def cv_param_search(estimator, valid_portion=0.2, n_cv_folds=10, n_jobs=1):
df_train, df_valid = train_valid_split(valid_portion)
X_train, Y_train = target_feature_split(df_train,
'log_ret_1',
filter_nan=True)
selected_params = estimator.fit_by_cv(X_train,
Y_train,
n_folds=n_cv_folds,
n_jobs=n_jobs)
return selected_params
def main():
# load data
df = make_overall_eurostoxx_df()
X, Y = target_feature_split(df, 'log_ret_1', filter_nan=True)
X, Y = np.array(X), np.array(Y)
if PCA_FEATURES:
X = pca_comp(X, n_components=4)
ndim_x, ndim_y = X.shape[1], 1
mdn = MixtureDensityNetwork('mdn_empirical_no_pca',
ndim_x,
ndim_y,
n_centers=20,
n_training_epochs=500,
random_seed=22,
x_noise_std=0.2,
y_noise_std=0.1)
mdn.fit(X, Y)
mdn.plot2d(x_cond=[np.mean(X, axis=0)], ylim=(-0.02, 0.02))
def main():
# load data
df = make_overall_eurostoxx_df()
X, Y, features = target_feature_split(df,
'log_ret_1',
filter_nan=True,
return_features=True)
X, Y = np.array(X), np.array(Y)
ndim_x, ndim_y = X.shape[1], 1
mdn = MixtureDensityNetwork('mdn_empirical_no_pca',
ndim_x,
ndim_y,
n_centers=20,
n_training_epochs=10,
random_seed=28,
x_noise_std=0.2,
y_noise_std=0.1)
mdn.fit(X, Y)
X_mean = np.mean(X, axis=0)
X_std = np.mean(X, axis=0)
# individual plots
for i, feature in enumerate(features):
factor = np.zeros(X_std.shape)
factor[i] = X_std[i]
x_cond = [
X_mean - 2 * factor, X_mean - 1 * factor, X_mean,
X_mean + 1 * factor, X_mean + 2 * factor
]
mdn.plot2d(x_cond=x_cond, ylim=(-0.04, 0.04), show=False)
plt.legend(
['mean-2*std', 'mean-1*std', 'mean', 'mean+1*std', 'mean+2*std'])
plt.title(feature)
fig_path = os.path.join(DATA_DIR,
'plots/feature_selection/' + feature + '.png')
plt.savefig(fig_path)
# one large plot
resolution = 100
ncols = 3
fig, axes = plt.subplots(nrows=5, ncols=ncols, figsize=(12, 16))
y = np.linspace(-0.04, 0.04, resolution)
n = 0
for i, feature in enumerate(features):
if n == 2:
n += 1
factor = np.zeros(X_std.shape)
factor[i] = X_std[i]
x_cond = [
X_mean + 2 * factor, X_mean + 1 * factor, X_mean,
X_mean - 1 * factor, X_mean - 2 * factor
]
for j in range(len(x_cond)):
x = np.array([x_cond[j] for _ in range(resolution)])
z = mdn.pdf(x, y)
axes[n // ncols][n % ncols].plot(y, z)
axes[n // ncols][n % ncols].set_title(feature)
n += 1
axes[0][0].set_xlabel('log return')
axes[0][0].set_ylabel('probability density log-returns')
# make top right plot disappear
axes[0, 2].tick_params(colors='white')
for spine in axes[0, 2].spines.values():
spine.set_color('white')
fig.legend(
['mean+2*std', 'mean+1*std', 'mean', 'mean-1*std', 'mean-2*std'],
loc=(.77, 0.88))
fig.tight_layout()
fig_path = os.path.join(
DATA_DIR,
'plots/feature_selection/feature_selection_all_variables.png')
fig.savefig(fig_path)
fig_path = os.path.join(
DATA_DIR,
'plots/feature_selection/feature_selection_all_variables.pdf')
fig.savefig(fig_path)
def empirical_evaluation(estimator,
valid_portion=0.2,
moment_r2=True,
eval_by_fc=False,
fit_by_cv=False):
"""
Fits the estimator and, based on a left out validation splot, computes the
Root Mean Squared Error (RMSE) between realized and estimated mean and std
Args:
estimator: estimator object
valid_portion: portion of dataset to be separated as validation set
moment_r2: (bool) whether to compute the rmse of mean and variance
Returns:
(likelihood, mu_rmse, std_rmse)
"""
# get data and split into train and valid set
df_train, df_valid = train_valid_split(valid_portion)
X_train, Y_train = target_feature_split(df_train,
'log_ret_1',
filter_nan=True)
X_valid, Y_valid = target_feature_split(df_valid,
'log_ret_1',
filter_nan=True)
# realized moments
mu_realized = df_valid['log_ret_last_period'][1:]
std_realized_intraday = np.sqrt(df_valid['RealizedVariation'][1:])
# fit density model
if eval_by_fc and not fit_by_cv:
raise NotImplementedError
elif not eval_by_fc and fit_by_cv:
estimator.fit_by_cv(X_train, Y_train, n_folds=5)
else:
estimator.fit(X_train, Y_train)
# compute avg. log likelihood
mean_logli = np.mean(estimator.log_pdf(X_valid, Y_valid))
if moment_r2:
# predict mean and std
mu_predicted, std_predicted = estimator.mean_std(X_valid,
n_samples=N_SAMPLES)
mu_predicted = mu_predicted.flatten()[:-1]
std_predicted = std_predicted.flatten()[:-1]
assert mu_realized.shape == mu_predicted.shape
assert std_realized_intraday.shape == std_realized_intraday.shape
# compute realized std
std_realized = np.abs(mu_predicted - mu_realized)
# compute RMSE
mu_rmse = np.sqrt(np.mean((mu_realized - mu_predicted)**2))
std_rmse = np.sqrt(np.mean((std_realized - std_predicted)**2))
std_intraday_rmse = np.sqrt(
np.mean((std_realized_intraday - std_predicted)**2))
else:
mu_rmse, std_rmse, std_intraday_rmse = None, None, None
return mean_logli, mu_rmse, std_rmse, std_intraday_rmse
def main():
if COMPUTE_MOMENTS:
# 1) load data
df = make_overall_eurostoxx_df()
X, Y, features = target_feature_split(df,
'log_ret_1',
filter_nan=True,
return_features=True)
X, Y = np.array(X), np.array(Y)
ndim_x, ndim_y = X.shape[1], 1
# 2) Fite density model
mdn = MixtureDensityNetwork('mdn_empirical_no_pca',
ndim_x,
ndim_y,
n_centers=20,
n_training_epochs=10,
random_seed=22,
x_noise_std=0.2,
y_noise_std=0.1)
mdn.fit(X, Y)
# 3) estimate moments
n_samples = 10**7
print('compute mean')
mean = np.squeeze(mdn.mean_(x_cond=X, n_samples=n_samples))
print('compute cov')
cov = np.squeeze(mdn.covariance(x_cond=X, n_samples=n_samples))
print('compute skewness')
skew = mdn._skewness_mc(x_cond=X, n_samples=n_samples)
print('compute kurtosis')
kurt = mdn._kurtosis_mc(x_cond=X, n_samples=n_samples)
# 4) save data
data = np.stack([mean, cov, skew, kurt], axis=-1)
moments_df = pd.DataFrame(
data=data,
index=df.dropna().index,
columns=['mean', 'variance', 'skewness', 'kurtosis'])
print(moments_df)
# dump csv
if not os.path.exists(dump_dir):
os.makedirs(dump_dir)
moments_df.to_csv(dump_file_path)
else:
moments_df = pd.read_csv(dump_file_path, index_col=0)
#5) plot moment timeseries
fig, axes = plt.subplots(nrows=4, ncols=1, figsize=(15, 20))
x = moments_df.index
for i in range(4):
label = moments_df.columns[i]
y = moments_df.ix[:, i]
axes[i].plot(x, y)
axes[i].set_title(label)
plt.savefig(os.path.join(dump_dir, 'moments_time_series.png'))
print("Saved figure")