如何通过先知交叉验证获得更好的预测结果？

Question

我有10年的每日需求数据，有一个积极的趋势。https://gofile.io/?c=PS3YCO

在一年的最后三个月里，总是会有1到2的需求冲击。以及15-16岁。当月(晋升)。

我试着预测了：

未来= m.make_future_dataframe(periods=365)

预测=m.predict(未来)

但结果并不如我所料。我能得到的最好的MSE是6681。但是当我尝试交叉验证时，结果与6690几乎相同。

• 有什么可以通过我的设置来改变以获得更好的结果吗？
• 会帮助获得一个更大的数据集，比方说50年吗？
• ：我是否正确地理解了它，要么使用“预测=m.predict(未来)”，要么使用交叉验证的选项？

。

另外:当我使用"from fbprophet.diagnostics import performance_metrics“来计算MSE时，它会给出测试数据的值，但会给出更长的时间。我怎么才能计算去年的MSE呢？

非常感谢你的帮助:)

import numpy as np
from fbprophet import Prophet
import matplotlib.pyplot as plt
from fbprophet.diagnostics import cross_validation

df = pd.read_excel('Dataset2.3_kurz.xls')   

promotions = pd.DataFrame({ 
 'holiday': 'winter_promotion',
    'ds': pd.to_datetime(['2009-10-1','2009-10-2','2009-10-15','2009-10-16',
                          '2009-11-1','2009-11-2','2009-11-15','2009-11-16',
                          '2009-12-1','2009-12-2','2009-12-15','2009-12-16',
                          '2010-10-1','2010-10-2','2010-10-15','2010-10-16',
                          '2010-11-1','2010-11-2','2010-11-15','2010-11-16',
                          '2010-12-1','2010-12-2','2010-12-15','2010-12-16',
                          '2011-10-1','2011-10-2','2011-10-15','2011-10-16',
                          '2011-11-1','2011-11-2','2011-11-15','2011-11-16',
                          '2011-12-1','2011-12-2','2011-12-15','2011-12-16',
                          '2012-10-1','2012-10-2','2012-10-15','2012-10-16',
                          '2012-11-1','2012-11-2','2012-11-15','2012-11-16',
                          '2012-12-1','2012-12-2','2012-12-15','2012-12-16',
                          '2013-10-1','2013-10-2','2013-10-15','2013-10-16',
                          '2013-11-1','2013-11-2','2013-11-15','2013-11-16',
                          '2013-12-1','2013-12-2','2013-12-15','2013-12-16',
                          '2014-10-1','2014-10-2','2014-10-15','2014-10-16',
                          '2014-11-1','2014-11-2','2014-11-15','2014-11-16',
                          '2014-12-1','2014-12-2','2014-12-15','2014-12-16',
                          '2015-10-1','2015-10-2','2015-10-15','2015-10-16',
                          '2015-11-1','2015-11-2','2015-11-15','2015-11-16',
                          '2015-12-1','2015-12-2','2015-12-15','2015-12-16',
                          '2016-10-1','2016-10-2','2016-10-15','2016-10-16',
                          '2016-11-1','2016-11-2','2016-11-15','2016-11-16',
                          '2016-12-1','2016-12-2','2016-12-15','2016-12-16',
                          '2017-10-1','2017-10-2','2017-10-15','2017-10-16',
                          '2017-11-1','2017-11-2','2017-11-15','2017-11-16',
                          '2017-12-1','2017-12-2','2017-12-15','2017-12-16',
                          '2018-10-1','2018-10-2','2018-10-15','2018-10-16',
                          '2018-11-1','2018-11-2','2018-11-15','2018-11-16',
                          '2018-12-1','2018-12-2','2018-12-15','2018-12-16',
                          '2019-10-1','2019-10-2','2019-10-15','2019-10-16',
                          '2019-11-1','2019-11-2','2019-11-15','2019-11-16',
                          '2019-12-1','2019-12-2','2019-12-15','2019-12-16']),
    'lower_window': 0, 
    'upper_window': 0, 
})

#model
m = Prophet( growth='linear',
           holidays=promotions,
           seasonality_mode='multiplicative',

           holidays_prior_scale=10,  
           seasonality_prior_scale=10,            
            yearly_seasonality=True,
           )

m.fit(df)  

df_cv = cross_validation(m, initial='732 days', period='365 days', horizon = '366 days')

from fbprophet.diagnostics import performance_metrics
df_p = performance_metrics(df_cv)
df_p = df_p[-365:]
df_p.tail()

Answer 1

尝试网格搜索并调优超参数。使用changepoint_prior_scale参数可能也有帮助。这只是个人的观点，但将每年的季节性设置为假，并在外部加上一个新的傅立叶顺序和优先级也会有帮助。关于cross_validation，请查看this link以获得更好的理解。