def load_train_test(self, indexed_df, enable_load_model, rebuild_model=False, enable_train_if_load_not_suss=True,
enable_train_even_load_succ=False, enable_test=False):
if rebuild_model:
self.get_model(rebuild_model=True)
trade_date = str_2_date(indexed_df.index[-1])
# 加载模型
if enable_load_model:
is_load = self.load_model_if_exist(trade_date)
else:
is_load = False
if enable_train_even_load_succ or (enable_train_if_load_not_suss and not is_load):
factor_df_dic = get_factor(indexed_df, ohlcav_col_name_list=self.ohlcav_col_name_list,
trade_date_series=self.trade_date_series,
delivery_date_series=self.delivery_date_series, do_multiple_factors=True)
factor_df = factor_df_dic[1]
num = 0
while True:
num += 1
if num > 1:
self.get_model(rebuild_model=True)
# 训练模型
train_acc, val_acc = self.train(factor_df_dic, predict_test_random_state=num)
if self.over_fitting_train_acc is not None and train_acc > self.over_fitting_train_acc:
self.logger.warning('第 %d 次训练，训练集精度 train_acc=%.2f%% 过高，可能存在过拟合，重新采样训练',
num, train_acc * 100)
continue
if self.validation_accuracy_base_line is not None:
if val_acc < self.validation_accuracy_base_line:
self.logger.warning('第 %d 次训练，训练结果不及预期，重新采样训练', num)
continue
# elif train_acc - val_acc > 0.15 and val_acc < 0.75:
# self.logger.warning('第 %d 次训练，train_acc=%.2f%%, val_acc=%.2f%% 相差大于 15%% 且验证集正确率小于75%%，重新采样训练',
# num, train_acc * 100, val_acc * 100)
# continue
else:
break
else:
break
self.save_model(trade_date)
self.trade_date_last_train = trade_date
else:
factor_df = get_factor(indexed_df, ohlcav_col_name_list=self.ohlcav_col_name_list,
trade_date_series=self.trade_date_series,
delivery_date_series=self.delivery_date_series)
train_acc, val_acc = self.valid_model_acc(factor_df)
self.trade_date_acc_list[trade_date] = [train_acc, val_acc]
# enable_test 默认为 False
# self.valid_model_acc(factor_df) 以及完全取代 self.predict_test
# self.predict_test 仅用于内部测试使用
if enable_test:
self.predict_test(factor_df)
return factor_df
def _test_account():
# 建立相关数据
n_step = 60
ohlcav_col_name_list = ["open", "high", "low", "close", "amount", "volume"]
from ibats_common.example.data import load_data
md_df = load_data('RB.csv').set_index('trade_date')[ohlcav_col_name_list]
md_df.index = pd.DatetimeIndex(md_df.index)
from ibats_common.backend.factor import get_factor, transfer_2_batch
factors_df = get_factor(md_df, dropna=True)
df_index, df_columns, data_arr_batch = transfer_2_batch(factors_df,
n_step=n_step)
md_df = md_df.loc[df_index, :]
shape = [
data_arr_batch.shape[0], 5,
int(n_step / 5), data_arr_batch.shape[2]
]
data_factors = np.transpose(data_arr_batch.reshape(shape), [0, 2, 3, 1])
print(data_arr_batch.shape, '->', shape, '->', data_factors.shape)
# 建立 Account
env = Account(md_df, data_factors)
next_observation = env.reset()
print('next_observation.shape:', next_observation.shape)
assert next_observation.shape == (1, 12, 78, 5)
next_state, reward, done = env.step(1)
assert next_observation.shape == (1, 12, 78, 5)
assert not done
def on_min1(self, md_df, context):
if self.do_nothing_on_min_bar: # 仅供调试使用
return
# 数据整理
indexed_df = md_df.set_index('trade_date').drop('instrument_type', axis=1)
indexed_df.index = pd.DatetimeIndex(indexed_df.index)
# 获取最新交易日
trade_date = str_2_date(indexed_df.index[-1])
days_after_last_train = (trade_date - self.trade_date_last_train).days
if self.retrain_period is not None and 0 < self.retrain_period < days_after_last_train:
# 重新训练
self.logger.info('当前日期 %s 距离上一次训练 %s 已经过去 %d 天，重新训练',
trade_date, self.trade_date_last_train, days_after_last_train)
factor_df = self.load_train_test(indexed_df, rebuild_model=True,
enable_load_model=self.enable_load_model_if_exist)
else:
factor_df = get_factor(indexed_df, ohlcav_col_name_list=self.ohlcav_col_name_list,
trade_date_series=self.trade_date_series,
delivery_date_series=self.delivery_date_series)
# 预测
pred_mark = self.predict_latest(factor_df)
is_holding, is_buy, is_sell = pred_mark == 0, pred_mark == 1, pred_mark == 2
# self.logger.info('%s is_buy=%s, is_sell=%s', trade_date, str(is_buy), str(is_sell))
close = md_df['close'].iloc[-1]
instrument_id = context[ContextKey.instrument_id_list][0]
if is_buy: # is_buy
position_date_pos_info_dic = self.get_position(instrument_id)
no_target_position = True
if position_date_pos_info_dic is not None:
for position_date, pos_info in position_date_pos_info_dic.items():
direction = pos_info.direction
if direction == Direction.Short:
self.close_short(instrument_id, close, pos_info.position)
elif direction == Direction.Long:
no_target_position = False
if no_target_position:
self.open_long(instrument_id, close, self.unit)
else:
self.logger.debug("%s %s %.2f holding", self.trade_agent.curr_timestamp, instrument_id, close)
if is_sell: # is_sell
position_date_pos_info_dic = self.get_position(instrument_id)
no_holding_target_position = True
if position_date_pos_info_dic is not None:
for position_date, pos_info in position_date_pos_info_dic.items():
direction = pos_info.direction
if direction == Direction.Long:
self.close_long(instrument_id, close, pos_info.position)
elif direction == Direction.Short:
no_holding_target_position = False
if no_holding_target_position:
self.open_short(instrument_id, close, self.unit)
else:
self.logger.debug("%s %s %.2f holding", self.trade_agent.curr_timestamp, instrument_id, close)
if is_holding:
self.logger.debug("%s %s * * %.2f holding", self.trade_agent.curr_timestamp, instrument_id, close)
def _test_factor_analysis():
from ibats_common.example.data import load_data
from ibats_common.backend.factor import get_factor
df = load_data("RB.csv",
index_col='trade_date',
parse_index_to_datetime=True).drop(['instrument_type'],
axis=1)
factor_df = get_factor(df, price_factor_kwargs={
'with_diff_n': False
}).dropna()
ana_dic = factor_analysis(factor_df)
def _test_summary_md_2_docx(auto_open_file=True):
from ibats_common.example.data import load_data
instrument_type = 'RU' # 'RB' 'RU'
file_name = f"{instrument_type}.csv"
factor_df = load_data(file_name).set_index('trade_date').drop(
'instrument_type', axis=1)
factor_df.index = pd.DatetimeIndex(factor_df.index)
column_list_oraginal = list(factor_df.columns)
ohlcav_col_name_list = ["open", "high", "low", "close", "amount", "volume"]
from ibats_common.backend.factor import get_factor
from ibats_common.example.data import get_trade_date_series
from ibats_common.example.data import get_delivery_date_series
factor_df = get_factor(
factor_df,
ohlcav_col_name_list=ohlcav_col_name_list,
trade_date_series=get_trade_date_series(),
delivery_date_series=get_delivery_date_series(instrument_type))
col_transfer_dic = {'return': ['open', 'high', 'low', 'close', 'volume']}
file_path = summary_md_2_docx(
factor_df,
enable_show_plot=False,
enable_save_plot=True,
close_key='close',
name=instrument_type,
func_kwargs_dic={
"hist": {
"figure_4_each_col": False,
"columns": column_list_oraginal,
"col_transfer_dic": col_transfer_dic,
},
"drawdown": {
"col_name_list": ['close'],
},
"rr": {
"col_name_list": ['close'],
},
"hist_future_n_rr": {
'n_days': [3, 5],
"columns": ['close'],
},
"rr_quantile": {
'columns': ['close']
},
"validation": {
'trade_date_max_gap': 10
},
# "": {},
})
if auto_open_file:
open_file_with_system_app(file_path)
def _test_quote_market():
n_step = 60
ohlcav_col_name_list = ["open", "high", "low", "close", "amount", "volume"]
from ibats_common.example.data import load_data
md_df = load_data('RB.csv').set_index('trade_date')[ohlcav_col_name_list]
md_df.index = pd.DatetimeIndex(md_df.index)
from ibats_common.backend.factor import get_factor, transfer_2_batch
factors_df = get_factor(md_df, dropna=True)
df_index, df_columns, data_arr_batch = transfer_2_batch(factors_df,
n_step=n_step)
md_df = md_df.loc[df_index, :]
# 建立 QuotesMarket
qm = QuotesMarket(md_df=md_df[['close', 'open']],
data_factors=data_arr_batch,
state_with_flag=True)
next_observation = qm.reset()
assert len(next_observation) == 2
assert next_observation[0].shape[0] == n_step
assert next_observation[1] == 0
next_observation, reward, done = qm.step(1)
assert len(next_observation) == 2
assert next_observation[1] == 1
assert not done
next_observation, reward, done = qm.step(0)
assert next_observation[1] == 0
assert reward != 0
next_observation, reward, done = qm.step(0)
assert next_observation[1] == 0
assert reward == 0
next_observation, reward, done = qm.step(3)
assert next_observation[1] == 0
assert reward == 0
next_observation, reward, done = qm.step(2)
assert next_observation[1] == -1
assert not done
next_observation, reward, done = qm.step(3)
assert next_observation[1] == -1
assert reward != 0
try:
qm.step(4)
except ValueError:
print('is ok for not supporting action>3')
def _test_quote_market():
import os
n_step = 60
from ibats_common.example.data import load_data
md_df = load_data(
'RB.csv',
folder_path=os.path.join(os.pardir, os.pardir, os.pardir, 'example', 'data') # r'..\..\..\example\data'
).set_index('trade_date')[DEFAULT_MD_OHLCVA_LABELS]
md_df.index = pd.DatetimeIndex(md_df.index)
from ibats_common.backend.factor import get_factor, transfer_2_batch
factors_df = get_factor(md_df, dropna=True)
df_index, df_columns, data_arr_batch = transfer_2_batch(factors_df, n_step=n_step)
md_df = md_df.loc[df_index, :]
# 建立 QuotesMarket
qm = QuotesMarket(md_df=md_df[['close', 'open']], data_factors=data_arr_batch, state_with_flag=True)
next_observation = qm.reset()
assert len(next_observation) == 3
assert next_observation[0].shape[0] == n_step
assert next_observation[1] == FLAG_EMPTY
next_observation, reward, done = qm.step(ACTION_LONG)
assert len(next_observation) == 3
assert next_observation[1] == FLAG_LONG
assert not done
next_observation, reward, done = qm.step(ACTION_CLOSE)
assert next_observation[1] == FLAG_EMPTY
assert reward != 0
next_observation, reward, done = qm.step(ACTION_CLOSE)
assert next_observation[1] == FLAG_EMPTY
assert reward == 0
next_observation, reward, done = qm.step(ACTION_KEEP)
assert next_observation[1] == FLAG_EMPTY
assert reward == 0
next_observation, reward, done = qm.step(ACTION_SHORT)
assert next_observation[1] == FLAG_SHORT
assert not done
next_observation, reward, done = qm.step(ACTION_KEEP)
assert next_observation[1] == FLAG_SHORT
assert reward != 0
try:
qm.step(4)
except ValueError:
print('is ok for not supporting action>3')
def get_xy(df: pd.DataFrame,
target_n_bars=5,
get_factor_kwargs=None,
get_y_kwargs=None):
"""
生成 X Y 数据并进行筛选，对齐等
:param df:
:param target_n_bars:
:param get_factor_kwargs:
:param get_y_kwargs:
:return:
"""
get_factor_kwargs = {} if get_factor_kwargs is None else get_factor_kwargs
factor_df = get_factor(df, **get_factor_kwargs)
get_y_kwargs = {} if get_y_kwargs is None else get_y_kwargs
get_y_kwargs.update({"target_n_bars": target_n_bars})
y_s = get_y(df, **get_factor_kwargs)
# 数据切片
hist_bar_df, factor_df = df.iloc[:
-target_n_bars], factor_df.iloc[:
-target_n_bars]
logger.info("hist_bar_df.shape=%s, factor_df.shape=%s", hist_bar_df.shape,
factor_df.shape)
logger.info("y_s.shape=%s", y_s.shape)
assert factor_df.shape[0] == y_s.shape[0], \
f"因子数据 x 长度 {factor_df.shape[0]} 要与训练目标 y 数据长度 {y_s.shape[0]} 一致"
# 剔除无效数据，并根据 target_n_bars 进行数据切片
is_available = ~(np.isinf(y_s.to_numpy())
| np.isnan(y_s.to_numpy())
| np.any(np.isnan(factor_df.to_numpy()), axis=1)
| np.any(np.isinf(factor_df.to_numpy()), axis=1))
available_df = hist_bar_df[is_available]
available_factor_df = factor_df[is_available]
x_arr = available_factor_df.to_numpy()
y_arr = y_s[is_available]
assert x_arr.shape[0] == y_arr.shape[0], \
f"因子数据 x 长度 {x_arr.shape[0]} 要与训练目标 y 数据长度 {y_arr.shape[0]} 一致"
logger.info("x_arr.shape=%s, y_arr.shape=%s", x_arr.shape, y_arr.shape)
return available_df, available_factor_df, x_arr, y_arr
def generate_factors(self):
"""整理缓存数据，生成相应的因子"""
df = pd.DataFrame([{key: getattr(_, key)
for key in BAR_ATTRIBUTES}
for _ in self._hist_bar_list]).set_index('datetime')
df.index = pd.to_datetime(df.index)
# 重置缓冲区状态
self._hist_bar_list = []
self._hist_bar_days = 0
# 扩展 hist_bar_df
if self.hist_bar_df is None:
self.hist_bar_df = df
else:
self.hist_bar_df = self.hist_bar_df.append(df).sort_index()
# 生成因子
self._factor_df = get_factor(self.hist_bar_df,
ohlcav_col_name_list=[
'open_price', 'high_price',
'low_price', 'close_price', None,
'volume'
],
dropna=False)
def _test_account2():
"""测试 plot_data 返回数据是否符合预期"""
n_step = 60
ohlcav_col_name_list = ["open", "high", "low", "close", "amount", "volume"]
from ibats_common.example.data import load_data
md_df = load_data('RB.csv').set_index('trade_date')[ohlcav_col_name_list]
md_df.index = pd.DatetimeIndex(md_df.index)
from ibats_common.backend.factor import get_factor, transfer_2_batch
factors_df = get_factor(md_df, dropna=True)
df_index, df_columns, data_arr_batch = transfer_2_batch(factors_df,
n_step=n_step)
md_df = md_df.loc[df_index, :]
shape = [
data_arr_batch.shape[0], 5,
int(n_step / 5), data_arr_batch.shape[2]
]
data_factors = np.transpose(data_arr_batch.reshape(shape), [0, 2, 3, 1])
print(data_arr_batch.shape, '->', shape, '->', data_factors.shape)
# 建立 Account
env = Account(md_df, data_factors)
next_observation = env.reset()
# 做空
env.step(2)
for n in range(int(md_df.shape[0] / 2)):
env.step(3)
# 做多
next_observation, reward, done = env.step(1)
while not done:
next_observation, reward, done = env.step(3)
# 展示结果
reward_df = env.plot_data()
value_s = reward_df.iloc[:, 0]
from ibats_utils.mess import datetime_2_str
from datetime import datetime
dt_str = datetime_2_str(datetime.now(), '%Y-%m-%d %H_%M_%S')
title = f'test_account_{dt_str}'
from ibats_common.analysis.plot import plot_twin
plot_twin(value_s, md_df["close"], name=title)
def on_min1_release(self, md_df):
"""
增加模型对未来数据预测成功率走势图展示
:param md_df:
:return:
"""
if md_df is None or md_df.shape[0] == 0:
self.logger.warning('md_df is None or shape[0] == 0')
return
else:
self.logger.debug('md_df.shape= %s', md_df.shape)
# 获取各个模型训练时间点及路径
date_file_path_pair_list = self.get_date_file_path_pair_list()
if len(date_file_path_pair_list) > 0:
# 按日期排序
date_file_path_pair_list.sort(key=lambda x: x[0])
# 建立数据集
indexed_df = md_df.set_index('trade_date').drop('instrument_type', axis=1)
trade_date_end = indexed_df.index[-1]
factor_df = get_factor(indexed_df, ohlcav_col_name_list=self.ohlcav_col_name_list,
trade_date_series=self.trade_date_series,
delivery_date_series=self.delivery_date_series)
xs, ys_onehot, trade_date_index = self.get_x_y(factor_df)
ys = np.argmax(ys_onehot, axis=1)
data_len = len(trade_date_index)
if data_len == 0:
self.logger.warning('ys 长度为0，请检查是否存在数据错误')
return
trade_date2_list = [_[0] for _ in date_file_path_pair_list][1:]
trade_date2_list.append(None)
# 预测结果
self.logger.info("按日期分段验证检验预测结果")
pred_ys_tot, real_ys_tot, img_meta_dic_list = [], [], []
# 根据模型 trade_date_last_train 进行分段预测，并将结果记录到 pred_ys
for num, ((trade_date_last_train, file_path, predict_test_random_state), trade_date_next) in enumerate(zip(
date_file_path_pair_list, trade_date2_list)):
# 以模型训练日期为基准，后面的数据作为验证集数据（样本外数据）
# 获取有效的日期范围 from - to
range_from_arr = trade_date_index >= pd.to_datetime(trade_date_last_train)
range_from_len = len(range_from_arr)
if range_from_len == 0: # range_from_len 应该与 trade_date_list_count 等长度，所以这个条件应该永远不会满足
self.logger.error('总共%d条数据，%s 开始后面没有可验证数据', data_len, trade_date_last_train)
continue
true_count = sum(range_from_arr)
self.logger.debug("len(range_from)=%d, True Count=%d", len(range_from_arr), true_count)
if true_count == 0:
self.logger.warning('总共%d条数据，%s 开始后面没有可验证数据', data_len, trade_date_last_train)
continue
# 自 trade_date_last_train 起的所有有效日期
trade_date_list_sub = trade_date_index[range_from_arr]
# 获取 in_range，作为 range_from, range_to 的交集
if trade_date_next is None:
in_range_arr = None
in_range_count = true_count
else:
in_range_arr = trade_date_list_sub < pd.to_datetime(trade_date_next)
in_range_count = sum(in_range_arr)
if in_range_count == 0:
self.logger.warning('总共%d条数据，[%s - %s) 之间没有可用数据',
data_len, trade_date_last_train, trade_date_next)
continue
else:
self.logger.debug('总共%d条数据，[%s - %s) 之间有 %d 条数据将被验证 model path:%s',
data_len, trade_date_last_train, trade_date_next, in_range_count, file_path)
# 获取当前时段对应的 xs
# 进行验证时，对 range_from 开始的全部数据进行预测，按照 range_to 为分界线分区着色显示
xs_sub, real_ys = xs[range_from_arr, :, :], ys[range_from_arr]
close_df = indexed_df.loc[trade_date_list_sub, 'close']
# 加载模型
is_load = self.load_model_if_exist(trade_date_last_train, enable_load_model_if_exist=True)
if not is_load:
self.logger.error('%s 模型加载失败：%s', trade_date_last_train, file_path)
continue
# 预测
pred_ys_one_hot = self.model.predict(xs_sub)
pred_ys = np.argmax(pred_ys_one_hot, axis=1)
if in_range_arr is not None and in_range_count > 0:
pred_ys_tot.extend(pred_ys[in_range_arr])
else:
pred_ys_tot.extend(pred_ys)
# 为每一个时段单独验证成功率，以当前模型为基准，验证后面全部历史数据成功率走势
if trade_date_next is None:
split_point_list = None
else:
split_point_list = [close_df.index[0], trade_date_next, close_df.index[-1]]
base_line_list = self.trade_date_acc_list[trade_date_last_train]
img_file_path = show_dl_accuracy(real_ys, pred_ys, close_df, split_point_list,
base_line_list=base_line_list)
img_meta_dic_list.append({
'img_file_path': img_file_path,
'trade_date_last_train': trade_date_last_train,
'module_file_path': file_path,
'predict_test_random_state': predict_test_random_state,
'split_point_list': split_point_list,
'in_range_count': in_range_count,
'trade_date_end': trade_date_end,
})
pred_ys_tot = np.array(pred_ys_tot)
trade_date_last_train_first = pd.to_datetime(date_file_path_pair_list[0][0])
split_point_list = [_[0] for _ in date_file_path_pair_list]
split_point_list.append(trade_date_index[-1])
# 获取 real_ys
real_ys = ys[trade_date_index >= trade_date_last_train_first]
close_df = indexed_df.loc[trade_date_index[trade_date_index >= trade_date_last_train_first], 'close']
img_file_path = show_dl_accuracy(real_ys, pred_ys_tot, close_df, split_point_list)
img_meta_dic_list.append({
'img_file_path': img_file_path,
'trade_date_last_train': trade_date_last_train_first,
'module_file_path': date_file_path_pair_list[0][1],
'predict_test_random_state': date_file_path_pair_list[0][2],
'split_point_list': split_point_list,
'in_range_count': close_df.shape[0],
'trade_date_end': trade_date_end,
})
is_output_docx = True
if is_output_docx:
title = f"[{self.stg_run_id}] predict accuracy trend report"
file_path = summary_release_2_docx(title, img_meta_dic_list)
copy_file_to(file_path, self.base_folder_path)
