1_prophet_model_evaluate.py

from prophet import Prophet
from hkjournalist import Journalist
import os
import sys
import pandas as pd
import matplotlib.pyplot as plt
import seaborn as sns

SCRIPT_DIR = os.path.dirname(os.path.abspath(__file__))
sys.path.append(os.path.dirname(SCRIPT_DIR))


# data and models from fbprophet tutorials https://facebook.github.io/prophet/docs/quick_start.html#python-api
# markdown report template './reports/simple_model_report_template.md'

sns.set(style="darkgrid")

# maintain a dict to record all variables used in reports
config = {}

# load data
df = pd.read_csv('./data/example_wp_log_peyton_manning.csv')
df['ds'] = pd.to_datetime(df['ds'])

# plot full data set (time series)
ax = df.set_index('ds').plot(figsize=(20, 10))
plt.tight_layout()
config['data_plot'] = ax

# train/test split
train_end_date = '20131231'
config['train_end_date'] = train_end_date
train_df = df[df['ds'] <= train_end_date]
test_df = df[df['ds'] > train_end_date]

test_df['year'] = test_df['ds'].dt.year
test_df['month'] = test_df['ds'].dt.month

# model build
model = Prophet(weekly_seasonality=True, yearly_seasonality=True)
model.add_seasonality('monthly', period=30.5, fourier_order=12, prior_scale=10)
model.add_seasonality('quarter', period=364.5 / 4,
                      fourier_order=10, prior_scale=5)

# model train & test
model.fit(train_df)
config['seasonality'] = pd.DataFrame(model.seasonalities)
test_df['y_pred'] = model.predict(test_df[['ds']])['yhat'].values
ax = test_df[['ds', 'y', 'y_pred']].set_index(
    'ds').plot(figsize=(20, 10))  # plot predict result
plt.tight_layout()
config['pred_plot'] = ax


# define a new kpi metric according to rules
def kpi_mape(df, y_true, y_pred):
    # mape group by yearmonth
    df[y_pred].clip(0, None, inplace=True)
    df['diff'] = abs(df[y_true] - df[y_pred])
    mape_df = df.groupby(['year', 'month']).agg(
        {'diff': 'sum', y_true: 'sum'}).reset_index()
    mape_df['mape'] = mape_df['diff'] / mape_df[y_true]
    res_df = mape_df.pivot(index='month', columns='year', values='mape')
    return res_df


config['metric_func'] = kpi_mape

# evaluate prediction on test set
kpi_df = kpi_mape(test_df, 'y', 'y_pred')
plt.figure(figsize=(4, 6))
ax = sns.heatmap(kpi_df, annot=True, cmap='YlGn', linewidth=.5, fmt='.2f')
plt.tight_layout()
config['error_plot'] = ax

# record extra info about model
config['note'] = "Prophet with no holidays info"

# Besides maintaining config dict, all code you need to generate reports is simply 3 lines below:
report_journalist = Journalist(
    template_file='./reports/1_prophet_report_template.md')
report_journalist.hear(config)
report_journalist.report(
    output_file='./reports/1_prophet_report.pdf', beamer=True, overwrite=False)