def mySettings():
settings = get_settings()
# Set futures to trade
futures_list = get_futures_list(filter_insignificant_lag=2)
settings["markets"] = ["CASH", *futures_list]
# Load LSTM models
settings["lstm_models"] = {
ticker: load_lstm_model(ticker)
for ticker in futures_list
}
# Load XGBoost models
settings["xgb_models"] = {
ticker: load_xgb_model(ticker)
for ticker in futures_list
}
# Load ARIMA parameters
settings["arima_params"] = {
ticker: load_arima_parameters(ticker)
for ticker in futures_list
}
return settings
def mySettings():
settings = get_settings()
futures_list = get_futures_list(filter_insignificant_lag=2)
settings["markets"] = ["CASH", *futures_list]
settings["models"] = {
ticker: load_lstm_model(ticker)
for ticker in futures_list
}
return settings
def mySettings():
settings = get_settings()
futures_list = get_futures_list(filter_insignificant_lag=2)
settings["markets"] = ["CASH", *futures_list]
# Set parameters for GA
settings['population_size'] = 100
settings['tournament_size'] = 5
settings['crossover_rate'] = 0.3
settings['mutation_rate'] = 0.02
settings['n_iter'] = 100
return settings
def build_arima():
futures_list = get_futures_list(filter_insignificant_lag_1_acf=True)
for ticker in futures_list:
print(f"{futures_list.index(ticker) + 1}/{len(futures_list)}: {ticker}")
data = load_processed_data(ticker)
data = data.loc[:"2020-12-31"]
data = data.iloc[:-1]
data = data[['CLOSE']]
arima_model = pmdarima.auto_arima(data)
arima_model = arima_model.fit(data)
p, d, q = arima_model.order
arima_residuals = arima_model.arima_res_.resid
params = {"p": p, "q": q, "d":d, "residuals": list(arima_residuals)}
# Save model
dire = f"./models/arima/param2/"
os.makedirs(os.path.dirname(dire), exist_ok=True)
with open(f'{dire}/{ticker}_params.txt', 'w') as f:
json.dump(params, f, ensure_ascii=False)
print(f"Saved parameters for {ticker}")
from models.xgboost.training_util import train_model
from systems.systems_util import get_futures_list
from utils.data_loader import load_processed_data
# Cross validation
ts_crossval = TimeSeriesSplit(n_splits=5)
# Define search space for bayesian optimisation
XGB_param_hyperopt = {
"booster": hp.choice("booster", ["gblinear"]),
"max_depth": scope.int(hp.quniform("max_depth", 1, 5, 1)),
"subsample": hp.uniform("subsample", 0.4, 0.6),
"colsample_bytree": hp.uniform("colsample_bytree", 0.5, 0.6),
"colsample_bynode": hp.uniform("colsample_bynode", 0.4, 0.6),
"colsample_bylevel": hp.uniform("colsample_bylevel", 0.4, 0.6),
"gamma": hp.uniform("gamma", 0, 10),
"min_child_weight": hp.uniform("min_child_weight", 1.5, 2.3),
"n_estimators": 100,
"reg_lambda": hp.uniform("reg_lambda", 1, 8),
"reg_alpha": hp.uniform("reg_alpha", 0, 0.02),
}
futures = get_futures_list(filter_insignificant_lag=1)
# Pre-train models and save the weights
for ticker in futures:
data = load_processed_data(ticker)
data = data.loc[:"2020-12-31"]
data = data.iloc[:-1]
train_model(data, XGB_param_hyperopt, ts_crossval, ticker)
def execute():
for ticker in get_futures_list(filter_insignificant_lag=1):
if not is_model_trained_and_saved(ticker=ticker):
train_and_save_model(ticker=ticker)
else:
print(f"Model for {ticker} is already trained and saved!")
def mySettings():
settings = get_settings()
settings["markets"] = [
"CASH", *get_futures_list(filter_insignificant_lag_1_acf=True)
]
return settings
from models.lgbm.training_util import split, train_model
from systems.systems_util import get_futures_list
from utils.data_loader import load_processed_data
FUTURES_LIST = get_futures_list(filter_insignificant_lag=2)
params = {
'num_threads': 4,
'num_class': 3,
'objective': 'multiclassova',
'seed': 123,
'tree_learner': 'feature',
'is_unbalance': True
}
# tune these
params['learning_rate'] = 0.03
params['boosting_type'] = 'gbdt'
params['metric'] = 'multi_logloss'
params['num_leaves'] = 15
params['extra_trees'] = True
params['sub_feature'] = 0.7
params['min_data'] = 20
params['max_depth'] = 20
params['num_iterations'] = 500
params['max_bin'] = 500
params['num_boost_round'] = 175
params['early_stopping_rounds'] = 30
def mySettings():
settings = get_settings()
futures_list = get_futures_list(filter_insignificant_lag=2)
settings["markets"] = ["CASH", *futures_list]
return settings
from utils.data_loader import load_processed_data
from systems.systems_util import get_futures_list
from models.lgbm.training_util import split, train_model
FUTURES_LIST = get_futures_list(filter_insignificant_lag_1_acf=True)
params = {'num_threads': 4,
'num_class': 3,
'objective': 'multiclassova',
'seed': 123,
'tree_learner': 'feature',
'is_unbalance': True,
}
# tune these
params['learning_rate'] = 0.03
params['boosting_type'] = 'gbdt'
params['metric'] = 'multi_logloss'
params['num_leaves'] = 15
params['extra_trees'] = True
params['sub_feature'] = 0.7
params['min_data'] = 20
params['max_depth'] = 20
params['num_iterations'] = 500
params['max_bin'] = 500
params['num_boost_round'] = 175
params['early_stopping_rounds'] = 30
def train(ticker):
from models.xgboost.training_util import train_model
from systems.systems_util import get_futures_list
from utils.data_loader import load_processed_data
# Cross validation
ts_crossval = TimeSeriesSplit(n_splits=5)
# Define search space for bayesian optimisation
XGB_param_hyperopt = {
"booster": hp.choice("booster", ["gblinear"]),
"max_depth": scope.int(hp.quniform("max_depth", 1, 5, 1)),
"subsample": hp.uniform("subsample", 0.4, 0.6),
"colsample_bytree": hp.uniform("colsample_bytree", 0.5, 0.6),
"colsample_bynode": hp.uniform("colsample_bynode", 0.4, 0.6),
"colsample_bylevel": hp.uniform("colsample_bylevel", 0.4, 0.6),
"gamma": hp.uniform("gamma", 0, 10),
"min_child_weight": hp.uniform("min_child_weight", 1.5, 2.3),
"n_estimators": 100,
"reg_lambda": hp.uniform("reg_lambda", 1, 8),
"reg_alpha": hp.uniform("reg_alpha", 0, 0.02),
}
futures = get_futures_list(filter_insignificant_lag_1_acf=True)
# Pre-train models and save the weights
for ticker in futures:
data = load_processed_data(ticker)
data = data.loc[:"2020-12-31"]
data = data.iloc[:-1]
train_model(data, XGB_param_hyperopt, ts_crossval, ticker)