def test_ngboost(X_train, y_train, X_val, y_val, X_test, y_test, alpha):
# Parameter Grid Search
b5 = DecisionTreeRegressor(criterion='friedman_mse', max_depth=5)
b10 = DecisionTreeRegressor(criterion='friedman_mse', max_depth=10)
b15 = DecisionTreeRegressor(criterion='friedman_mse', max_depth=15)
b20 = DecisionTreeRegressor(criterion='friedman_mse', max_depth=20)
b25 = DecisionTreeRegressor(criterion='friedman_mse', max_depth=25)
b30 = DecisionTreeRegressor(criterion='friedman_mse', max_depth=30)
b35 = DecisionTreeRegressor(criterion='friedman_mse', max_depth=35)
param_grid = {
'minibatch_frac': [1.0, 0.75, 0.5],
'Base': [b5, b10, b15, b20, b25, b30, b35],
'learning_rate': [1e-3, 5e-3, 1e-2, 5e-2, .1]
}
model = ngboost.NGBRegressor(Dist=Normal, Score=LogScore, verbose=False)
grid_search = GridSearchCV(model, param_grid=param_grid, cv=2, n_jobs=-1)
grid_search.fit(X_train, y_train)
# Obtain optimum parameters from grid search and train model with validaton score
best_model = ngboost.NGBRegressor(
Dist=Normal,
verbose=False,
Base=grid_search.best_params_['Base'],
minibatch_frac=grid_search.best_params_['minibatch_frac'],
learning_rate=grid_search.best_params_['learning_rate'],
).fit(X_train, y_train, X_val, y_val, early_stopping_rounds=40)
# Test
y_pred = best_model.pred_dist(X_test,
max_iter=best_model.best_val_loss_itr)
# Model Metrics
rmse = np.mean((y_test - y_pred.params['loc'])**2.)**0.5
test_ll = sum(ll(y_test, y_pred.params['loc'], y_pred.params['scale']))
# Confidence Intervals
on_target = 0
ranges = []
for loc, scale, y in zip(y_pred.params['loc'], y_pred.params['scale'],
y_test):
# Compute Interval
conf_int = stats.norm.interval(1 - alpha, loc=loc, scale=scale)
# Update Metrics
if (y >= conf_int[0] and y <= conf_int[1]): on_target += 1
ranges.append(abs(conf_int[0]) + abs(conf_int[1]))
coverage = on_target / len(ranges)
avg_range = np.mean(ranges)
# Done.
return (coverage, avg_range, test_ll)
def final_model(train_valid_setup):
train = train_valid_setup[0]
test = train_valid_setup[1]
dist = train_valid_setup[2]
eta = train_valid_setup[3]
it = train_valid_setup[4]
if dist == "Laplace":
dist_ = ngboost.distns.Laplace
elif dist == "T":
dist_ = ngboost.distns.TFixedDf
ngb = ngboost.NGBRegressor(Dist=dist_,
Score=ngboost.scores.LogScore,
Base=ExtraTreeRegressor(max_depth=3,
min_samples_split=2),
n_estimators=it,
learning_rate=eta,
minibatch_frac=1.0,
col_sample=1.0,
verbose=False,
verbose_eval=500,
tol=0.0001,
random_state=2021)
ngb.fit(train[features], train["rr"])
Y_dists = ngb.pred_dist(test[features])
return Y_dists.dist.ppf(0.01)[0], Y_dists.dist.ppf(0.025)[0]
def mp_cv(train_valid_setup):
train = train_valid_setup[0]
test = train_valid_setup[1]
if train_valid_setup[2] == 'ET':
model = ExtraTreeRegressor(max_depth=3)
elif train_valid_setup[2] == "DT":
model = DecisionTreeRegressor(max_depth=3)
elif train_valid_setup[2] == "Ridge":
model = Ridge(alpha=0.25)
ngb = ngboost.NGBRegressor(Dist=ngboost.distns.Laplace,
Score=ngboost.scores.LogScore,
Base=model,
n_estimators=500,
learning_rate=0.01,
minibatch_frac=1.0,
col_sample=1.0,
verbose=False,
verbose_eval=500,
tol=0.0001,
random_state=2021)
ngb.fit(train[features], train["rr"])
Y_dists = ngb.pred_dist(test[features])
return [Y_dists.dist.ppf(0.01)[0], Y_dists.dist.ppf(0.025)[0]]
def test_ngboost():
try:
import ngboost
except:
print("Skipping test_ngboost!")
return
X,y = shap.datasets.boston()
model = ngboost.NGBRegressor(n_estimators=20).fit(X, y)
explainer = shap.TreeExplainer(model, model_output=0)
assert np.max(np.abs(explainer.shap_values(X).sum(1) + explainer.expected_value - model.predict(X))) < 1e-5
def ngb_pipeline():
base_model = ngb.NGBRegressor(Dist=Normal,
Score=LogScore,
Base=learner,
n_estimators=500,
learning_rate=0.04,
col_sample=1.0,
minibatch_frac=1.0,
verbose=False,
natural_gradient=True)
return base_model
def final_model_switching(train_valid_setup):
train = train_valid_setup[0]
test = train_valid_setup[1]
dist = train_valid_setup[2]
eta = train_valid_setup[3]
it = train_valid_setup[4]
switch = train_valid_setup[5]
if dist == "Laplace":
dist_ = ngboost.distns.Laplace
elif dist == "T":
dist_ = ngboost.distns.TFixedDf
ngb = ngboost.NGBRegressor(Dist=dist_,
Score=ngboost.scores.LogScore,
Base=ExtraTreeRegressor(max_depth=3,
min_samples_split=2),
n_estimators=it,
learning_rate=eta,
minibatch_frac=1.0,
col_sample=1.0,
verbose=False,
verbose_eval=500,
tol=0.0001,
random_state=2021)
ngb.fit(train[features], train["rr"])
Y_dists = ngb.pred_dist(test[features])
VaR1_NGB, VaR25_NGB = Y_dists.dist.ppf(0.01)[0], Y_dists.dist.ppf(0.025)[0]
am = arch_model(train["rr"],
vol="GARCH",
p=1,
o=0,
q=1,
dist="normal",
mean="AR",
lags=1)
res = am.fit(disp="off")
forecasts = res.forecast(horizon=1)
cond_mean = forecasts.mean.iloc[-1, 0]
cond_var = forecasts.variance.iloc[-1, 0]
stan_resid = res.resid[1:] / res.conditional_volatility[1:]
q = np.percentile(stan_resid, [1, 2.5])
VaR1_GARCH, VaR25_GARCH = cond_mean + np.sqrt(cond_var) * q
if switch == 1:
return VaR1_GARCH, VaR25_GARCH
else:
return VaR1_NGB, VaR25_NGB
def mp_cv_ET_hyp_tuning(train_valid_setup):
train = train_valid_setup[0]
test = train_valid_setup[1]
ngb = ngboost.NGBRegressor(Dist=ngboost.distns.Laplace,
Score=ngboost.scores.LogScore,
Base=ExtraTreeRegressor(
max_depth=train_valid_setup[2],
min_samples_split=train_valid_setup[3]),
n_estimators=500,
learning_rate=0.01,
minibatch_frac=1.0,
col_sample=1.0,
verbose=False,
verbose_eval=500,
tol=0.0001,
random_state=2021)
ngb.fit(train[features], train["rr"])
Y_dists = ngb.pred_dist(test[features])
return [Y_dists.dist.ppf(0.01)[0], Y_dists.dist.ppf(0.025)[0]]
def final_model_ver2(train_valid_setup):
train = train_valid_setup[0]
test = train_valid_setup[1]
dist = train_valid_setup[2]
eta = train_valid_setup[3]
it = train_valid_setup[4]
if dist == "Laplace":
dist_ = ngboost.distns.Laplace
elif dist == "T":
dist_ = ngboost.distns.TFixedDf
ngb = ngboost.NGBRegressor(Dist=dist_,
Score=ngboost.scores.LogScore,
Base=ExtraTreeRegressor(max_depth=3,
min_samples_split=2),
n_estimators=it,
learning_rate=eta,
minibatch_frac=1.0,
col_sample=1.0,
verbose=False,
verbose_eval=500,
tol=0.0001,
random_state=2021)
ngb.fit(train[features], train["rr"])
Y_dists = ngb.pred_dist(test[features])
Y_dists_train = ngb.pred_dist(train[features])
residuals = np.array(train.rr) - ngb.predict(train[features])
standardized_residuals = (residuals) / np.sqrt(
2 * (Y_dists_train.params["scale"])**2)
standardized_residuals_df = pd.DataFrame({"t": standardized_residuals})
VaR1 = Y_dists.params["loc"][0] + np.sqrt(
2 * (Y_dists.params["scale"][0])**
2) * standardized_residuals_df.t.quantile(0.01)
VaR25 = Y_dists.params["loc"][0] + np.sqrt(
2 * (Y_dists.params["scale"][0])**
2) * standardized_residuals_df.t.quantile(0.025)
return VaR1, VaR25