时间序列预测（一）基于Prophet的销售额预测

小O：小H，有没有什么方法能快速的预测下未来的销售额啊

小H：Facebook曾经开源了一款时间序列预测算法fbprophet，简单又快速～

传统的时间序列算法很多，例如AR、MA、ARIMA等，对于非专业人员来说显得很难上手。而Prophet相对来说就友好多了，而且预测效果又很不错，所以用它来预测时间序列数据再适合不过了。本文主要参考基于facebook的时间序列预测框架prophet的实战应用。

Prophet的安装需要先安装pystan

conda install pystan # 终端上安装，需要执行procced选择y

pip install fbprophet

数据探索

import matplotlib.pyplot as plt
import pandas as pd
import numpy as np
import seaborn as sns
from fbprophet import Prophet
from sklearn.metrics import mean_squared_error
from math import sqrt
import datetime
from xgboost import XGBRegressor
from sklearn.metrics import explained_variance_score, mean_absolute_error, \
mean_squared_error, r2_score  # 批量导入指标算法

from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import LSTM
from tensorflow.keras.layers import Dense, Dropout
from sklearn.preprocessing import MinMaxScaler
from keras.wrappers.scikit_learn import KerasRegressor
from sklearn.model_selection import GridSearchCV

以下数据如果有需要的同学可关注公众号HsuHeinrich，回复【数据挖掘-时间序列01】自动获取～

# 读取数据
raw_data = pd.read_csv('train.csv')
raw_data.head()

# 转化为日期
raw_data['datetime'] = raw_data['datetime'].apply(pd.to_datetime)

# 查看历史销售趋势
plt.figure(figsize = (15,8))
sns.lineplot(x = 'datetime', y = 'count', data = raw_data, err_style=None)
plt.show()

output_10_0

特征工程

# 构造prophet需要的ds/y数据
df_model = raw_data[['datetime', 'count']].rename(columns = {'datetime': 'ds','count': 'y'})

模型拟合

# 模型拟合
model_fb = Prophet(interval_width = 0.95).fit(df_model)
# 构造预测日期
future_dates = model_fb.make_future_dataframe(periods = 100, freq='H')
# 预测结果
forecast = model_fb.predict(future_dates)

INFO:fbprophet:Disabling yearly seasonality. Run prophet with yearly_seasonality=True to override this.

# 预测最后几周的日期
forecast[['ds', 'yhat', 'yhat_lower', 'yhat_upper']].tail()

结果展示

# 观察预测效果
model_fb.plot(forecast);

output_16_0

# 观察趋势因素
model_fb.plot_components(forecast);

output_17_0

# 模型评估 MSE
metric_df = forecast.set_index('ds')[['yhat']].join(df_model.set_index('ds').y).reset_index()
metric_df.dropna(inplace=True)
error = mean_squared_error(metric_df.y, metric_df.yhat)
print('The MSE is {}'. format(error))

The MSE is 12492.842870220222

添加假期因素

# 定义假期因素
def is_school_holiday_season(ds):    
    date = pd.to_datetime(ds)
    starts = datetime.date(date.year, 7, 1)
    ends = datetime.date(date.year, 9, 9)
    return starts < date.to_pydatetime().date() < ends

df_model['school_holiday_season'] = df_model['ds'].apply(is_school_holiday_season)
df_model['not_school_holiday_season'] = ~df_model['ds'].apply(is_school_holiday_season)
model_fb = Prophet(interval_width=0.95)

# 添加假期因素
model_fb.add_seasonality(name='school_holiday_season', period=365, fourier_order=3, condition_name='school_holiday_season')
model_fb.add_seasonality(name='not_school_holiday_season', period=365, fourier_order=3, condition_name='not_school_holiday_season')
model_fb.fit(df_model)

INFO:fbprophet:Disabling yearly seasonality. Run prophet with yearly_seasonality=True to override this.

<fbprophet.forecaster.Prophet at 0x7ff4e48833d0>

# 构造日期
future_dates = model_fb.make_future_dataframe(periods=100, freq='H')
future_dates['school_holiday_season'] = future_dates['ds'].apply(is_school_holiday_season)
future_dates['not_school_holiday_season'] = ~future_dates['ds'].apply(is_school_holiday_season)

# 预测
forecast = model_fb.predict(future_dates)

plt.figure(figsize=(10, 5))
model_fb.plot(forecast);

<Figure size 720x360 with 0 Axes>

output_24_1

# 观察趋势因素
model_fb.plot_components(forecast);

output_25_0

# 模型评估 MSE
metric_df = forecast.set_index('ds')[['yhat']].join(df_model.set_index('ds').y).reset_index()
metric_df.dropna(inplace=True)
error = mean_squared_error(metric_df.y, metric_df.yhat)
print('The MSE is {}'. format(error))