print '\nFeature importance for different lead-lag relations:'
print '----------------------------------------------------\n'
if config.do_model_fit == True:
start_T = time.time() # for time taking
# initialisations for results
feat_imp, feat_imp_sd, target_feat_corr =\
np.zeros((len(config.time_shifts),\
len(config.features))),np.zeros((len(config.time_shifts),len(config.features))),\
np.zeros((len(config.time_shifts),len(config.features)))
# loop over horizon lengths (shift values)
for t, hor in enumerate(config.time_shifts):
df_train = data_func.data_framer(data.raw_data.copy(),config.target,config.features,config.time_var,\
config.start_time,config.end_time,shift=hor,\
trafos=data.trafos,name_trafo=False)
m_test_2 = int(np.round(len(df_train) *
config.test_fraction)) # training set size
# model fit
out_dict = ml_func.ML_train_tester(df_train,config.target,config.features,config.method,\
is_class=config.is_class,m_test=m_test_2,n_boot=config.n_boot,\
to_norm=config.to_norm,counter_fact=config.counter_fact,\
verbose=config.verbose)
# print model specification and progress
if t == 0:
print '\tModel specs:\n\n\t', out_dict['models'][0], '\n'
print '\tHorizon ({0}):'.format(config.unit),
print hor, '..',
i_t = int(np.where(proj_index==end)[0])
i_t_hor = i_t+config.horizon
# TRAINING
# --------
# start time
if config.fixed_start==False and t>0:
i_s += config.time_step_size
start = data.data_shifted.index[i_s]
else:
i_s = 0
start = data.data_shifted.index[i_s]
# training data
df_train = data_func.data_framer(data.data_shifted, config.target, config.features,\
index=config.time_var, start_i=start, end_i=end, name_trafo=False)
m_test_2 = int(np.round(len(df_train)*config.test_fraction)) # test data set fraction of total
# model fit
out_dict = ml_func.ML_train_tester(df_train, config.target, config.features, config.method,\
is_class=config.is_class, m_test=m_test_2, n_boot=config.n_boot,\
to_norm=config.to_norm, counter_fact=config.counter_fact,\
verbose=config.verbose)
# return model specification
if t==0:
print '\n\tModel specs:\n\n\t',out_dict['models'][0],'\n'
# get variable importance
p = ml_func.get_feat_importance(out_dict['feat_weights'])
feat_imp[t,:], feat_imp_sd[t,:] = p[0], p[1]
i_cat_trafo = int(np.where(np.array(config.features) == cat)[0])
cat_trafo = trafos[i_cat_trafo]
del trafos[i_cat_trafo]
# get indicator frames
cat_rawData = pat.dmatrix(cat, raw_data, return_type='dataframe').iloc[:,
1:]
# append columns
for col in cat_rawData.columns:
raw_data[col] = cat_rawData[col]
config.features.append(col)
trafos.append(cat_trafo)
config.features.remove(cat)
#%% get transformed data
data_shifted = data_func.data_framer(data=raw_data.copy(),target=config.target,features=config.features,\
index=config.time_var,start_i=config.start_time,end_i=config.end_time,\
shift=config.horizon,trafos=trafos,name_trafo=False,drop_missing=True)
# number of observations trasining data length
M = len(data_shifted)
if M < 10: # thin dataset warning
print 'Warning: Dataset has less than 20 observations.\n'
if config.init_train_period == 0: # use full data set from the start
config.init_train_period = M
# test data, where features have not been shifted relative to target (needed for future projections).
data_no_shift = data_func.data_framer(data=raw_data.copy(),target=config.target,features=config.features,\
index=config.time_var,start_i=config.start_time,end_i=config.end_time,\
shift=0,trafos=trafos,name_trafo=False,drop_missing=True)
# data & settings
ref_time, cond, models = '2015Q4', True, pro.results_dict['models']
title = '(un)conditioned fan chart for future projection'
fig_name = config.fig_path + 'fan_chart_{0}_ref-{1}-{2}.{3}'.format(
data.ID_short, ref_time, cond, config.fig_format)
# go one horizon length back and into the future
proj_dates = pro.projections.index.values
x_dates, y_dates = proj_dates[-3 * config.horizon -
1:-config.horizon], proj_dates[
-2 * config.horizon - 1:]
df_X = data.data_no_shift[config.features].loc[x_dates]
df_X.index = proj_dates[-2 * config.horizon -
1:] # shift index to prediction period
df_Y = data_func.data_framer(data.raw_data,config.target,config.features,index=config.time_var,\
start_i=config.start_time,end_i='2016Q4',\
shift=0,trafos=data.trafos,name_trafo=False).loc[y_dates]
# plot
ml_plot.cond_fan_chart(df_X,df_Y,models,ref_time,h_ref_line=ref_line,cond=cond,\
x_label=x_label,y_label=y_label,title=title,\
save=config.save_plots,save_name=fig_name)
# feature importance (full sample, fixed frequency, no time series)
# -----------------------------------------------------------------
if (config.counter_fact == True) or (config.method.split('-')[0]
in ['Tree', 'Forest']):
# data & settings
importance, impo_sd = pro.results_dict['feat_imp'], pro.results_dict[
'feat_imp_sd']