnls), 1000)
summary_writer.add_scalar('shifted_mean/mean_test_pnl', np.mean(test_pnls), 1000)
summary_writer.add_scalar('shifted_mean/var_test_pnl', np.var(test_pnls), 1000)
summary_writer.add_scalar('shifted_mean/mean_train_eval', np.mean(train_evals), 1000)
summary_writer.add_scalar('shifted_mean/var_train_eval', np.var(train_evals), 1000)
summary_writer.add_scalar('shifted_mean/mean_test_eval', np.mean(test_evals), 1000)
summary_writer.add_scalar('shifted_mean/var_test_eval', np.var(test_evals), 1000)