fbprophet - tutorials

Quick Start¶

In [10]:

import pandas as pd
from fbprophet import Prophet
from fbprophet.plot import add_changepoints_to_plot
import matplotlib.pyplot as plt
import matplotlib as mpl
import fbprophet
import pystan

In [12]:

assert fbprophet.__version__ >= '0.7.1', 'fbprophet 0.7.1 이상 필요'
assert pd.__version__ >= '1.0.4', 'pandas 1.0.4 이상 필요'
assert pystan.__version__ >= '2.19.1.1', 'pystan 2.19.1.1 이상 필요'

Prophet.make_future_dataframe

In [ ]:

df = pd.read_csv("./data_1907_2020.csv")

m = Prophet()
m.fit(df[-3650:])

future = m.make_future_dataframe(periods=30)
#future.tail()

INFO:numexpr.utils:NumExpr defaulting to 2 threads.
INFO:fbprophet:Disabling daily seasonality. Run prophet with daily_seasonality=True to override this.

Prophet.predict

yhat: 예측값

In [ ]:

forecast = m.predict(future)
#forecast.tail()

Prophet.plot

예측 데이터를 이용해 그래프를 그림

In [ ]:

fig1 = m.plot(forecast)

Prophet.plot_components 데이터의 구성요소를 연, 월, 주 단위로 나눔

In [ ]:

fig2 = m.plot_components(forecast)

Prophet.plot.plot_plotly 또한 plotly를 이용해서 그래프를 그릴 수 있음

In [ ]:

from fbprophet.plot import plot_plotly, plot_components_plotly

plot_plotly(m, forecast)

In [ ]:

plot_components_plotly(m, forecast)

Saturating Forecasts¶

상한, 하한을 정해서 예측

Forecasting Growth¶

cap을 이용해 상한을 제한

In [ ]:

df_2 = df
df_2['cap'] = 50

In [ ]:

m = Prophet(growth='logistic')
m.fit(df_2[-3650:])

INFO:fbprophet:Disabling daily seasonality. Run prophet with daily_seasonality=True to override this.

Out[ ]:

<fbprophet.forecaster.Prophet at 0x7f26784a5eb8>

In [ ]:

future = m.make_future_dataframe(periods=30)
future['cap'] = 50
fcst = m.predict(future)
fig = m.plot(fcst)

Saturating Minimum¶

floor를 이용해 하한값을 제한

In [ ]:

df_2['floor'] = -30

m = Prophet(growth='logistic')
m.fit(df_2[-3650:])

future = m.make_future_dataframe(periods=30)
future['cap'] = 50
future['floor'] = -30

fcst = m.predict(future)
fig = m.plot(fcst)

INFO:fbprophet:Disabling daily seasonality. Run prophet with daily_seasonality=True to override this.

Trend ChangePoints¶

그래프의 트렌드가 변하는 지점을 자동으로 감지 및 직접 제어

Automatic changepoint detection in Prophet¶

변경점 자동 감지¶

데이터의 해당 %의 크기만큼에서 잠재적 변경점 지정, 기본 80%¶

In [ ]:

from fbprophet.plot import add_changepoints_to_plot

m = Prophet(changepoint_range=0.9)
m.fit(df[-3650:])
future = m.make_future_dataframe(periods=30)
fcst = m.predict(future)

INFO:fbprophet:Disabling daily seasonality. Run prophet with daily_seasonality=True to override this.

In [ ]:

fig = m.plot(fcst)
a = add_changepoints_to_plot(fig.gca(), m, fcst)

Adjusting trend flexibility¶

트렌드가 Overfit, Underfit 인 경우 유연성을 조절 가능¶

높을 수록 트렌드가 유연해지며 Underfit 해결 가능¶

기본 0.05¶

In [ ]:

m = Prophet(changepoint_prior_scale=0.5)
fcst = m.fit(df[-3650:]).predict(future)

INFO:fbprophet:Disabling daily seasonality. Run prophet with daily_seasonality=True to override this.

In [ ]:

fig = m.plot(fcst)
a = add_changepoints_to_plot(fig.gca(), m, fcst)

Specifying the locations of the changepoints¶

변경점을 직접 추가

In [ ]:

m = Prophet(changepoints=['2018-05-01'])
fcst = m.fit(df[-3650:]).predict(future)

INFO:fbprophet:Disabling daily seasonality. Run prophet with daily_seasonality=True to override this.

In [ ]:

fig = m.plot(fcst)
a = add_changepoints_to_plot(fig.gca(), m, fcst)

Seasonality, Holiday Effects, And Regressors¶

Modeling Holidays and Special Events¶

공휴일이나 특별한 이벤트를 설정하여 예측에 도움을 줌¶

비나 태풍을 추가하면 괜찮을 듯 함¶

lower_window = 포함시킬 이전 일수¶

upper_window = 포함시킬 이후 일수¶

Built-in Country Holidays¶

내장된 나라별 공휴일 설정

In [ ]:

m = Prophet()
m.add_country_holidays(country_name='KR')
m.fit(df[-3650:])

INFO:fbprophet:Disabling daily seasonality. Run prophet with daily_seasonality=True to override this.

Out[ ]:

<fbprophet.forecaster.Prophet at 0x7f26790a4908>

In [ ]:

m.train_holiday_names

Out[ ]:

0                                       New Year's Day
1            The day preceding of Lunar New Year's Day
2                                 Lunar New Year's Day
3               The second day of Lunar New Year's Day
4          Alternative holiday of Lunar New Year's Day
5                            Independence Movement Day
6                               Birthday of the Buddha
7                                       Children's Day
8                                           Labour Day
9                                         Memorial Day
10                                      Liberation Day
11                        The day preceding of Chuseok
12                                             Chuseok
13                           The second day of Chuseok
14                             National Foundation Day
15                                         Hangeul Day
16                                       Christmas Day
17               Alternative holiday of new Year's Day
18               Alternative holiday of Children's Day
19                      Alternative holiday of Chuseok
20    Alternative holiday of Independence Movement Day
dtype: object

In [ ]:

future = m.make_future_dataframe(periods=30)
forecast = m.predict(future)
fig = m.plot_components(forecast)

Fourier Order for Seasonalities¶

푸리에 급수를 이용한 것.

연간 트렌드의 유연성 설정. 강도를 높이면 Overfit 발생

In [ ]:

from fbprophet.plot import plot_yearly
m = Prophet().fit(df[-3650:])
a = plot_yearly(m)

INFO:fbprophet:Disabling daily seasonality. Run prophet with daily_seasonality=True to override this.

In [ ]:

m = Prophet(yearly_seasonality=100).fit(df[-3650:])
a = plot_yearly(m)

INFO:fbprophet:Disabling daily seasonality. Run prophet with daily_seasonality=True to override this.

Specifying Custom Seasonalities¶

사용자 설정 트렌드 생성

기본적으로 weekly와 yearly 만 있으므로 monthly를 추가로 설정.

한달의 주기를 30.5일로 설정

In [ ]:

m = Prophet(weekly_seasonality=False)

m.add_seasonality(name='monthly', period=30.5, fourier_order=5)

forecast = m.fit(df[-3650:]).predict(future)

fig = m.plot_components(forecast)

INFO:fbprophet:Disabling daily seasonality. Run prophet with daily_seasonality=True to override this.

Seasonalities that depend on other factors¶

Seasonalities는 다른 요인에 의해 달라질 수 있음

숙박업 매출의 경우, 비수기와 성수기에 따라서 달라질 수 있으므로 이를 조건부 계절성을 사용하여 설정 가능

In [ ]:

# 8월부터 2월까지는 True, 그 외는 False로 설정

def is_nfl_season(ds):
    date = pd.to_datetime(ds)
    return (date.month > 8 or date.month < 2)

df_3 = df

df_3['on_season'] = df_3['ds'].apply(is_nfl_season)
df_3['off_season'] = ~df_3['ds'].apply(is_nfl_season)

df_3

Out[ ]:

	ds	y	cap	floor	on_season	off_season
0	1907-10-01	13.5	50	-30	True	False
1	1907-10-02	16.2	50	-30	True	False
2	1907-10-03	16.2	50	-30	True	False
3	1907-10-04	16.5	50	-30	True	False
4	1907-10-05	17.6	50	-30	True	False
...	...	...	...	...	...	...
42807	2020-09-13	21.5	50	-30	True	False
42808	2020-09-14	21.3	50	-30	True	False
42809	2020-09-15	23.0	50	-30	True	False
42810	2020-09-16	22.5	50	-30	True	False
42811	2020-09-17	21.3	50	-30	True	False

42812 rows × 6 columns

In [ ]:

m = Prophet(weekly_seasonality=False)
m.add_seasonality(name='weekly_on_season', period=7, fourier_order=3, condition_name='on_season')
m.add_seasonality(name='weekly_off_season', period=7, fourier_order=3, condition_name='off_season')

future['on_season'] = future['ds'].apply(is_nfl_season)
future['off_season'] = ~future['ds'].apply(is_nfl_season)
forecast = m.fit(df_3[-3650:]).predict(future)
fig = m.plot_components(forecast)

INFO:fbprophet:Disabling daily seasonality. Run prophet with daily_seasonality=True to override this.

Prior scale for holidays and seasonality¶

Seasonality 또는 holidays가 overfit, underfit 된 경우 holidays_prior_scale을 조절하여 해결 가능.

기본값 = 10

아래는 데이터가 없어 작동하지 않는 예

In [ ]:

m = Prophet(holidays=holidays, holidays_prior_scale=0.05).fit(df)
forecast = m.predict(future)
forecast[(forecast['playoff'] + forecast['superbowl']).abs() > 0][
    ['ds', 'playoff', 'superbowl']][-10:]

Seasnonality는 prior_scale을 조절하여 유연성 설정 가능

In [ ]:

m = Prophet(weekly_seasonality=False)
m.add_seasonality(name='weekly_on_season', period=7, fourier_order=3, condition_name='on_season', prior_scale=0.1)
m.add_seasonality(name='weekly_off_season', period=7, fourier_order=3, condition_name='off_season', prior_scale=0.1)

future['on_season'] = future['ds'].apply(is_nfl_season)
future['off_season'] = ~future['ds'].apply(is_nfl_season)
forecast = m.fit(df_3[-3650:]).predict(future)
fig = m.plot_components(forecast)

INFO:fbprophet:Disabling daily seasonality. Run prophet with daily_seasonality=True to override this.

Additional regressors¶

특정 요일에 추가 효과를 줄 수 있음

아래는 NFL 시즌의 일요일에 추가 효과를 주는 것으로 작동만 되며 의미는 없음

In [ ]:

def nfl_sunday(ds):
    date = pd.to_datetime(ds)
    if date.weekday() == 6 and (date.month > 8 or date.month < 2):
        return 1
    else:
        return 0
df['nfl_sunday'] = df['ds'].apply(nfl_sunday)

m = Prophet()
m.add_regressor('nfl_sunday')
m.fit(df[-3650:])

future['nfl_sunday'] = future['ds'].apply(nfl_sunday)

forecast = m.predict(future)
fig = m.plot_components(forecast)

INFO:fbprophet:Disabling daily seasonality. Run prophet with daily_seasonality=True to override this.

Multiplicative Seasonality¶

증가하거나 점점 범위가 넓어지는 계절성을 가진 모델을 예측할 때 유용

계절별 온도차가 심해지는 경우에 유용

In [ ]:

m = Prophet(seasonality_mode='multiplicative')
m.fit(df[-3650:])
forecast = m.predict(future)
fig = m.plot(forecast)

INFO:fbprophet:Disabling daily seasonality. Run prophet with daily_seasonality=True to override this.

Uncertainty Intervals¶

불확실성의 범위

Uncertainty in the trend¶

예측 값의 불확실성 범위

기본 80%

그래프가 2번 뜨는 이유는 모르겠음

In [ ]:

m = Prophet(interval_width=0.95)
m.fit(df[-3650:])
future = m.make_future_dataframe(periods=30)
fcst = m.predict(future)
m.plot(fcst)

INFO:fbprophet:Disabling daily seasonality. Run prophet with daily_seasonality=True to override this.

Out[ ]:

In [ ]:

fig = m.plot_components(fcst)

Uncertainty in seasonality¶

계절성의 불확실성

베이지안 샘플링 이용

mcmc_samples

기본값 0

하단의 예는 최초 300일에 대해 적용하는 예시

In [ ]:

m = Prophet(mcmc_samples=300)

In [ ]:

m.fit(df[-3650:]) # 실행시간 오래걸림

INFO:fbprophet:Disabling daily seasonality. Run prophet with daily_seasonality=True to override this.
WARNING:pystan:Maximum (flat) parameter count (1000) exceeded: skipping diagnostic tests for n_eff and Rhat.
To run all diagnostics call pystan.check_hmc_diagnostics(fit)

Out[ ]:

<fbprophet.forecaster.Prophet at 0x7f2677e934e0>

In [ ]:

forecast = m.predict(future) 

In [ ]:

fig = m.plot_components(forecast)

In [ ]:

m.predictive_samples(future)

Outliers¶

이상치가 예측에 영향을 미치므로 제거하는 방법

1. 특정 기간에 이상치가 집중된 경우

   특정 구간을 None으로 설정

In [ ]:

df.loc[(df['ds'] > '2010-01-01') & (df['ds'] < '2011-01-01'), 'y'] = None

model = Prophet().fit(df)

fig = model.plot(model.predict(future))

1. 특정 범위를 넘어가는 이상치가 여러 군데에 분포한 경우

   특정 값을 넘어가는 데이터만 None으로 설정

Non-Daily Data¶

Sub-daily data¶

5분 간격의 데이터도 예측에 이용 가능

Data with regular gaps¶

매일 5분 간격으로 00시부터 06시까지의 데이터만 가지고도 예측 가능

Monthly data¶

매월 1일에 대한 데이터만 있어도 예측 가능

Holidays with aggregated data¶

Diagnostics¶

교차검증기능이 있음

아래 그림은 최초 5년간의 데이터를 이용해 그 다음 1년을 예측하고, 실제 데이터를 이용해 검증하는 것

In [ ]:

from fbprophet.diagnostics import cross_validation
df = pd.read_csv("./data_1907_2020.csv")
m = Prophet()
m.fit(df[-3650:])

INFO:fbprophet:Disabling daily seasonality. Run prophet with daily_seasonality=True to override this.

Out[ ]:

<fbprophet.forecaster.Prophet at 0x7f2678b051d0>

In [ ]:

df_cv = cross_validation(m, initial='730 days', period='180 days', horizon = '365 days')
df_cv.head()

INFO:fbprophet:Making 15 forecasts with cutoffs between 2012-10-24 00:00:00 and 2019-09-18 00:00:00

Out[ ]:

	ds	yhat	yhat_lower	yhat_upper	y	cutoff
0	2012-10-25	14.446122	10.435999	18.222661	15.2	2012-10-24
1	2012-10-26	14.293630	10.373073	18.243430	15.9	2012-10-24
2	2012-10-27	14.319158	10.352468	18.037690	14.3	2012-10-24
3	2012-10-28	14.282673	10.426169	17.926619	14.3	2012-10-24
4	2012-10-29	13.935930	10.224008	17.663029	11.7	2012-10-24

cutoff 날짜를 임의로 입력 가능

In [ ]:

cutoffs = pd.to_datetime(['2013-02-15', '2013-08-15', '2014-02-15'])
df_cv2 = cross_validation(m, cutoffs=cutoffs, horizon='365 days')
df_cv2.head()

Out[ ]:

	ds	yhat	yhat_lower	yhat_upper	y	cutoff
0	2013-02-16	-0.857983	-4.753327	3.605245	-2.6	2013-02-15
1	2013-02-17	-0.468900	-4.470028	3.794405	2.3	2013-02-15
2	2013-02-18	-0.291445	-4.469484	3.542425	0.5	2013-02-15
3	2013-02-19	0.129143	-3.848923	4.356698	-2.9	2013-02-15
4	2013-02-20	0.203922	-3.795085	4.051412	-4.8	2013-02-15

RMSE: Root Mean Square Error
MSE: Mean Square Error
MAE: Mean Absolute Error
MDAPE: Median Absolute Prediction Error

mse, mae 등의 용어는 다음 링크 참고

https://partrita.github.io/posts/regression-error/

오차 및 오차율 확인 가능

In [ ]:

from fbprophet.diagnostics import performance_metrics
df_p = performance_metrics(df_cv)
df_p.head()

INFO:fbprophet:Skipping MAPE because y close to 0

Out[ ]:

	horizon	mse	rmse	mae	mdape	coverage
0	37 days	9.835698	3.136192	2.493288	0.133392	0.776234
1	38 days	9.903961	3.147056	2.509502	0.135919	0.771725
2	39 days	9.914303	3.148699	2.517180	0.135919	0.771603
3	40 days	9.996804	3.161772	2.523359	0.136254	0.772456
4	41 days	10.144623	3.185062	2.540350	0.136254	0.766971

In [ ]:

from fbprophet.plot import plot_cross_validation_metric
fig = plot_cross_validation_metric(df_cv, metric='mape')

Parallelizing cross validation¶

교차검증의 병렬화는 다음 파라미터를 조절함으로써 설정 가능

parallel¶

• None : 기본, 병렬처리 하지 않음
• 'processes' : Parallelize with concurrent.futures.ProcessPoolExectuor.
• 'threads' : Parallelize with concurrent.futures.ThreadPoolExecutor. Note that some operations currently hold Python's Global Interpreter Lock, so parallelizing with threads may be slower than training sequentially.
• 'dask': Dask로 병렬처리. dask.distributed Client 이 필요함
• object : Any instance with a .map method. This method will be called with single_cutoff_forecast and a sequence of iterables where each element is the tuple of arguments to pass to single_cutoff_forecast

In [ ]:

from fbprophet.diagnostics import cross_validation
df = pd.read_csv("./data_1907_2020.csv")
m = Prophet()
m.fit(df[-3650:])

INFO:fbprophet:Disabling daily seasonality. Run prophet with daily_seasonality=True to override this.

Out[ ]:

<fbprophet.forecaster.Prophet at 0x7f268e00cb00>

In [ ]:

from dask.distributed import Client

client = Client()  # dask 사용시 필요

df_cv = cross_validation(m, initial='730 days', period='180 days', horizon = '365 days', parallel="processes")

/usr/local/lib/python3.6/dist-packages/distributed/bokeh/core.py:57: UserWarning:


Port 8787 is already in use. 
Perhaps you already have a cluster running?
Hosting the diagnostics dashboard on a random port instead.

INFO:fbprophet:Making 15 forecasts with cutoffs between 2012-10-24 00:00:00 and 2019-09-18 00:00:00
INFO:fbprophet:Applying in parallel with <concurrent.futures.process.ProcessPoolExecutor object at 0x7f2678373a90>

Hyperparameter tuning¶

In [ ]:

import itertools
import numpy as np
import pandas as pd

param_grid = {  
    'changepoint_prior_scale': [0.001, 0.01, 0.1, 0.5],
    'seasonality_prior_scale': [0.01, 0.1, 1.0, 10.0],
}

# Generate all combinations of parameters
all_params = [dict(zip(param_grid.keys(), v)) for v in itertools.product(*param_grid.values())]
rmses = []  # Store the RMSEs for each params here

# Use cross validation to evaluate all parameters
for params in all_params:
    m = Prophet(**params).fit(df)  # Fit model with given params
    df_cv = cross_validation(m, cutoffs=cutoffs, horizon='30 days', parallel="processes")
    df_p = performance_metrics(df_cv, rolling_window=1)
    rmses.append(df_p['rmse'].values[0])

# Find the best parameters
tuning_results = pd.DataFrame(all_params)
tuning_results['rmse'] = rmses
print(tuning_results)

INFO:fbprophet:Disabling daily seasonality. Run prophet with daily_seasonality=True to override this.
INFO:fbprophet:Applying in parallel with <concurrent.futures.process.ProcessPoolExecutor object at 0x7f2677361e80>
INFO:fbprophet:Disabling daily seasonality. Run prophet with daily_seasonality=True to override this.
INFO:fbprophet:Applying in parallel with <concurrent.futures.process.ProcessPoolExecutor object at 0x7f267736eb38>
INFO:fbprophet:Disabling daily seasonality. Run prophet with daily_seasonality=True to override this.
INFO:fbprophet:Applying in parallel with <concurrent.futures.process.ProcessPoolExecutor object at 0x7f2671c11fd0>
INFO:fbprophet:Disabling daily seasonality. Run prophet with daily_seasonality=True to override this.
INFO:fbprophet:Applying in parallel with <concurrent.futures.process.ProcessPoolExecutor object at 0x7f2678603da0>
INFO:fbprophet:Disabling daily seasonality. Run prophet with daily_seasonality=True to override this.
INFO:fbprophet:Applying in parallel with <concurrent.futures.process.ProcessPoolExecutor object at 0x7f2678611ef0>
INFO:fbprophet:Disabling daily seasonality. Run prophet with daily_seasonality=True to override this.
INFO:fbprophet:Applying in parallel with <concurrent.futures.process.ProcessPoolExecutor object at 0x7f2671c11160>
INFO:fbprophet:Disabling daily seasonality. Run prophet with daily_seasonality=True to override this.
INFO:fbprophet:Applying in parallel with <concurrent.futures.process.ProcessPoolExecutor object at 0x7f2676687208>
INFO:fbprophet:Disabling daily seasonality. Run prophet with daily_seasonality=True to override this.
INFO:fbprophet:Applying in parallel with <concurrent.futures.process.ProcessPoolExecutor object at 0x7f267471f978>
INFO:fbprophet:Disabling daily seasonality. Run prophet with daily_seasonality=True to override this.
INFO:fbprophet:Applying in parallel with <concurrent.futures.process.ProcessPoolExecutor object at 0x7f2678c2a860>
INFO:fbprophet:Disabling daily seasonality. Run prophet with daily_seasonality=True to override this.
INFO:fbprophet:Applying in parallel with <concurrent.futures.process.ProcessPoolExecutor object at 0x7f26750bcda0>
INFO:fbprophet:Disabling daily seasonality. Run prophet with daily_seasonality=True to override this.
INFO:fbprophet:Applying in parallel with <concurrent.futures.process.ProcessPoolExecutor object at 0x7f2671c112b0>
INFO:fbprophet:Disabling daily seasonality. Run prophet with daily_seasonality=True to override this.
INFO:fbprophet:Applying in parallel with <concurrent.futures.process.ProcessPoolExecutor object at 0x7f26773fcef0>
INFO:fbprophet:Disabling daily seasonality. Run prophet with daily_seasonality=True to override this.
INFO:fbprophet:Applying in parallel with <concurrent.futures.process.ProcessPoolExecutor object at 0x7f26750bcda0>
INFO:fbprophet:Disabling daily seasonality. Run prophet with daily_seasonality=True to override this.
INFO:fbprophet:Applying in parallel with <concurrent.futures.process.ProcessPoolExecutor object at 0x7f26773b7278>
INFO:fbprophet:Disabling daily seasonality. Run prophet with daily_seasonality=True to override this.
INFO:fbprophet:Applying in parallel with <concurrent.futures.process.ProcessPoolExecutor object at 0x7f2676687710>
INFO:fbprophet:Disabling daily seasonality. Run prophet with daily_seasonality=True to override this.
INFO:fbprophet:Applying in parallel with <concurrent.futures.process.ProcessPoolExecutor object at 0x7f2676687710>

    changepoint_prior_scale  seasonality_prior_scale      rmse
0                     0.001                     0.01  2.753411
1                     0.001                     0.10  2.762134
2                     0.001                     1.00  2.763271
3                     0.001                    10.00  2.763607
4                     0.010                     0.01  2.688957
5                     0.010                     0.10  2.688815
6                     0.010                     1.00  2.695567
7                     0.010                    10.00  2.691588
8                     0.100                     0.01  2.752182
9                     0.100                     0.10  2.762239
10                    0.100                     1.00  2.761342
11                    0.100                    10.00  2.759737
12                    0.500                     0.01  2.770102
13                    0.500                     0.10  2.777792
14                    0.500                     1.00  2.774974
15                    0.500                    10.00  2.773694

In [ ]:

best_params = all_params[np.argmin(rmses)]
print(best_params)

Additional Topics¶

Saving models¶

나중에 불러와서 사용할 수 있도록 Prophet 모델을 저장할 수 있음

In [ ]:

import json
from fbprophet.serialize import model_to_json, model_from_json

with open('serialized_model.json', 'w') as fout:
    json.dump(model_to_json(m), fout)  # Save model

with open('serialized_model.json', 'r') as fin:
    m = model_from_json(json.load(fin))  # Load model

Flat trend and custom trends¶

트렌드 = 연간 기온 상승

트렌드보다 계절성의 성향이 강한경우 트렌드의 성장률을 균일하게 유지할 수 있음

growth 파라미터를 조절하여 설정가능

In [ ]:

m = Prophet(growth='flat')

Updating fitted models¶

이미 fit 된 모델에 새 데이터를 업데이트 할 때 사용

In [ ]:

# Python
def stan_init(m):
    """Retrieve parameters from a trained model.
    
    Retrieve parameters from a trained model in the format
    used to initialize a new Stan model.
    
    Parameters
    ----------
    m: A trained model of the Prophet class.
    
    Returns
    -------
    A Dictionary containing retrieved parameters of m.
    
    """
    res = {}
    for pname in ['k', 'm', 'sigma_obs']:
        res[pname] = m.params[pname][0][0]
    for pname in ['delta', 'beta']:
        res[pname] = m.params[pname][0]
    return res

df = pd.read_csv('../examples/example_wp_log_peyton_manning.csv')
df1 = df.loc[df['ds'] < '2016-01-19', :]  # All data except the last day
m1 = Prophet().fit(df1) # A model fit to all data except the last day


%timeit m2 = Prophet().fit(df)  # Adding the last day, fitting from scratch
%timeit m2 = Prophet().fit(df, init=stan_init(m1))  # Adding the last day, warm-starting from m1

원리¶

Model Components¶

y(t) = g(t) + s(t) + h(t) + error¶

g: Growth
s: Seasonality
h: Holidays, Events

Growth¶

Linear Growth
Non-Linear Growth(Logistic Growth)

Seasonality¶

주기적으로 나타나는 패턴은 Fourier Series를 이용해서 근사치를 찾음

Holidays¶

주기성을 가지지는 않지만 전체 추세에 영향을 주는 이벤트로, 앞뒤로 window범위를 지정해서 해당 이벤트가 미치는 영향의 범위를 설정 할 수 있음

Model Fitting¶

Stan을 통해 모델을 학습시킴

MAP(Maximum A Posteriori), MCMC(MarkovChain Monte Carlo) 두 방식이 있음

MAP(Maximum A Posteriori)

 - 속도가 훨씬 빠름  

MCMC(MarkovChain Monte Carlo)

 - 모형의 변동성을 더 자세히 볼 수 있음

함수 정리¶

Prophet¶

함수 파라미터

맨 앞에 오는 것이 default

1. growth: 'linear', 'logistic'

1. changepoints: None

   List of dates at which to include potential changepoints. If not specified, potential changepoints are selected automatically.

2. n_changepoints: 25

   Number of potential changepoints to include. Not used if input changepoints is supplied. If changepoints is not supplied, then n_changepoints potential changepoints are selected uniformly from the first changepoint_range proportion of the history.

3. changepoint_range: 0.8

   changepoint의 자동감지 범위
   Proportion of history in which trend changepoints will be estimated. Defaults to 0.8 for the first 80%. Not used if changepoints is specified.

4. yearly_seasonality: 'auto', True, False, or a number of Fourier terms to generate.

   Fit yearly seasonality.

5. weekly_seasonality: 'auto', True, False, or a number of Fourier terms to generate.

   Fit weekly seasonality.

6. daily_seasonality: 'auto', True, False, or a number of Fourier terms to generate.

   Fit daily seasonality.

7. holidays: None

   공휴일 및 이벤트를 dataframe 형식으로 입력
   pd.DataFrame with columns holiday (string) and ds (date type) and optionally columns lower_window and upper_window which specify a range of days around the date to be included as holidays. lower_window=-2 will include 2 days prior to the date as holidays. Also optionally can have a column prior_scale specifying the prior scale for that holiday.

8. seasonality_mode: 'additive', 'multiplicative'.

   additive는 추세만 점점 증가하는 경우
   multiplicative는 추세뿐만 아니라 값의 범위도 증가하는 경우

9. seasonality_prior_scale: 10.0

   Parameter modulating the strength of the seasonality model. Larger values allow the model to fit larger seasonal fluctuations, smaller values dampen the seasonality. Can be specified for individual seasonalities using add_seasonality.

10. holidays_prior_scale: 10.0

    holidays 가 모델에 미치는 영향의 정도를 설정
    Parameter modulating the strength of the holiday components model, unless overridden in the holidays input.

11. changepoint_prior_scale: 0.05

    changepoint의 유연성 조절. 낮으면 underfit, 높으면 overfit 발생
    Parameter modulating the flexibility of the automatic changepoint selection. Large values will allow many changepoints, small values will allow few changepoints.

12. mcmc_samples: 0

    Integer, if greater than 0, will do full Bayesian inference with the specified number of MCMC samples. If 0, will do MAP estimation.

13. interval_width: 0.8

    예측 값의 불확실성 범위 설정
    Float, width of the uncertainty intervals provided for the forecast. If mcmc_samples=0, this will be only the uncertaintyin the trend using the MAP estimate of the extrapolated generative model. If mcmc.samples>0, this will be integrated over all model parameters, which will include uncertainty in seasonality.

14. uncertainty_samples: 1000

    Number of simulated draws used to estimate uncertainty intervals. Settings this value to 0 or False will disable uncertainty estimation and speed up the calculation.

15. stan_backend: None

    str as defined in StanBackendEnum default: None - will try to iterate over all available backends and find the working one

Prophet.fit¶

함수 파라미터

1. df: 필수입력

   pd.DataFrame containing the history. Must have columns ds (date type) and y, the time series. If self.growth is 'logistic', then df must also have a column cap that specifies the capacity at each ds.

2. kwargs:

   Additional arguments passed to the optimizing or sampling functions in Stan.

Prophet.predict¶

함수 파라미터

1. df: None

   make_future_dataframe 의 return 값을 파라미터로 받음
   pd.DataFrame with dates for predictions (column ds), and capacity (column cap) if logistic growth. If not provided, predictions are made on the history.

Prophet.plot¶

함수 파라미터

1. m: 필수입력

   Prophet model.

2. fcst: 필수입력

   pd.DataFrame output of m.predict.

3. ax: None

   Optional matplotlib axes on which to plot.

4. uncertainty: True

   Optional boolean to plot uncertainty intervals, which will only be done if m.uncertainty_samples > 0.

5. plot_cap: True

   Optional boolean indicating if the capacity should be shown in the figure, if available.

6. xlabel: 'ds'

   x 축으로 사용할 데이터의 이름

7. ylabel: 'y'

   y 축으로 사용할 데이터의 이름

8. figsize: (10, 6)

   Optional tuple width, height in inches.

Prophet.plot_components¶

함수 파라미터

1. fcst: 필수입력

   pd.DataFrame output of self.predict.

2. uncertainty: True

   Optional boolean to plot uncertainty intervals.

3. plot_cap: True

   Optional boolean indicating if the capacity should be shown in the figure, if available.

4. weekly_start: 0

   Optional int specifying the start day of the weekly seasonality plot. 0 (default) starts the week on Sunday. 1 shifts by 1 day to Monday, and so on.

5. yearly_start: 0

   Optional int specifying the start day of the yearly seasonality plot. 0 (default) starts the year on Jan 1. 1 shifts by 1 day to Jan 2, and so on.

6. figsize: None

   Optional tuple width, height in inches.

Prophet.make_future_dataframe¶

함수 파라미터

1. periods: 필수입력

   앞으로 예측할 기간

2. freq: 'D'

   Any valid frequency for pd.date_range, such as 'D' or 'M'.

3. include_history: True

   Boolean to include the historical dates in the dataframe for predictions.

참고자료¶

https://facebook.github.io/prophet/docs/outliers.html

https://zamezzz.tistory.com/category/IT/Ai

https://zzsza.github.io/data/2019/02/06/prophet/

https://gorakgarak.tistory.com/1255

https://dacon.io/codeshare/1527?dtype=recent

프레젠테이션 형식의 간단한 설명: https://www.slideshare.net/lumiamitie/facebook-prophet

https://be-favorite.tistory.com/64