如何在python中计算时间序列数据中组的SMAPE？

Question

我的数据如下所示，我正在使用 facebook FbProphet 进行预测。接下来我想为我的数据框中的每个组计算SMAPE。我找到了 Kaggle 用户 here 描述的功能，但我不确定如何在我当前的代码中实现。这样SMAPE 就可以为每个组计算。另外，我知道 fbProphet 有验证功能，但我想为每个组计算SMAPE。

注意：我是python新手，请提供代码解释。

数据集

import pandas as pd
data = {'Date':['2017-01-01', '2017-01-01','2017-01-01','2017-01-01','2017-01-01','2017-01-01','2017-01-01','2017-01-01',
               '2017-02-01', '2017-02-01','2017-02-01','2017-02-01','2017-02-01','2017-02-01','2017-02-01','2017-02-01'],'Group':['A','A','B','B','C','C','D','D','A','A','B','B','C','C','D','D'],
       'Amount':['12.1','13.2','15.1','10.7','12.9','9.0','5.6','6.7','4.3','2.3','4.0','5.6','7.8','2.3','5.6','8.9']}
df = pd.DataFrame(data)
print (df)

到目前为止的代码...

def get_prediction(df):
    prediction = {}
    df = df.rename(columns={'Date': 'ds','Amount': 'y', 'Group': 'group'})
    df=df.groupby(['ds','group'])['y'].sum()
    df=pd.DataFrame(df).reset_index()
    list_articles = df.group.unique()

    for group in list_articles:
        article_df = df.loc[df['group'] == group]
        # set the uncertainty interval to 95% (the Prophet default is 80%)
        my_model = Prophet(weekly_seasonality= True, daily_seasonality=True,seasonality_prior_scale=1.0)
        my_model.fit(article_df)
        future_dates = my_model.make_future_dataframe(periods=6, freq='MS')
        forecast = my_model.predict(future_dates)
        prediction[group] = forecast
        my_model.plot(forecast)
    return prediction

Answer 1

您仍然可以使用 fbprophet 自己的 cross_validation 函数，但使用您自己的评分。这是 uber 的一篇不错的博客，介绍了他们如何进行回测（滑动窗口和扩展窗口）：https://eng.uber.com/forecasting-introduction/

fbprophet 的 cv 函数在滑动窗口上运行。如果这没问题，您可以将其与自定义评分功能结合使用。我认为一个不错的方法是扩展 Prophet 并实现 .score() 方法。

这里是一个示例实现：

from fbprophet import Prophet
from fbprophet.diagnostics import cross_validation
import numpy as np

class ProphetEstimator(Prophet):
    """
        Wrapper with custom scoring
    """

    def __init__(self, *args, **kwargs):
        super(ProphetEstimator, self).__init__(*args, **kwargs)

    def score(self):
        # cross val score reusing prophets own cv implementation
        df_cv = cross_validation(self, horizon='6 days')
        # Here decide how you want to calculate SMAPE.
        # Here each sliding window is summed up, 
        # and the SMAPE is calculated over the sum of periods, for all windows.
        df_cv = df_cv.groupby('cutoff').agg({
            "yhat": "sum",
            'y': "sum"
        })
        smape = self.calc_smape(df_cv['yhat'], df_cv['y'])
        return smape

    def calc_smape(self, y_hat, y):
        return 100/len(y) * np.sum(2 * np.abs(y_hat - y) / (np.abs(y) + np.abs(y_hat)))


def get_prediction(df):
    prediction = {}
    df = df.rename(columns={'Date': 'ds','Amount': 'y', 'Group': 'group'})
    df=df.groupby(['ds','group'])['y'].sum()
    df=pd.DataFrame(df).reset_index()
    list_articles = df.group.unique()

    for group in list_articles:
        article_df = df.loc[df['group'] == group]
        # set the uncertainty interval to 95% (the Prophet default is 80%)
        my_model = ProphetEstimator(weekly_seasonality= True, daily_seasonality=True,seasonality_prior_scale=1.0)
        my_model.fit(article_df)
        smape = my_model.score() # store this somewhere
        future_dates = my_model.make_future_dataframe(periods=6, freq='MS')
        forecast = my_model.predict(future_dates)
        prediction[group] = (forecast, smape)
        my_model.plot(forecast)
    return prediction