def main():
model_specs = [['GBC', gbm_formula1, kw1]]
train1 = []
train2 = []
test1 = []
test2 = []
global df
for train, test, trial_i in utility.train_test_splitter(
df, 0.1, 'ORIG_DTE'):
train_pos = train[train['dflt_pct'] > 0]
# stage 1
train['PD_pred'], test['PD_pred'] = fit_stage1(model_specs[0], train,
test, True)
print('\nRunning bootstraps...')
btstrp_stage1 = bootstrap.get_btstrp(fit_stage1, model_specs[0], train,
test, 1)
vin_id, x, lo_name, hi_name = 'ORIG_DTE', 'AGE', '2.5%', '97.5%'
df_stage1 = pd.concat(
[test[[vin_id, x, 'did_dflt', 'PD_pred']], btstrp_stage1], axis=1)
for date, dall in df_stage1.groupby(vin_id):
print('\nGenerating plots for {0}...'.format(date))
ds = []
df_ = df_stage1
to_append = df_[df_[vin_id] == date]
del to_append[vin_id]
ds.append(to_append)
del dall[vin_id]
# fig, axes = plt.subplots(2, 2, figsize=(10, 10))
# plt.subplot(121)
d1 = ds[0]
fpr, tpr = utility.get_roc_curve(train['did_dflt'],
train['PD_pred'])
train1.append(fpr)
train2.append(tpr)
fpr_average1 = np.mean(train1)
tpr_average1 = np.mean(train2)
# roc_auc = utility.get_auc(fpr, tpr)
# plt.plot(fpr, tpr, label='Train AUC = {0:.2f}'.format(roc_auc))
# plt.subplot(122)
fpr, tpr = utility.get_roc_curve(d1['did_dflt'], d1['PD_pred'])
test1.append(fpr)
test2.append(tpr)
fpr_average2 = np.mean(test1)
tpr_average2 = np.mean(test2)
# roc_auc = utility.get_auc(fpr, tpr)
# plt.plot(fpr, tpr, label='Test AUC = {0:.2f}'.format(roc_auc))
plt.subplot(121)
plt.plot(fpr_average1,
tpr_average1,
label='Train AUC = {0:.2f}'.format(roc_auc))
plt.subplot(122)
plt.plot([-0.5, 1.5], [-0.5, 1.5])
plt.plot(fpr_average2,
tpr_average2,
label='Test AUC = {0:.2f}'.format(roc_auc))
plt.xlim([-0.05, 1.05])
plt.ylim([-0.05, 1.05])
plt.xlabel('False Positive Rate')
plt.ylabel('True Positive Rate')
plt.title('Probability of Default: ROC curve')
plt.legend(loc="lower right")
def main(args):
############################################################
# PATHS
############################################################
BASE_PATH = os.path.join(os.path.dirname(os.path.realpath(__file__)),
'../../data')
# BASE_PATH = os.path.join(os.path.dirname(os.path.realpath(__file__)),
# '../../data')
DIR_PATH = os.path.join(BASE_PATH, args.dataname)
DATA_PATH = os.path.join(DIR_PATH, 'vintage_analysis', 'data')
EXPORT_PATH = os.path.join(DIR_PATH, 'vintage_analysis', 'results')
ECON_PATH = os.path.join(BASE_PATH, 'economic')
FILENAME = args.filename
############################################################
# HELPER FXNS
############################################################
def plot_path(fname):
return os.path.join(EXPORT_PATH, fname)
############################################################
# READ/PROCESS/CLEAN DATA
############################################################
df = pd.read_csv(os.path.join(DATA_PATH, FILENAME),
parse_dates=['PRD', 'ORIG_DTE'])
# attach econ vars
df_econ = pd.read_csv(os.path.join(ECON_PATH, 'agg_ntnl_mnthly.csv'),
parse_dates=['DATE'])
df = df.merge(df_econ, how='left', left_on='PRD',
right_on='DATE', copy=False)
# delete unnecessary variables
del df['DATE'], df['Unnamed: 0']
# change date format
df.loc[:, 'PRD'] = df.loc[:, 'PRD'].dt.to_period('m')
df.loc[:, 'ORIG_DTE'] = df.loc[:, 'ORIG_DTE'].dt.to_period('m')
# create age and other vars
df['AGE'] = (df['PRD'] - df['ORIG_DTE']).astype(int)
df['dflt_pct'] = df['DFLT_AMT'] / df['ORIG_AMT_sum']
df['min_dflt'] = 1*(df['dflt_pct'] > 0)
df['net_loss_pct'] = 0
df.loc[df['min_dflt'] == 1,
'net_loss_pct'] = df['NET_LOSS_AMT'] / df['DFLT_AMT']
# what we ultiamtely want to predict
df['final_loss_pct'] = df['NET_LOSS_AMT'] / df['ORIG_AMT_sum']
# # create other variables
# regex = re.compile('_wv$')
# wghted_cols = [regex.sub('', c) for c in df.columns if regex.search(c)]
# for col in wghted_cols:
# df['{0}_cv'.format(col)] = np.sqrt(
# df['{0}_wv'.format(col)])/df['{0}_wm'.format(col)]
# drop columns with too many NA values
df = utility.drop_NA_cols(df)
# remove all NA rows
a = df.shape
df.dropna(axis=0, how='any', inplace=True)
print('Reduced rows from {0} -> {1}'.format(a, df.shape))
# isolate covariates
regex = re.compile('^MI_TYPE|^MI_PCT')
all_vars = [
v for v in df.columns if v not in MAIN_VARS and not regex.search(v)]
############################################################
# PLOTTING
############################################################
yr2grp = {'00-01': [2000, 2001],
'02-05': [2002, 2003, 2004, 2005],
'06-08': [2006, 2007, 2008],
'09-10': [2009, 2010]}
grp2yr = {yr: k for k, v in yr2grp.items() for yr in v}
# categorize years by yr2grp
df['ORIG_YR'] = df.ORIG_DTE.apply(lambda x: x.year)
df['ORIG_YR_GRP'] = df.ORIG_YR.apply(lambda yr: grp2yr[yr] if yr in grp2yr
else 'OTHER')
df_pos = df[df['min_dflt'] == 1]
#################### ag vs. dflt ####################
to_plot = df[['AGE', 'dflt_pct']]
# find median loss per age
meds = to_plot.groupby('AGE').median()
x_uniq, y_median = meds.index, meds.values
# fit bspline of deg = 3
sfit = splrep(x=meds.index, y=meds,
k=3, t=np.quantile(x_uniq, [0.33,0.66]))
x, y = to_plot['AGE'], to_plot['dflt_pct']
all_x = np.arange(x.min(), x.max() + 1)
fig, ax = plt.subplots(figsize=(8,8))
ax.scatter(x=x.values, y=y.values, s=1, c='#F8C471')
ax.plot(all_x, splev(all_x, sfit), c='r', lw=3,
label='B-Spline Fit')
ax.set_xlabel('AGE')
ax.set_ylabel('dflt_pct')
ax.legend()
ax.set_ylim(-0.0003, None)
fig.suptitle('% Default over the Age of Vintage')
fig.tight_layout(rect=[0, 0.03, 1, 0.95])
fig.savefig(plot_path('dflt_age.png'))
plt.close('all')
#################### age vs. dflt by year ####################
fig, ax = plt.subplots(figsize=(8,8))
ax.scatter(x=df['AGE'].values, y=df['dflt_pct'].values, s=1, c='#F8C471')
for orig_yr, df_yr in df.groupby('ORIG_YR'):
meds = df_yr.groupby('AGE')['dflt_pct'].median()
x_uniq, y_median = meds.index, meds.values
sfit = splrep(x=x_uniq, y=y_median,
k=3, t=np.quantile(x_uniq, [0.33,0.66]))
x, y = df_yr['AGE'], df_yr['dflt_pct']
all_x = np.arange(x.min(), x.max() + 1)
ax.plot(all_x, splev(all_x, sfit), lw=2,
label=str(orig_yr))
ax.set_xlabel('AGE')
ax.set_ylabel('dflt_pct')
ax.set_ylim(-0.0003, None)
ax.legend()
fig.suptitle('% Default over Age of Vintage by Year')
fig.tight_layout(rect=[0, 0.03, 1, 0.95])
fig.savefig(plot_path('dflt_age_by_year.png'))
plt.close('all')
#################### age vs. dflt by year up to < 2007 ####################
df_prerec = df[df.PRD < pd.Period('2007-01', freq='m')]
fig, ax = plt.subplots(figsize=(8,8))
ax.scatter(x=df_prerec['AGE'].values, y=df_prerec['dflt_pct'].values, s=1,
c='#F8C471')
for orig_yr, df_yr in df_prerec.groupby('ORIG_YR'):
meds = df_yr.groupby('AGE')['dflt_pct'].median()
x_uniq, y_median = meds.index, meds.values
sfit = splrep(x=x_uniq, y=y_median,
k=3, t=np.quantile(x_uniq, [0.33,0.66]))
x, y = df_yr['AGE'], df_yr['dflt_pct']
all_x = np.arange(x.min(), x.max() + 1)
ax.plot(all_x, splev(all_x, sfit), lw=2,
label=str(orig_yr))
ax.set_xlabel('AGE')
ax.set_ylabel('dflt_pct')
ax.set_ylim(-0.0003, None)
ax.legend()
fig.suptitle('% Default over the Age of Vintage pre-recession')
fig.tight_layout(rect=[0, 0.03,1, 0.95])
fig.savefig(plot_path('dflt_age_by_year_prerec.png'))
plt.close('all')
#################### age vs. dflt by year post rec ####################
df_postrec = df[df.ORIG_YR >= 2009]
fig, ax = plt.subplots(figsize=(8,8))
ax.scatter(x=df_postrec['AGE'].values, y=df_postrec['dflt_pct'].values, s=1,
c='#F8C471')
for orig_yr, df_yr in df_postrec.groupby('ORIG_YR'):
meds = df_yr.groupby('AGE')['dflt_pct'].median()
x_uniq, y_median = meds.index, meds.values
sfit = splrep(x=x_uniq, y=y_median,
k=3, t=np.quantile(x_uniq, [0.33,0.66]))
x, y = df_yr['AGE'], df_yr['dflt_pct']
all_x = np.arange(x.min(), x.max() + 1)
ax.plot(all_x, splev(all_x, sfit), lw=2,
label=str(orig_yr))
ax.set_xlabel('AGE')
ax.set_ylabel('dflt_pct')
ax.set_ylim(-0.0003, None)
ax.legend()
fig.suptitle('% Default over the Age of Vintage post-recession')
fig.tight_layout(rect=[0, 0.03, 1, 0.95])
fig.savefig(plot_path('dflt_age_by_year_postrec.png'))
plt.close('all')
# the dflt curve does peak early and slope down
# but hard to untangle age effects from recession/macro effects
############################################################
# dflt_pct
############################################################
#################### plot distribtiosn ####################
to_hist, yr_grps = [], []
for orig_yr_grp, df_yr in df.groupby('ORIG_YR_GRP'):
data = df_yr['dflt_pct']
to_hist.append(data)
yr_grps.append('{0}: {1:.1f}% = no dflt'
.format(orig_yr_grp,
100 * (data==0).mean()))
fig, ax = plt.subplots(figsize=(8, 8))
ax.hist(to_hist, label=yr_grps,
# normalize the histograms
density=True)
ax.set_xlabel('dflt_pct')
ax.legend(loc='best')
fig.suptitle('Distribution of % Default per Group')
fig.tight_layout(rect=[0, 0.03, 1, 0.95])
fig.savefig(plot_path('dflt_pct dist by grp.png'))
plt.close('all')
#################### plot relationships ####################
to_plot = df[['dflt_pct', 'CSCORE_MN_wm', 'DTI_wm', 'LOAN_ID_count',
'ORIG_YR_GRP']]
sns.pairplot(to_plot, hue='ORIG_YR_GRP', markers='.', height=2)
plt.tight_layout()
plt.savefig(plot_path('pairplot of dflt_pct'))
to_plot = df[['dflt_pct', 'MR', 'rGDP', 'AGE', 'ORIG_YR_GRP']]
sns.pairplot(to_plot, hue='ORIG_YR_GRP', markers='.', height=2)
plt.tight_layout()
plt.savefig(plot_path('pairplot of dflt_pct2'))
############################################################
# net_loss_pct
############################################################
#################### distributions ####################
fig, ax = plt.subplots(figsize=(8, 8))
for orig_yr_grp, df_yr in df_pos.groupby('ORIG_YR'):
ax.hist(df_yr['net_loss_pct'], label=str(orig_yr_grp),
# normalize the histograms
alpha=0.5,
density=True)
ax.set_xlabel('net_loss_pct')
ax.legend(loc='best')
fig.suptitle('Distribution: % of Default Amount Lost per Group')
fig.tight_layout(rect=[0, 0.03, 1, 0.95])
fig.savefig(plot_path('net_loss_pct dist by yr'))
plt.close('all')
#################### relationships ####################
to_plot = df[['net_loss_pct', 'CSCORE_MN_wm', 'DTI_wm', 'LOAN_ID_count',
'ORIG_YR_GRP']]
sns.pairplot(to_plot, hue='ORIG_YR_GRP', markers='.', height=2)
plt.tight_layout()
plt.savefig(plot_path('pairplot of net_loss_pct'))
to_plot = df[['net_loss_pct', 'MR', 'rGDP', 'AGE', 'ORIG_YR_GRP']]
sns.pairplot(to_plot, hue='ORIG_YR_GRP', markers='.', height=2)
plt.tight_layout()
plt.savefig(plot_path('pairplot of net_loss_pct2'))
# not as strong of relationship as dflt_pct
# but still some structue (Ex. MR)
############################################################
# Low default analysis
############################################################
cutoff_date = pd.Period('2009-06-01', freq='m')
df_low = df[df.ORIG_DTE >= cutoff_date]
df_high = df[df.ORIG_DTE < cutoff_date]
fig, ax = plt.subplots(figsize=(8, 8))
ax.hist([df_low['dflt_pct'],
df_high['dflt_pct']],
label=['Low Period', 'High Period'],
density=True)
ax.set_xlabel('dflt_pct')
ax.legend(loc='best')
fig.suptitle('Distribution: % Default pre- & post- 2009')
fig.tight_layout(rect=[0, 0.03, 1, 0.95])
fig.savefig(plot_path('low_dflt histogram'))
plt.close('all')
#################### plot relationships with dflt_pct ####################
to_plot = df_low[['dflt_pct', 'CSCORE_MN_wm', 'DTI_wm', 'LOAN_ID_count',
'ORIG_YR_GRP']]
sns.pairplot(to_plot, markers='.', height=2)
plt.tight_layout()
plt.savefig(plot_path('low_dflt: pairplot of dflt_pct'))
to_plot = df[['dflt_pct', 'MR', 'rGDP', 'AGE', 'ORIG_YR_GRP']]
sns.pairplot(to_plot, markers='.', height=2)
plt.tight_layout()
plt.savefig(plot_path('low_dflt: pairplot of dflt_pct2'))
# fortunately, there is still some relationship btw dflt_pct and vars
#################### plot relationships with net_loss_pct ####################
to_plot = df_low[['net_loss_pct', 'CSCORE_MN_wm', 'DTI_wm', 'LOAN_ID_count',
'ORIG_YR_GRP']]
sns.pairplot(to_plot, markers='.', height=2)
plt.tight_layout()
plt.savefig(plot_path('low_dflt: pairplot of net_loss_pct'))
to_plot = df[['net_loss_pct', 'MR', 'rGDP', 'AGE', 'ORIG_YR_GRP']]
sns.pairplot(to_plot, markers='.', height=2)
plt.tight_layout()
plt.savefig(plot_path('low_dflt: pairplot of net_loss_pct2'))
############################################################
# MODEL SPECS
############################################################
# Stage 1
# selected = ['UNEMP', 'HPI', 'rGDP', 'ORIG_AMT_sum', 'ORIG_CHN_R_wv',
# 'LOAN_ID_count', 'MR', 'LIBOR', 'DTI_wm', 'AGE', 'CPI',
# 'ORIG_RT_wm', 'ORIG_CHN_R_wm', 'ORIG_RT_wv',
# 'DTI_wv', 'OCLTV_wv',
# 'ORIG_CHN_C_wv',
# 'CSCORE_MN_wv', 'CSCORE_MN_wm']
selected = ['CSCORE_MN_wm', 'CSCORE_MN_wv',
'DTI_wm', 'DTI_wv', 'LOAN_ID_count',
'OCLTV_wm', 'OCLTV_wv',
'ORIG_AMT_sum',
'ORIG_CHN_C_wm',
'ORIG_CHN_R_wm',
'ORIG_RT_wm', 'ORIG_RT_wv',
'PURPOSE_P_wm', 'PURPOSE_P_wv',
'PURPOSE_R_wm', 'PURPOSE_R_wv',
'CPI', 'UNEMP', 'HPI', 'rGDP', 'AGE']
gbm_formula1 = 'min_dflt ~ -1 + {0}'.format(' + '.join(selected))
kw1 = {'learning_rate': 0.1, 'max_depth': 3, 'max_features': 0.5,
'min_impurity_decrease': 0.0001, 'n_estimators': 10, 'subsample': 0.1}
# Stage 2
rfr_formula2 = ('np.log(dflt_pct) ~ -1 + {0}'
.format(' + '.join(selected)))
kw2 = {'bootstrap': False, 'max_features': 0.333,
'min_impurity_decrease': 1e-10, 'n_estimators': 20}
# Stage 3
rfr_formula3 = ('net_loss_pct ~ -1 + {0}'
.format(' + '.join(selected)))
kw3 = {'n_estimators': 50, 'max_features': 0.1,
'min_impurity_decrease': 1e-10}
# define model_choice
model_specs = [['GBC', gbm_formula1, kw1],
['RFR', rfr_formula2, kw2],
['RFR', rfr_formula3, kw3]]
# create new folder if not exist
if not os.path.exists(EXPORT_PATH):
print('\nCreating results directory...')
os.makedirs(EXPORT_PATH)
output_dir_name = '.'.join([m[0] for m in model_specs] +
[FILENAME] +
[datetime.datetime.now()
.strftime("%y%m%d%H%M%S")])
output_dir_path = os.path.join(EXPORT_PATH, output_dir_name)
if not os.path.exists(output_dir_path):
print('\nCreating directory at export location...')
os.makedirs(output_dir_path)
for train, test, trial_i in utility.train_test_splitter(df_low,
0.1, 'ORIG_DTE'):
train_pos = train[train['dflt_pct'] > 0]
# stage 1
train['PD_pred'], test['PD_pred'], model1 = fit_stage1(model_specs[0],
train, test, True,
return_model=True)
# stage 2
train['EAD_pred'], test['EAD_pred'], model2 = fit_stage2(model_specs[1],
train_pos, test,
True, train,
return_model=True)
# stage 3
train['LGD_pred'], test['LGD_pred'], model3 = fit_stage3(model_specs[2],
train_pos, test,
True, train,
return_model=True)
# plot model summaries (just the first 10 most important)
print('\nGenerating summary plot...')
summ1 = model1.summary()[:10]
summ2 = model2.summary()[:10]
summ3 = model3.summary()[:10]
fig, axes = plt.subplots(2, 2, figsize=(8, 8))
ax = axes[0, 0]
ax.bar(summ1['Name'], summ1['Importance'])
ax.set_xticklabels(summ1['Name'], rotation=90)
ax.set_ylabel('Feature Importance')
ax.set_title('GBC model for Probability of Default')
ax = axes[0, 1]
ax.bar(summ2['Name'], summ2['Importance'],
yerr=summ2['Std'])
ax.set_xticklabels(summ3['Name'], rotation=90)
ax.set_ylabel('Feature Importance')
ax.set_title('RFR model for Exposure at Default')
ax = axes[1, 0]
ax.bar(summ3['Name'], summ3['Importance'],
yerr=summ3['Std'])
ax.set_xticklabels(summ3['Name'], rotation=90)
ax.set_ylabel('Feature Importance')
ax.set_title('RFR model for Loss given Default')
ax = axes[1, 1]
ax.set_axis_off()
# make a dataframe of parameters for each model
param_list = [pd.DataFrame(
{'Model': '{0}{1}'
.format(model.__class__.__name__, i+1),
'Param':
list(model.params().keys()),
'Value':
list(model.params().values()),
})
for i, model in
enumerate([model1, model2, model3])]
num_rows = 5
param_df = pd.concat([df[:num_rows] for df in param_list])
colors = [['#F4D03F'] * 3] * num_rows +\
[['#F1948A'] * 3] * num_rows +\
[['#7DCEA0'] * 3] * num_rows
tab = ax.table(cellText=param_df.values,
bbox=[0,0,1,1.05],
cellColours=colors,
colLabels=param_df.columns,
colWidths=[0.05, 0.1, 0.1])
tab.auto_set_font_size(False)
tab.set_fontsize(9)
fig.tight_layout()
plotname = '{0}.model summary.png'.format(trial_i)
fig.savefig(os.path.join(output_dir_path,
plotname))
plt.close('all')
# bootstrapping
train['L_pred'] = train['PD_pred'] * \
train['EAD_pred'] * train['LGD_pred']
test['L_pred'] = test['PD_pred'] * test['EAD_pred'] * test['LGD_pred']
test_pos = test[test['dflt_pct'] > 0]
# get bootstraps
print('\nRunning bootstraps...')
btstrp_size = 30
btstrp_3stage = bootstrap.get_btstrp(fit_3stages, model_specs,
train, test, btstrp_size)
# stage 1 bootstraps
btstrp_stage1 = bootstrap.get_btstrp(fit_stage1, model_specs[0],
train, test, btstrp_size)
# stage 2 bootstraps
btstrp_stage2 = bootstrap.get_btstrp(fit_stage2, model_specs[1],
train_pos, test_pos, btstrp_size)
# stage 2 bootstraps
btstrp_stage3 = bootstrap.get_btstrp(fit_stage3, model_specs[2],
train_pos, test_pos, btstrp_size)
vin_id, x, lo_name, hi_name = 'ORIG_DTE', 'AGE', '2.5%', '97.5%'
# compile bootstrapped data for each stages
df_all = pd.concat([test[[vin_id, x, 'final_loss_pct', 'L_pred']],
btstrp_3stage], axis=1)
df_stage1 = pd.concat([test[[vin_id, x, 'min_dflt', 'PD_pred']],
btstrp_stage1], axis=1)
df_stage2 = pd.concat([test_pos[[vin_id, x, 'dflt_pct', 'EAD_pred']],
btstrp_stage2], axis=1)
df_stage3 = pd.concat([test_pos[[vin_id, x,
'net_loss_pct', 'LGD_pred']],
btstrp_stage3], axis=1)
for date, dall in df_all.groupby(vin_id):
print('\nGenerating plots for {0}...'.format(date))
ds = []
for df_ in [df_stage1, df_stage2, df_stage3]:
to_append = df_[df_[vin_id] == date]
# remove vin_id column
del to_append[vin_id]
ds.append(to_append)
del dall[vin_id]
fig, axes = plt.subplots(2, 2, figsize=(10, 10))
# stage 1 roc
ax, d1 = axes[0, 0], ds[0]
fpr, tpr = utility.get_roc_curve(train['min_dflt'], train['PD_pred'])
roc_auc = utility.get_auc(fpr, tpr)
ax.plot(fpr, tpr, label='Train AUC = {0:.2f}'.format(roc_auc),
**PLOT_PARAMS['train roc curve'])
fpr, tpr = utility.get_roc_curve(d1['min_dflt'], d1['PD_pred'])
roc_auc = utility.get_auc(fpr, tpr)
ax.plot(fpr, tpr, label='Test AUC = {0:.2f}'.format(roc_auc),
**PLOT_PARAMS['test roc curve'])
ax.plot([-0.5, 1.5], [-0.5, 1.5], **PLOT_PARAMS['actual line'])
ax.set_xlim([-0.05, 1.05])
ax.set_ylim([-0.05, 1.05])
ax.set_xlabel('False Positive Rate')
ax.set_ylabel('True Positive Rate')
ax.set_title('Probability of Default: ROC curve')
ax.legend(loc="lower right")
# stage 2 predicted vs. actual
ax, d2 = axes[0, 1], ds[1]
ax.plot([-0.5, 1.5], [-0.5, 1.5], **PLOT_PARAMS['actual line'])
c, d = train['EAD_pred'], train['dflt_pct']
ax.scatter(c, d,
label='Train RMSE: {0:.2E}'.format(((c-d)**2).mean()),
**PLOT_PARAMS['train'])
a, b = d2['EAD_pred'], d2['dflt_pct']
ax.scatter(a, b,
label='Test RMSE: {0:.2E}'.format(((a-b)**2).mean()),
**PLOT_PARAMS['test'])
xymax = min(0.5, max(a.max(), b.max()) * 1.05)
xymin = max(-0.01, min(a.min(), b.min()) * 0.95)
ax.set_xlim([xymin, xymax])
ax.set_ylim([xymin, xymax])
ax.set_xlabel('Predicted')
ax.set_ylabel('Actual')
ax.set_title('Exposure at Default: % balance at default')
ax.legend(loc="lower right")
# # stage 2 bootstrap actual vs. predicted
# ax, d2 = axes[0, 1], ds[1]
# bootstrap.plot_btstrp(d2, 'dflt_pct', 'dflt_pct', 'EAD_pred',
# lo_name, hi_name, ax=ax)
# ax.set_xlabel('Actual')
# ax.set_ylabel('Predicted')
# ax.set_title('Exposure at Default: % balance at default')
# ax.legend(loc="lower right")
# stage 3 predicted vs. actual
ax, d3 = axes[1, 0], ds[2]
ax.plot([-10, 10], [-10, 10], **PLOT_PARAMS['actual line'])
c, d = train['LGD_pred'], train['net_loss_pct']
ax.scatter(c, d,
label='Train RMSE: {0:.2E}'.format(((c-d)**2).mean()),
**PLOT_PARAMS['train'])
a, b = d3['LGD_pred'], d3['net_loss_pct']
ax.scatter(a, b,
label='Test RMSE: {0:.2E}'.format(((a-b)**2).mean()),
**PLOT_PARAMS['test'])
xymax = min(2, max(a.max(), b.max()) * 1.05)
xymin = max(-0.5, min(a.min(), b.min()) * 0.95)
ax.set_xlim([xymin, xymax])
ax.set_ylim([xymin, xymax])
ax.set_xlabel('Predicted')
ax.set_ylabel('Actual')
ax.set_title('Loss given Default: % lost at default')
ax.legend(loc="lower right")
# # stage 3 bootstrap actual vs. predicted
# ax, d3 = axes[1, 0], ds[2]
# bootstrap.plot_btstrp(d3, 'net_loss_pct', 'net_loss_pct', 'LGD_pred',
# lo_name, hi_name, ax=ax)
# ax.set_xlabel('Actual')
# ax.set_ylabel('Predicted')
# ax.set_title('Loss given Default: % lost at default')
# ax.legend(loc="lower right")
# 3 stage bootstrap cumulative
ax = axes[1, 1]
cumsum = dall.set_index(x).cumsum()
bootstrap.plot_btstrp(cumsum.reset_index(), x, 'final_loss_pct',
'L_pred', lo_name, hi_name, ax=ax)
ax.set_xlabel('Age of loan')
ax.set_ylabel('% of ORIG_AMT lost')
loan_count = test.loc[test.ORIG_DTE ==
date, 'LOAN_ID_count'].values[0]
ax.set_title(
'Projected Loss: {0} loans from {1}'.format(loan_count, date))
ax.legend(loc="lower right")
# plot settings
plt.tight_layout()
plotname = '{2}.{1}.{0}.png'.format(loan_count, date, trial_i)
plt.savefig(os.path.join(output_dir_path, plotname),
dpi=300, bbox_inches='tight')
# plt.show()
plt.close('all')
# Test kmf on test set
kmf = KaplanMeierFitter()
cph = CoxPHFitter()
# right censoring must be independent of dflt
# check if loans that didn't dflt all go to the last period
df_cox[df_cox.did_dflt == 0].PRD.value_counts()
############################################################
# NOTE: Cox is always underpredicting bc many in test set are censored
# Thus not had the chance to experience the full extent
############################################################
ts = int(df_cox.shape[0] * 0.1)
ts = 4
for train, test, trial_i in utility.train_test_splitter(df_cox,
5,
ts,
key='ORIG_DTE'):
break
kmf.fit(train["AGE"], event_observed=train["did_dflt"])
test['year'] = test.ORIG_DTE.apply(lambda x: x.year)
year_cutoff = 2005
rtest = test[test.year <= year_cutoff]
rtest = test
T, E = rtest['AGE'], rtest['did_dflt']
did_die = E == 1
for time_cutoff in range(72, 132, 12):
# remove loans censored before time_cutoff
def main(args):
############################################################
# READ/PROCESS/CLEAN DATA
############################################################
BASE_PATH = os.path.join(ROOT_PATH, 'data')
DIR_PATH = os.path.join(BASE_PATH, args.dataname)
DATA_PATH = os.path.join(DIR_PATH, 'vintage_analysis', 'data')
EXPORT_PATH = os.path.join(DIR_PATH, 'vintage_analysis', 'results')
ECON_PATH = os.path.join(BASE_PATH, 'economic')
FILENAME = args.filename
savemodel = args.DO_SAVE
df = pd.read_csv(os.path.join(DATA_PATH, FILENAME),
parse_dates=['PRD', 'ORIG_DTE'])
# attach econ vars
df_econ = pd.read_csv(os.path.join(ECON_PATH, 'agg_ntnl_mnthly.csv'),
parse_dates=['DATE'])
df = df.merge(df_econ,
how='left',
left_on='PRD',
right_on='DATE',
copy=False)
# delete unnecessary variables
del df['DATE'], df['Unnamed: 0']
# change date format
df.loc[:, 'PRD'] = df.loc[:, 'PRD'].dt.to_period('m')
df.loc[:, 'ORIG_DTE'] = df.loc[:, 'ORIG_DTE'].dt.to_period('m')
# create age and other vars
df['AGE'] = (df['PRD'] - df['ORIG_DTE']).astype(int)
df['dflt_pct'] = df['DFLT_AMT'] / df['ORIG_AMT_sum']
df['did_dflt'] = 1 * (df['dflt_pct'] > 0)
df['net_loss_pct'] = 0
df.loc[df['did_dflt'] == 1,
'net_loss_pct'] = df['NET_LOSS_AMT'] / df['DFLT_AMT']
# what we ultiamtely want to predict
df['dflt_loss_pct'] = df['NET_LOSS_AMT'] / df['ORIG_AMT_sum']
# create other variables
regex = re.compile('_wv$')
wghted_cols = [regex.sub('', c) for c in df.columns if regex.search(c)]
for col in wghted_cols:
df['{0}_cv'.format(col)] = np.sqrt(
df['{0}_wv'.format(col)]) / df['{0}_wm'.format(col)]
# drop columns with too many NA values
df = utility.drop_NA_cols(df)
# remove all NA rows
a = df.shape
df.dropna(axis=0, how='any', inplace=True)
print('Reduced rows from {0} -> {1}'.format(a, df.shape))
# isolate covariates
regex = re.compile('^MI_TYPE|^MI_PCT')
all_vars = [
v for v in df.columns if v not in MAIN_VARS and not regex.search(v)
]
############################################################
# MODEL SPECS
############################################################
# Stage 1
selected = ['UNEMP', 'HPI', 'rGDP', 'ORIG_AMT_sum']
gbm_formula1 = 'did_dflt ~ -1 + {0}'.format(' + '.join(selected))
kw1 = {
'learning_rate': 0.1,
'max_depth': 3,
'max_features': 0.5,
'min_impurity_decrease': 0.0001,
'n_estimators': 10,
'subsample': 0.1
}
# Stage 2
# selected2 = ['np.log(ORIG_AMT_sum)*cr(AGE, df=5)', 'PROP_TYP_MH_wm*DTI_wm',
# 'LIBOR', 'ORIG_TRM_cv', 'ORIG_RT_wm*MR',
# 'ORIG_CHN_C_wm*np.log(ORIG_AMT_sum)', 'DTI_wm*UNEMP',
# '(OCC_STAT_P_wm+OCC_STAT_S_wm)', '(PURPOSE_P_wm+PURPOSE_R_wm)',
# '(ORIG_CHN_C_wm+ORIG_CHN_R_wm)',
# '(PROP_TYP_CP_wm+PROP_TYP_MH_wm+PROP_TYP_PU_wm+PROP_TYP_SF_wm)']
# gam_formula2 = ('np.log(dflt_pct) ~ {0}'.format(' + '.join(selected2)))
rfr_formula2 = ('np.log(dflt_pct) ~ -1 + {0}'.format(' + '.join(all_vars)))
kw2 = {
'bootstrap': False,
'max_features': 0.333,
'min_impurity_decrease': 1e-10,
'n_estimators': 20
}
# Stage 3
# selected3 = ['AGE*(UNEMP + MR)', '(DTI_wm + DTI_wv) * UNEMP', 'HPI', 'LIBOR',
# 'PROP_TYP_CP_wm + PROP_TYP_MH_wm', 'np.sqrt(LOAN_ID_count)']
# gam_formula3 = ('net_loss_pct ~ {0}'.format(' + '.join(selected3)))
rfr_formula3 = ('net_loss_pct ~ -1 + {0}'.format(' + '.join(all_vars)))
kw3 = {
'n_estimators': 50,
'max_features': 0.1,
'min_impurity_decrease': 1e-10
}
# define model_choice
model_specs = [['GBC', gbm_formula1, kw1], ['RFR', rfr_formula2, kw2],
['RFR', rfr_formula3, kw3]]
# create new folder if not exist
if not os.path.exists(EXPORT_PATH):
print('\nCreating results directory...')
os.makedirs(EXPORT_PATH)
output_dir_name = '.'.join(
[m[0] for m in model_specs] + [FILENAME] +
[datetime.datetime.now().strftime("%y%m%d%H%M%S")])
output_dir_path = os.path.join(EXPORT_PATH, output_dir_name)
if not os.path.exists(output_dir_path):
print('\nCreating directory at export location...')
os.makedirs(output_dir_path)
path_index = 1
for train, test, trial_i in utility.train_test_splitter(
df, 0.1, 'ORIG_DTE'):
train_pos = train[train['dflt_pct'] > 0]
# stage 1
train['PD_pred'], test['PD_pred'] = fit_stage1(model_specs[0], train,
test, True)
# model 1
model1 = get_fitted_model(train, test, model_specs[0])
# stage 2
train['EAD_pred'], test['EAD_pred'] = fit_stage2(
model_specs[1], train_pos, test, True, train)
# model 2
model2 = get_fitted_model(train_pos, test, model_specs[1])
# stage 3
train['LGD_pred'], test['LGD_pred'] = fit_stage3(
model_specs[2], train_pos, test, True, train)
# model 3
model3 = get_fitted_model(train_pos, test, model_specs[2])
train['L_pred'] = train['PD_pred'] * \
train['EAD_pred'] * train['LGD_pred']
test['L_pred'] = test['PD_pred'] * test['EAD_pred'] * test['LGD_pred']
test_pos = test[test['dflt_pct'] > 0]
pathname = str(path_index)
train.to_csv(pathname)
if savemodel == True:
utility.save_3stages(model1, model2, model3, pathname)
path_index += 1
# get bootstraps
print('\nRunning bootstraps...')
btstrp_size = 4
btstrp_3stage = bootstrap.get_btstrp(fit_3stages, model_specs, train,
test, btstrp_size)
# stage 1 bootstraps
btstrp_stage1 = bootstrap.get_btstrp(fit_stage1, model_specs[0], train,
test, 1)
# stage 2 bootstraps
btstrp_stage2 = bootstrap.get_btstrp(fit_stage2, model_specs[1],
train_pos, test_pos, 1)
# stage 2 bootstraps
btstrp_stage3 = bootstrap.get_btstrp(fit_stage3, model_specs[2],
train_pos, test_pos, 1)
vin_id, x, lo_name, hi_name = 'ORIG_DTE', 'AGE', '2.5%', '97.5%'
# compile bootstrapped data for each stages
df_all = pd.concat(
[test[[vin_id, x, 'dflt_loss_pct', 'L_pred']], btstrp_3stage],
axis=1)
df_stage1 = pd.concat(
[test[[vin_id, x, 'did_dflt', 'PD_pred']], btstrp_stage1], axis=1)
df_stage2 = pd.concat(
[test_pos[[vin_id, x, 'dflt_pct', 'EAD_pred']], btstrp_stage2],
axis=1)
df_stage3 = pd.concat(
[test_pos[[vin_id, x, 'net_loss_pct', 'LGD_pred']], btstrp_stage3],
axis=1)
for date, dall in df_all.groupby(vin_id):
print('\nGenerating plots for {0}...'.format(date))
ds = []
for df_ in [df_stage1, df_stage2, df_stage3]:
to_append = df_[df_[vin_id] == date]
# remove vin_id column
del to_append[vin_id]
ds.append(to_append)
del dall[vin_id]
fig, axes = plt.subplots(2, 2, figsize=(10, 10))
# stage 1 roc
ax, d1 = axes[0, 0], ds[0]
fpr, tpr = utility.get_roc_curve(train['did_dflt'],
train['PD_pred'])
roc_auc = utility.get_auc(fpr, tpr)
ax.plot(fpr,
tpr,
label='Train AUC = {0:.2f}'.format(roc_auc),
**PLOT_PARAMS['train roc curve'])
fpr, tpr = utility.get_roc_curve(d1['did_dflt'], d1['PD_pred'])
roc_auc = utility.get_auc(fpr, tpr)
ax.plot(fpr,
tpr,
label='Test AUC = {0:.2f}'.format(roc_auc),
**PLOT_PARAMS['test roc curve'])
ax.plot([-0.5, 1.5], [-0.5, 1.5], **PLOT_PARAMS['actual line'])
ax.set_xlim([-0.05, 1.05])
ax.set_ylim([-0.05, 1.05])
ax.set_xlabel('False Positive Rate')
ax.set_ylabel('True Positive Rate')
ax.set_title('Probability of Default: ROC curve')
ax.legend(loc="lower right")
# stage 2 predicted vs. actual
ax, d2 = axes[0, 1], ds[1]
ax.plot([-0.5, 1.5], [-0.5, 1.5], **PLOT_PARAMS['actual line'])
c, d = train['EAD_pred'], train['dflt_pct']
ax.scatter(c,
d,
label='Train RMSE: {0:.2E}'.format(((c - d)**2).mean()),
**PLOT_PARAMS['train'])
a, b = d2['EAD_pred'], d2['dflt_pct']
ax.scatter(a,
b,
label='Test RMSE: {0:.2E}'.format(((a - b)**2).mean()),
**PLOT_PARAMS['test'])
xymax = min(0.5, max(a.max(), b.max()) * 1.05)
xymin = max(-0.01, min(a.min(), b.min()) * 0.95)
ax.set_xlim([xymin, xymax])
ax.set_ylim([xymin, xymax])
ax.set_xlabel('Predicted')
ax.set_ylabel('Actual')
ax.set_title('Exposure at Default: % balance at default')
ax.legend(loc="lower right")
# # stage 2 bootstrap actual vs. predicted
# ax, d2 = axes[0, 1], ds[1]
# bootstrap.plot_btstrp(d2, 'dflt_pct', 'dflt_pct', 'EAD_pred',
# lo_name, hi_name, ax=ax)
# ax.set_xlabel('Actual')
# ax.set_ylabel('Predicted')
# ax.set_title('Exposure at Default: % balance at default')
# ax.legend(loc="lower right")
# stage 3 predicted vs. actual
ax, d3 = axes[1, 0], ds[2]
ax.plot([-10, 10], [-10, 10], **PLOT_PARAMS['actual line'])
c, d = train['LGD_pred'], train['net_loss_pct']
ax.scatter(c,
d,
label='Train RMSE: {0:.2E}'.format(((c - d)**2).mean()),
**PLOT_PARAMS['train'])
a, b = d3['LGD_pred'], d3['net_loss_pct']
ax.scatter(a,
b,
label='Test RMSE: {0:.2E}'.format(((a - b)**2).mean()),
**PLOT_PARAMS['test'])
xymax = min(2, max(a.max(), b.max()) * 1.05)
xymin = max(-0.5, min(a.min(), b.min()) * 0.95)
ax.set_xlim([xymin, xymax])
ax.set_ylim([xymin, xymax])
ax.set_xlabel('Predicted')
ax.set_ylabel('Actual')
ax.set_title('Loss given Default: % lost at default')
ax.legend(loc="lower right")
# # stage 3 bootstrap actual vs. predicted
# ax, d3 = axes[1, 0], ds[2]
# bootstrap.plot_btstrp(d3, 'net_loss_pct', 'net_loss_pct', 'LGD_pred',
# lo_name, hi_name, ax=ax)
# ax.set_xlabel('Actual')
# ax.set_ylabel('Predicted')
# ax.set_title('Loss given Default: % lost at default')
# ax.legend(loc="lower right")
# 3 stage bootstrap cumulative
ax = axes[1, 1]
cumsum = dall.set_index(x).cumsum()
bootstrap.plot_btstrp(cumsum.reset_index(),
x,
'dflt_loss_pct',
'L_pred',
lo_name,
hi_name,
ax=ax)
ax.set_xlabel('Age of loan')
ax.set_ylabel('% of ORIG_AMT lost')
loan_count = test.loc[test.ORIG_DTE == date,
'LOAN_ID_count'].values[0]
ax.set_title('Projected Loss: {0} loans from {1}'.format(
loan_count, date))
ax.legend(loc="lower right")
# plot settings
plt.tight_layout()
plotname = '{2}.{1}.{0}.png'.format(loan_count, date, trial_i)
plt.savefig(os.path.join(output_dir_path, plotname),
dpi=300,
bbox_inches='tight')
plt.show()
plt.clf()