def main(cutoff_date):
s_ut.my_print('loading to sup.dim_cx_ticket_forecast the forecast with cutoff date ' + str(cutoff_date))
t_file = os.path.expanduser('~/Forecasts/par/' + 'table_output_' + str(cutoff_date) + '.par')
s_ut.my_print('table file: ' + str(t_file))
if os.path.isfile(t_file):
df = p_ut.read_df(t_file)
p_ut.set_week_start(df, tcol='ds') # week_starting patch
# week_starting patch
df_cols_ = df.columns
if 'ds_week_ending' in df_cols_ and 'ds_week_starting' not in df_cols_:
df['ds_week_ending'] = pd.to_datetime(df['ds_week_ending'])
df['ds_week_starting'] = df['ds_week_ending'] - pd.to_timedelta(6, unit='D')
s_ut.my_print('data file: ' + str(t_file) + ' rows: ' + str(len(df)) + ' to table')
partition = {'ds': str(cutoff_date)}
table = 'sup.dim_cx_ticket_forecast'
ap.hive.push(df, table=table, if_exists='replace', partition=partition,
table_props={'abb_retention_days': '-1', 'abb_retention_days_reason': 'fact table. No pii'}
)
return 0
else:
s_ut.my_print('ERROR: failed to load: file ' + t_file + ' is missing')
return -1
def table_load(dr_, cutoff_date_, m_adj=1.0):
# load to adjusted data to the table
# read table files '~/Forecasts/par/' + 'table_output_'
gcols = [
'dim_business_unit', 'dim_language', 'dim_tier', 'dim_channel',
'time_interval'
]
t_list, max_ds_ = list(), None
for d_ in dr_:
fname_ = os.path.expanduser('~/Forecasts/par/table_output_' +
str(d_.date())) + '.par'
s_ut.my_print('rolling date: ' + str(d_.date()) + ' fname: ' +
str(fname_))
if os.path.exists(fname_):
fx = p_ut.read_df(fname_)
p_ut.set_week_start(
fx, tcol='fcst_date_inv_endings') # week_starting patch
# week_starting patch
df_cols__ = fx.columns
if 'ds_week_ending' in df_cols__ and 'ds_week_starting' not in df_cols__:
fx['ds_week_ending'] = pd.to_datetime(fx['ds_week_ending'])
fx['ds_week_starting'] = fx[
'ds_week_ending'] - pd.to_timedelta(6, unit='D')
fv = process_w_df(fx, cutoff_date_, 'fcst_date_inv_ending',
gcols + ['run_date_inv_ending'])
max_ds_ = fv['fcst_date_inv_ending'].max(
) if max_ds_ is None else min(max_ds_,
fv['fcst_date_inv_ending'].max())
t_list.append(fv)
tdf = pd.concat(t_list, axis=0)
t_fdf = tdf.groupby(gcols + ['fcst_date_inv_ending']).apply(
lambda x: x['ticket_count'].mean()).reset_index()
t_fdf.rename(columns={0: 'ticket_count'}, inplace=True)
avg_tdf = t_fdf[t_fdf['fcst_date_inv_ending'] <= max_ds_].copy()
avg_tdf['run_date_inv_ending'] = str(cutoff_date_.date())
avg_tdf.reset_index(inplace=True)
avg_tdf.rename(columns={'index': 'fcst_horizon'}, inplace=True)
avg_tdf['fcst_date_inv_ending'] = avg_tdf[
'fcst_date_inv_ending'].dt.date.astype(str)
avg_tdf['ticket_count'] *= m_adj
print(
'******* saving data to load to sup.dim_cx_ticket_forecast >>>>>>>>>>>>>>'
)
p_ut.save_df(avg_tdf, '~/my_tmp/tab_data_' + str(cutoff_date_.date()))
print(888888888888888888888888)
print('---------------- SKIPPING TABLE ---------------------')
ret = -1
# ret = t2t.to_table(avg_tdf, str(cutoff_date_.date()), 'sup.dim_cx_ticket_forecast') # 'josep.dim_ticket_facst_test # 'sup.dim_cx_ticket_forecast'
if ret == -1:
s_ut.my_print('ERROR: table push failed')
return ret
def get_fcast_(ts_name, cutoff_date, e_date):
# get forecast
froot = '~/my_tmp/fbp/'
fname = froot + 'lang_fcast_' + ts_name + '_' + str(cutoff_date.date())
fcast_df = p_ut.read_df(fname)
if fcast_df is None:
s_ut.my_print('ERROR: no forecasts for ' + str(ts_name) +
' and cutoff date ' + str(cutoff_date.date()))
sys.exit()
elif fcast_df['ds'].max() <= e_date:
s_ut.my_print('ERROR: no forecasts for ' + str(ts_name) +
' and cutoff date ' + str(cutoff_date.date()) +
' and horizon ' + str(e_date.date()))
sys.exit()
else:
p_ut.set_week_start(fcast_df, tcol='ds') # week_starting patch
return fcast_df
def get_cfg_data(ts_name, cfg_cols, p_col):
# read all the cfgs and set the cfg_idx
t_name = 'sup.fct_cx_forecast_config'
qry = 'select * from ' + t_name + ';'
q_file = '/tmp/read_cfg_' + ts_name + '.hql'
with open(q_file, 'w') as f:
f.write(qry)
s_ut.my_print('pid: ' + str(os.getpid()) + ' in query file: ' + q_file)
fout = None
ret = hql.run_hql((q_file, q_file), fout)
if ret == -1:
s_ut.my_print('pid: ' + str(os.getpid()) +
' ERROR: Query failed. No configs found')
sys.exit()
s_ut.my_print('pid: ' + str(os.getpid()) + ' fcast cfg file: ' + ret)
cfg_df = p_ut.read_df(ret, sep='\t')
p_ut.set_week_start(cfg_df, tcol='ds') # week_starting patch
if cfg_df is None or len(cfg_df) == 0:
s_ut.my_print('pid: ' + str(os.getpid()) +
' ERROR: no data for query: ' + str(qry))
sys.exit()
dcol = {
x: x.replace(t_name.split('.')[-1] + '.', '')
for x in cfg_df.columns
}
cfg_df.rename(columns=dcol, inplace=True)
cfg_df['cutoff'] = pd.to_datetime(cfg_df['cutoff'])
cfg_df = cfg_df[(cfg_df['ts_name'] == ts_name)]
cfg_df = cfg_df[cfg_df[p_col] > 0.0].copy()
cfg_df.fillna('None', inplace=True)
cfg_df['cfg_str'] = cfg_df.apply(
lambda x: json.dumps(x[cfg_cols].to_dict()), axis=1)
z = cfg_df['cfg_str'].drop_duplicates()
zf = pd.DataFrame(z)
zf.reset_index(inplace=True, drop=True)
zf.reset_index(inplace=True)
zf.columns = ['cfg_idx', 'cfg_str']
df = cfg_df.merge(zf, on=['cfg_str'], how='left')
df['language'].replace(['Mandarin_Offshore', 'Mandarin_Onshore'],
'Mandarin',
inplace=True) # Mandarin need to be fixed later
df.drop_duplicates(inplace=True)
p_ut.save_df(df, '~/my_tmp/rk_df_' + ts_name)
# df = p_ut.read_df('~/my_tmp/rk_df_' + ts_name)
return df
def get_actuals_(ts_name, ts_cfg, e_date):
# get actuals
froot = ts_cfg['data_path'].split('/')
a_dir = os.path.expanduser('/'.join(froot[:-1])) + '/'
fname = froot[-1]
actuals_df = None
for f in os.listdir(a_dir):
if fname in f:
s_ut.my_print('actuals file: ' + a_dir + str(f.split('.')[0]))
actuals_df = p_ut.read_df(a_dir + f.split('.')[0])
if actuals_df is not None and actuals_df['ds'].max() >= e_date:
break
if actuals_df is None:
s_ut.my_print('ERROR: no actuals for ' + ts_name + ' and horizon ' +
str(e_date.date()))
sys.exit()
else:
p_ut.set_week_start(actuals_df, tcol='ds') # week_starting patch
return actuals_df
def get_fcast_cfg(ts_name, cutoff_date):
fdir = '~/my_tmp/cfg_sel/'
fperf = os.path.expanduser(fdir + 'cfg_sel_' + ts_name + '_' + cutoff_date)
fidx = os.path.expanduser(fdir + 'cfg_idx_' + ts_name + '_' + cutoff_date)
df_cfg = p_ut.read_df(fidx)
p_ut.set_week_start(df_cfg, tcol='ds') # week_starting patch
dfp = p_ut.read_df(fperf)
p_ut.set_week_start(dfp, tcol='ds') # week_starting patch
f_list = list()
for l, f in dfp.groupby('language'):
tf = f.nsmallest(n=1, columns=['f_err'])
cfg_list = list(tf.loc[tf.index[0], 'cfg_idx'][0])
print(l)
print(cfg_list)
fi = df_cfg[(df_cfg['language'] == l)
& (df_cfg['cfg_idx'].isin(cfg_list))]
fi.drop('f_err', axis=1, inplace=True)
fi.drop_duplicates(inplace=True)
print(fi)
f_list.append(fi)
return pd.concat(f_list) if len(f_list) > 0 else None
def get_fcast(cutoff_date_, froot, months=3):
# get the fcast issued <months> ago
f_month = 1 + (cutoff_date_.month - months) % 12 # fcast issue month
yr = cutoff_date_.year if f_month < cutoff_date_.month else cutoff_date_.year - 1
dm = pd.to_datetime(str(yr) + '-' + str(f_month) +
'-01') # 1st day of issue month
wd = dm.weekday()
fcast_sat = dm - pd.to_timedelta(
wd + 2, unit='D') if wd < 5 else dm - pd.to_timedelta(wd - 5, unit='D')
fcast_f = froot + str(fcast_sat.date())
try:
fdf = p_ut.read_df(os.path.expanduser(fcast_f))
except OSError:
s_ut.my_print('file not found: ' + froot)
return None
if fdf is None:
return None
else:
p_ut.set_week_start(cfg_df, tcol='ds') # week_starting patch
fdf.rename(columns={'ticket_count': 'forecasted_count'}, inplace=True)
s_ut.my_print('getting forecast from ' + fcast_f)
return fdf
def prepare_regs(r_name, rcfg, cutoff_date, fcast_days, init_date):
s_ut.my_print('pid: ' + str(os.getpid()) + ' preparing regressor ' +
str(r_name))
in_file, r_col_dict, key_cols = rcfg['data_path'], rcfg['r_col'], rcfg.get(
'key_cols', None)
# regressors: set deterministic indicators
if r_name == 'peak': # peak season indicator. No clean up, imputation or forecast
r_col = list(r_col_dict.keys())[0] # peaks
df = pd.DataFrame({
'ds':
pd.date_range(start=pd.to_datetime(init_date),
end=pd.to_datetime(cutoff_date) +
pd.to_timedelta(fcast_days, unit='D'),
freq='D')
})
df[r_col] = df['ds'].apply(
lambda x: 1 if x.month_name() in ['July', 'August'] else 0)
regressors.IndicatorRegressor('peak',
'peak',
'ds',
init_date,
cutoff_date, ['July', 'August'],
fcast_days,
dim_cols=None)
return df
# other regressors: clean up, imputation and forecast (later)
r_file = d_proc.get_data_file(rcfg['data_path'], cutoff_date)
if r_file is None:
s_ut.my_print('pid: ' + str(os.getpid()) + ' WARNING: date ' +
str(cutoff_date.date()) + ' has no data for regressor ' +
r_name)
return None
else:
s_ut.my_print('pid: ' + str(os.getpid()) +
' found data file for regressor ' + r_name +
' and date ' + str(cutoff_date.date()) + ': ' + r_file)
rdf = p_ut.read_df(r_file)
p_ut.set_week_start(rdf, tcol='ds') # week_starting patch
rdf = rdf[(rdf['ds'] >= pd.to_datetime(init_date))
& (rdf['ds'] <= pd.to_datetime(cutoff_date))].copy()
if key_cols is not None: # get only relevant data
for c, v in key_cols.items():
rdf = rdf[rdf[c] == v]
rdf['ceiling'] = rcfg.get('ceiling', 1)
rdf['floor'] = rcfg.get('floor', 0)
if r_name == 'contact-rate':
if len(rdf) > 0:
dim_cols = 'language' if 'language' in rdf.columns else None
regressors.Regressor('contact-rate',
'contact_rate',
'ds',
rdf,
rcfg,
init_date,
cutoff_date,
fcast_days,
dim_cols=dim_cols)
if 'language' in rdf.columns:
return rdf[['ds', 'language', 'contact_rate', 'ceiling', 'floor']]
else:
return rdf[['ds', 'contact_rate', 'ceiling', 'floor']]
elif r_name == 'tenure':
if len(rdf) > 0:
regressors.Regressor('tenure',
'tenure_days',
'ds',
rdf,
rcfg,
init_date,
cutoff_date,
fcast_days,
dim_cols=['language'])
return rdf[['ds', 'language', 'tenure_days']]
elif r_name == 'bookings' or r_name == 'checkins':
if len(rdf) > 0:
regressors.Regressor(r_name,
r_name[:-1] + '_count',
'ds',
rdf,
rcfg,
init_date,
cutoff_date,
fcast_days,
dim_cols=['language'])
return rdf
else:
s_ut.my_print('pid: ' + str(os.getpid()) +
' WARNING: unknown regressor: ' + str(r_name))
return None
def main(ts_name_, cutoff_date_):
cfg_cols = ['growth', 'y_mode', 'w_mode', 'r_mode', 'xform', 'h_mode', 'training', 'do_res', 'changepoint_range']
upr_horizon, lwr_horizon = 112, 84
lbl = ts_name_ + '_' + cutoff_date_
if ts_name_ == 'phone-inbound-vol':
fname = dtp.get_data_file('~/my_tmp/cleaned/phone-vol_cleaned_', cutoff_date_)
interaction_type = 'inbound'
else:
fname = dtp.get_data_file('~/my_tmp/cleaned/phone-aht_cleaned_', cutoff_date_)
interaction_type = 'inbound' if 'inbound' in ts_name_ else 'outbound'
# ######################################################################
# ######################################################################
# actuals
s_ut.my_print('pid: ' + str(os.getpid()) + ' actuals file: ' + str(fname))
q_df = pd.read_parquet(fname)
# week_starting patch
df_cols_ = q_df.columns
if 'ds_week_ending' in df_cols_ and 'ds_week_starting' not in df_cols_:
q_df['ds_week_ending'] = pd.to_datetime(q_df['ds_week_ending'])
q_df['ds_week_starting'] = q_df['ds_week_ending'] - pd.to_timedelta(6, unit='D')
q_df['ds'] = pd.to_datetime(q_df['ds'].values)
w_col = 'y' if 'vol' in ts_name_ else 'calls'
# daily actuals (language level)
if ts_name_ == 'phone-inbound-vol':
q_df = q_df.groupby(['ds','language']).agg({'offered': np.sum, 'accepted': np.sum, 'abandons': np.sum}).reset_index()
else:
q_df = q_df[q_df['interaction_type'] == interaction_type].copy()
q_df = q_df.groupby(['ds', 'language']).agg({'calls': np.sum, 'agent_mins': np.sum}).reset_index()
a_ddf, ctype = set_demand(q_df.copy(), 10, ts_name_)
w_df = a_ddf[a_ddf['ds'] <= cutoff_date_].groupby('language').agg({w_col: np.sum}).reset_index()
w_df.columns = ['language', 'weight']
p_ut.save_df(a_ddf, '~/my_tmp/a_daily_df_' + lbl)
# weekly level: use week starting
if ts_name_ == 'phone-inbound-vol':
m_df = q_df.groupby(pd.Grouper(key='ds', freq='W-SUN')).agg({'offered': np.sum, 'accepted': np.sum, 'abandons': np.sum}).reset_index()
else:
m_df = q_df.groupby(pd.Grouper(key='ds', freq='W-SUN')).agg({'calls': np.sum, 'agent_mins': np.sum}).reset_index()
a_wdf_, ctype = set_demand(m_df, 10, ts_name_)
a_wdf = a_wdf_.copy()
horizon_date = min(pd.to_datetime(cutoff_date_) + pd.to_timedelta(upr_horizon, unit='D'), a_wdf['ds'].max())
a_wdf['ds_week_ending'] = a_wdf['ds'] + pd.to_timedelta(6, unit='D') # switch to week ending so that we do not have incomplete weeks at end
a_wdf = a_wdf[(a_wdf['ds_week_ending'] <= horizon_date) & (a_wdf['ds_week_ending'] > cutoff_date_)].copy()
a_wdf.drop('ds', axis=1, inplace=True)
a_wdf['ts_name'] = ts_name_
p_ut.save_df(a_wdf, '~/my_tmp/a_weekly_df_' + lbl)
# ######################################################################
# ######################################################################
# ######################################################################
# ######################################################################
# DS forecasts: select the top fcast cfgs for each language, score them based on past performance and forecast them
sdir = '~/my_tmp/cfg_sel/'
df_best = p_ut.read_df(sdir + 'cfg_best_' + ts_name_ + '_' + cutoff_date_) # best ensembles by idx
p_ut.set_week_start(df_best, tcol='ds') # week_starting patch
z = df_best[['language', 'cfg_idx']].copy()
z.set_index('language', inplace=True)
dx = z.to_dict()['cfg_idx']
dx = {k: list(v) for k, v in dx.items()}
df_idx = p_ut.read_df(sdir + 'cfg_idx_' + ts_name_ + '_' + cutoff_date_) # map cfg_idx to fcast cfg
p_ut.set_week_start(df_idx, tcol='ds') # week_starting patch
df_idx = df_idx[['language', 'cfg_idx'] + cfg_cols].copy()
df_idx.drop_duplicates(inplace=True)
# fix None for fcasts
for c in cfg_cols:
df_idx[c] = df_idx[c].apply(lambda x: None if x == 'None' else x)
df_idx['h_mode'] = df_idx['h_mode'].apply(lambda x: True if x == 1 else False)
df_idx['do_res'] = df_idx['do_res'].apply(lambda x: True if x == 1 else False)
cfg_df = pd.concat([lf[lf['cfg_idx'].isin(dx[l])] for l, lf in df_idx.groupby('language')], axis=0)
# run the fcasts for the selected cfg's
file_out = lfc.main(ts_name_, cutoff_date_, cfg_cols, to_db=False, df_cfg=cfg_df.copy(), is_mp=True) # , is_fcast=False)
if file_out is None:
s_ut.my_print('pid: ' + str(os.getpid()) + ' ERROR: no fcast file returned')
sys.exit()
s_ut.my_print('pid: ' + str(os.getpid()) + ' +++++++++++++ completed forecasts:: file: ' + str(file_out) + ' +++++++++++++++ ')
fdf = pd.read_parquet(file_out)
# week_starting patch
df_cols_ = fdf.columns
if 'ds_week_ending' in df_cols_ and 'ds_week_starting' not in df_cols_:
fdf['ds_week_ending'] = pd.to_datetime(fdf['ds_week_ending'])
fdf['ds_week_starting'] = fdf['ds_week_ending'] - pd.to_timedelta(6, unit='D')
print(fdf.head())
# make sure fdf and perf_df have the same cfg's
# fdf_list = list(fdf['cfg_idx'].unique())
# perf_df = perf_df[perf_df['cfg_idx'].isin(fdf_list)].copy()
# cfg_df = cfg_df[cfg_df['cfg_idx'].isin(fdf_list)].copy()
# arr = [int(x * num_cfg_) for x in [1.0, 0.75, 0.5, 0.25, 0.125]]
# arg_list = [[fdf, a_ddf, w_df, cfg_df, ts_name_, cutoff_date_, horizon_date] for k in arr if k > 1]
# m_list = s_ut.do_mp(get_models, arg_list, is_mp=True, cpus=len(arr), do_sigkill=True) # list of dicts
# m_list = get_models(fdf, a_ddf, w_df, cfg_df, ts_name_, cutoff_date_, horizon_date)
m_list = get_models(fdf, a_ddf, w_df, ts_name_, cutoff_date_, horizon_date)
s_ut.my_print('pid: ' + str(os.getpid()) + ' ============== main: get_models complete. appending results ==========')
if len(m_list) > 0:
d_out = dict()
for dv in m_list:
for k, fname in dv.items():
if k not in d_out.keys():
d_out[k] = list()
fz = p_ut.read_df(fname)
p_ut.set_week_start(fz, tcol='ds') # week_starting patch
d_out[k].append(fz)
return {k: pd.concat(d_out[k], axis=0) for k, arr in d_out.items()}
else:
return None
def main(argv):
print(argv)
time_scale = 'W' # reset for daily ticket data
if len(sys.argv) == 2:
run_date = sys.argv[
1] # at least 3 days after last Saturday with actual data
with_bu = True
s_ut.my_print(
'WARNING: rerun not set in command line. Assuming no rerun')
elif len(sys.argv) == 3:
_, run_date, bu = sys.argv # at least 3 days after last Saturday with actual data
with_bu = bool(int(bu))
elif len(sys.argv) == 1:
with_bu = True
run_date = str(pd.to_datetime('today').date())
else:
print('invalid args: ' + str(sys.argv))
sys.exit()
cutoff_date = tm_ut.get_last_sat(run_date) # set to last saturday
if time_scale == 'W':
upr_horizon, lwr_horizon = 75, None
fcast_days = 7 * upr_horizon # regardless of time_scale
inc_start, inc_end = 4, 0
else:
upr_horizon, lwr_horizon = 75 * 7, None
fcast_days = upr_horizon
inc_start, inc_end = 28, 0
fcast_date = cutoff_date + pd.to_timedelta(fcast_days, unit='D')
# get actuals
act_df = p_ut.read_df('~/my_tmp/tix_act_df_' + str(cutoff_date.date()))
if act_df is None:
s_ut.my_print('ERROR: No actuals found')
sys.exit()
p_ut.set_week_start(act_df, tcol='ds') # week_starting patch
# get lang fcast
froot = '~/my_tmp/fbp_tix_'
fname = froot + 'lwbu_fcast_' if with_bu is True else froot + 'lnbu_fcast_'
fcast_df = p_ut.read_df(fname + str(cutoff_date.date()))
if fcast_df is None:
s_ut.my_print('ERROR: No fcast found')
sys.exit()
p_ut.set_week_start(fcast_df, tcol='ds') # week_starting patch
b_cols = ['agent_sector', 'channel']
g_cols = ['language']
if with_bu is False:
b_cols.append('business_unit')
fcast_df.drop('business_unit', inplace=True, axis=1) # all None
else:
g_cols.append('business_unit')
s_ut.my_print(
'------------------------- start biz level forecast from cutoff date '
+ str(cutoff_date.date()) + ' to forecast date ' +
str(fcast_date.date()) + ' with business columns: ' + str(b_cols) +
' ------------')
b_fcast = biz_fcast(fcast_df, act_df, g_cols, b_cols, cutoff_date,
time_scale, inc_start, inc_end)
if b_fcast is not None: # save all the fcats for eahc fcast cfg
froot = '~/my_tmp/fbp_tix_'
fname = froot + 'wbu_b_fcast_' if with_bu is True else froot + 'nbu_b_fcast_'
p_ut.save_df(b_fcast, fname + str(cutoff_date.date()))
else:
print('ERROR: no business fcast')
sys.exit()
# final fcast (ens_avg)
ens_df = ens_fcast(b_fcast, act_df, cutoff_date, g_cols, b_cols)
froot = '~/my_tmp/fbp_tix_'
fname = froot + 'wbu_e_fcast_' if with_bu is True else froot + 'nbu_e_fcast_'
p_ut.save_df(ens_df, fname + str(cutoff_date.date()))
print('++++++++++++++ Error Summary ++++++++++++')
# check for language error
fdf = ens_df.groupby(['ds', 'language']).agg({
'y_pred': np.sum
}).reset_index()
months = 3
m_start = pd.to_datetime(
str(cutoff_date.year) + '-' + str(cutoff_date.month) +
'-01') + pd.DateOffset(months=months + 1)
end_date = tm_ut.last_saturday_month(m_start) # max date for err check
collect_date = cutoff_date - pd.DateOffset(months=months)
start_date = end_date - pd.to_timedelta(
2, unit='W') # start date for error check
a_df, _ = t_ut.get_actuals(
end_date, collect_date) # actuals from collect date to end_date
fa = t_ut.set_act(a_df, ['language']) # clean TS for each language
fa = fa[(fa['ds'] > start_date) & (fa['ds'] <= end_date)].copy()
z = fa.merge(fdf, on=['ds', 'language'], how='left')
z = z[(z['y_pred'] > 0) & z['ticket_count'] > 0].copy()
z_lang = z.groupby('language').agg({
'ticket_count': np.sum,
'y_pred': np.sum
}).reset_index()
z_all = pd.DataFrame({
'language': ['All'],
'ticket_count': [z_lang['ticket_count'].sum()],
'y_pred': [z_lang['y_pred'].sum()]
})
z_lang = pd.concat([z_all, z_lang], axis=0)
z_lang['err'] = np.abs((z_lang['y_pred'] / z_lang['ticket_count']) - 1)
print(z_lang)
# t_ut.err_chk(ens_df, cutoff_date, [['language']], ycol='y_pred', months=3)
print('DONE')
def prepare_regs(r_name, rcfg, cutoff_date, fcast_days, int_type, init_date):
s_ut.my_print('pid: ' + str(os.getpid()) + ' preparing regressor ' +
str(r_name))
in_file, r_col_dict, key_cols = rcfg['data_path'], rcfg['r_col'], rcfg.get(
'key_cols', None)
# regressors: set deterministic indicators
if r_name == 'peak': # peak season indicator. No clean up, imputation or forecast
r_col = list(r_col_dict.keys())[0] # peaks
df = pd.DataFrame({
'ds':
pd.date_range(start=init_date,
end=cutoff_date +
pd.to_timedelta(fcast_days, unit='D'),
freq='D')
})
df[r_col] = df['ds'].apply(
lambda x: 1 if x.month_name() in ['July', 'August'] else 0)
return regressors.IndicatorRegressor('peak', 'peak', 'ds', init_date,
cutoff_date, ['July', 'August'],
fcast_days)
else: # other regressors
r_file = d_proc.get_data_file(rcfg['data_path'], cutoff_date)
if r_file is None:
s_ut.my_print('pid: ' + str(os.getpid()) + ' WARNING: date ' +
str(cutoff_date.date()) +
' has no data for regressor ' + r_name)
return None
else:
s_ut.my_print('pid: ' + str(os.getpid()) +
' found data file for regressor ' + r_name +
' and date ' + str(cutoff_date.date()) + ': ' +
r_file)
rdf = p_ut.read_df(r_file)
p_ut.set_week_start(rdf, tcol='ds') # week_starting patch
rdf = rdf[(rdf['ds'] >= pd.to_datetime(init_date))
& (rdf['ds'] <= cutoff_date)].copy()
if rdf['ds'].max() < pd.to_datetime(cutoff_date):
s_ut.my_print('WARNING: ' + r_name + ' max date (' +
str(rdf['ds'].max().date()) +
') is smaller than cutoff date (' +
str(cutoff_date.date()) + ')')
elif len(rdf) > 0:
if 'interaction_type' in rdf.columns:
rdf = rdf[rdf['interaction_type'] == int_type].copy()
if r_name == 'contact-rate':
if rdf['contact_rate'].min(
) == 0.0: # would mean no inbound tickets
zmin = rdf[rdf['contact_rate'] >
0.0]['contact_rate'].min() / 10.0
rdf['contact_rate'].replace(0.0, zmin, inplace=True)
return regressors.Regressor(
'contact-rate', 'contact_rate', 'ds',
rdf[['ds', 'language', 'contact_rate']], rcfg,
init_date, cutoff_date, fcast_days)
elif r_name == 'tenure':
rdf = rdf.groupby(['ds', 'language']).agg({
'tenure_days':
np.sum
}).reset_index()
return regressors.Regressor('tenure', 'tenure_days', 'ds',
rdf, rcfg, init_date,
cutoff_date, fcast_days)
elif r_name == 'bookings' or r_name == 'checkins':
return regressors.Regressor(r_name, r_name[:-1] + '_count',
'ds', rdf, rcfg, init_date,
cutoff_date, fcast_days)
else:
s_ut.my_print('pid: ' + str(os.getpid()) +
' WARNING: unknown regressor: ' +
str(r_name))
return None