本文完整项目代码已托管至github，下载地址：https://github.com/gitren111/Sales-Forecasting-using-the-Prophet-Algorithm-Rossmann-Store-Sales-Dataset-
销量预测项目简介
本项目流程主要包括数据导入与清洗、特征分析、时间序列趋势分析、使用Prophet模型进行销量预测。通过增加特征变量、处理缺失值，分析商店类型、促销等因素对销量的影响，最终实现对店铺未来6周销量的预测并可视化展示
Rossmann Store Sales数据集下载和详细介绍地址：https://www.kaggle.com/datasets/pratyushakar/rossmann-store-sales
导入数据
train = pd.read_csv('train.csv',
parse_dates=True,low_memory=False)
store = pd.read_csv('store.csv',low_memory=False)
pd.set_option('display.max_columns', None)
pd.set_option('display.max_rows', None)
pd.set_option('display.width', 1000)
print(f'train文件的标签列{train.index}')
数据清洗和初步分析
train数据集处理
增加变量
train['Date'] = pd.to_datetime(train['Date'])
train['Year'] = train['Date'].dt.year
train['Month'] = train['Date'].dt.month
train['Day'] = train['Date'].dt.day
train['WeekOfYear'] = train['Date'].apply(lambda x:x.isocalendar()[1])
train['SalePerCustomer'] = train['Sales']/train['Customers']
print(train['SalePerCustomer'].describe())
可视化：经验累积分布函数ECDF
- 直观呈现出数据在各个取值区间上的累计概率情况，帮助我们更好地了解数据的整体分布形态
- ECDF数据分布情况分析：
- 偏态分布：sales和CUSTOMER 80%集中在1000以下
- 接近20%的sales customer为0
- 零销售额原因预判：商店关门、没有顾客、数据记录问题、商品因素等
sns.set(style='ticks')
c = '#386B7F'
plt.figure(figsize=(12,6))
plt.subplot(311)
cdf = ECDF(train['Sales'])
plt.plot(cdf.x,cdf.y,label='data_models',color=c)
plt.xlabel('Sales')
plt.ylabel('ECDF')
plt.subplot(312)
cdf = ECDF(train['Customers'])
plt.plot(cdf.x,cdf.y,label='data_models',color=c)
plt.xlabel('Customers')
plt.ylabel('ECDF')
plt.subplot(313)
cdf = ECDF(train['SalePerCustomer'])
plt.plot(cdf.x,cdf.y,label='data_models',color=c)
plt.xlabel('SalePerCustomer')
plt.ylabel('ECDF')
plt.tight_layout()
plt.show()
零销售额原因分析和缺失值处理
- 零销售额原因分析：日人均零销售额数量为172871，其中商店关门占172817，商店开门但是0销售额占54
- 处理思路：将关店且0销售额的数据剔除，不参加销量预测
#3.1 商店关门:导致销售额为0
close_stores = train[(train['Open'] == 0) & (train['Sales'] == 0)]
print(close_stores.head())
print(f'关店的数量\n{close_stores.shape}')
#3.2 商店开门但是0销售额
open_zero_sales = train[(train['Open'] != 0) & (train['Sales'] == 0)]
print(open_zero_sales.head(5))
print(f'开店但零销售额数量\n{open_zero_sales.shape}')
#3.3 0人均销售额
zero_SalePerCustomer = train[(train['SalePerCustomer'] == 0) | (train['SalePerCustomer'].isna())]
print(zero_SalePerCustomer.head(5))
print(f'人均0销售额数量\n{zero_SalePerCustomer.shape}')
#3.4 剔除关店和0销售额情况，组成新的train数据
print('关店且0销售额的天数应该剔除，不参加预测')
train = train[(train['Open'] != 0) & (train['Sales'] != 0)]
print(f'剔除后新的train大小：\n{train.shape}')
关店且0销售额的天数应该剔除，不参加预测
剔除后新的train大小：
(844338, 14)
store数据集处理
print(f'store数据集预览\n{store.head()}')
store数据集预览
Store StoreType Assortment CompetitionDistance CompetitionOpenSinceMonth CompetitionOpenSinceYear Promo2 Promo2SinceWeek Promo2SinceYear PromoInterval
0 1 c a 1270.0 9.0 2008.0 0 NaN NaN NaN
1 2 a a 570.0 11.0 2007.0 1 13.0 2010.0 Jan,Apr,Jul,Oct
2 3 a a 14130.0 12.0 2006.0 1 14.0 2011.0 Jan,Apr,Jul,Oct
3 4 c c 620.0 9.0 2009.0 0 NaN NaN NaN
4 5 a a 29910.0 4.0 2015.0 0 NaN NaN NaN
查看数据缺失值情况
- CompetitionDistance、CompetitionOpenSinceMonth/Year和Promo2SinceWeek/Year、PromoInterval 存在空值
null = store.isnull().sum()
print(f'store空值预览：\n{null}')
tore空值预览：
Store 0
StoreType 0
Assortment 0
CompetitionDistance 3
CompetitionOpenSinceMonth 354
CompetitionOpenSinceYear 354
Promo2 0
Promo2SinceWeek 544
Promo2SinceYear 544
PromoInterval 544
CompetitionDistance缺失值处理
- CompetitionDistance缺失值分析：
- CompetitionDistance的空值数据打印可知，这部分数据是缺失了导致的空值
- 通过ECDF可知CompetitionDistance是偏态分布，所以用中位数填充空值
null_CompetitionDistance = store[pd.isnull(store['CompetitionDistance'])]
print(f'null_CompetitionDistance\n{null_CompetitionDistance}')
null_CompetitionDistance
Store StoreType Assortment CompetitionDistance CompetitionOpenSinceMonth CompetitionOpenSinceYear Promo2 Promo2SinceWeek Promo2SinceYear PromoInterval
290 291 d a NaN NaN NaN 0 NaN NaN NaN
621 622 a c NaN NaN NaN 0 NaN NaN NaN
878 879 d a NaN NaN NaN 1 5.0 2013.0 Feb,May,Aug,Nov
#3.5.2.1 ECDF查看分布情况，决定填充方式
sns.set(style='ticks')
c = '#386B7F'
plt.figure(figsize=(12,6))
cdf = ECDF(store['CompetitionDistance'])
plt.plot(cdf.x,cdf.y,label='store_ECDF',color=c)
plt.xlabel('CompetitionDistance')
plt.ylabel('ECDF')
plt.show()
#3.5.2.2 中位数填充
store['CompetitionDistance'].fillna(store['CompetitionDistance'].median(),inplace=True)
CompetitionOpenSinceMonth\Year缺失值处理
- CompetitionOpenSinceMonth、Year缺失值分析：
- CompetitionOpenSinceMonth、Year的空值数据打印可知，这部分数据是缺失了导致的空值，同时CompetitionOpenSinceMonth、Year的空值354个，
这354个数据的CompetitionDistance均不为空值，所以不能全部按0来填充 - 通过ECDF可知CompetitionOpenSinceMonth、Year是偏态分布，所以用中位数填充空值
- CompetitionOpenSinceMonth、Year的空值数据打印可知，这部分数据是缺失了导致的空值，同时CompetitionOpenSinceMonth、Year的空值354个，
#3.5.3 CompetitionOpenSinceMonth、Year缺失值处理
null_CompetitionOpenSinceMonth = store[pd.isnull(store['CompetitionOpenSinceMonth'])]
print(f'null_CompetitionOpenSinceMonth\n{null_CompetitionOpenSinceMonth.head(10)}')
null_CompetitionOpenSinceYear = store[pd.isnull(store['CompetitionOpenSinceYear'])]
print(f'null_CompetitionOpenSinceYear\n{null_CompetitionOpenSinceYear.head(10)}')
null_dis_null_CompetitionOpenSinceMonth = null_CompetitionOpenSinceMonth[null_CompetitionOpenSinceMonth['CompetitionDistance'] != 0 ]
print(f'null_dis_null_CompetitionOpenSinceMonth\n{null_dis_null_CompetitionOpenSinceMonth.shape}')
null_dis_null_CompetitionOpenSinceYear = null_CompetitionOpenSinceYear[null_CompetitionOpenSinceYear['CompetitionDistance'] != 0 ]
print(f'null_dis_null_CompetitionOpenSinceYear\n{null_dis_null_CompetitionOpenSinceYear.shape}')
null_CompetitionOpenSinceMonth
Store StoreType Assortment CompetitionDistance CompetitionOpenSinceMonth CompetitionOpenSinceYear Promo2 Promo2SinceWeek Promo2SinceYear PromoInterval
11 12 a c 1070.0 NaN NaN 1 13.0 2010.0 Jan,Apr,Jul,Oct
12 13 d a 310.0 NaN NaN 1 45.0 2009.0 Feb,May,Aug,Nov
15 16 a c 3270.0 NaN NaN 0 NaN NaN NaN
18 19 a c 3240.0 NaN NaN 1 22.0 2011.0 Mar,Jun,Sept,Dec
21 22 a a 1040.0 NaN NaN 1 22.0 2012.0 Jan,Apr,Jul,Oct
25 26 d a 2300.0 NaN NaN 0 NaN NaN NaN
28 29 d c 2170.0 NaN NaN 0 NaN NaN NaN
31 32 a a 2910.0 NaN NaN 1 45.0 2009.0 Feb,May,Aug,Nov
39 40 a a 180.0 NaN NaN 1 45.0 2009.0 Feb,May,Aug,Nov
40 41 d c 1180.0 NaN NaN 1 31.0 2013.0 Jan,Apr,Jul,Oct
null_CompetitionOpenSinceYear
Store StoreType Assortment CompetitionDistance CompetitionOpenSinceMonth CompetitionOpenSinceYear Promo2 Promo2SinceWeek Promo2SinceYear PromoInterval
11 12 a c 1070.0 NaN NaN 1 13.0 2010.0 Jan,Apr,Jul,Oct
12 13 d a 310.0 NaN NaN 1 45.0 2009.0 Feb,May,Aug,Nov
15 16 a c 3270.0 NaN NaN 0 NaN NaN NaN
18 19 a c 3240.0 NaN NaN 1 22.0 2011.0 Mar,Jun,Sept,Dec
21 22 a a 1040.0 NaN NaN 1 22.0 2012.0 Jan,Apr,Jul,Oct
25 26 d a 2300.0 NaN NaN 0 NaN NaN NaN
28 29 d c 2170.0 NaN NaN 0 NaN NaN NaN
31 32 a a 2910.0 NaN NaN 1 45.0 2009.0 Feb,May,Aug,Nov
39 40 a a 180.0 NaN NaN 1 45.0 2009.0 Feb,May,Aug,Nov
40 41 d c 1180.0 NaN NaN 1 31.0 2013.0 Jan,Apr,Jul,Oct
null_dis_null_CompetitionOpenSinceMonth
(354, 10)
null_dis_null_CompetitionOpenSinceYear
(354, 10)
#3.5.3.1 ECDF查看分布情况，决定填充方式
sns.set(style='ticks')
c = '#386B7F'
plt.figure(figsize=(12,6))
plt.subplot(211)
cdf = ECDF(store['CompetitionOpenSinceMonth'])
plt.plot(cdf.x,cdf.y,label='store_ECDF',color=c)
plt.xlabel('CompetitionOpenSinceMonth')
plt.ylabel('ECDF')
plt.subplot(212)
cdf = ECDF(store['CompetitionOpenSinceYear'])
plt.plot(cdf.x,cdf.y,label='store_ECDF',color=c)
plt.xlabel('CompetitionOpenSinceYear')
plt.ylabel('ECDF')
plt.tight_layout()
plt.show()
#3.5.3.2 中位数填充
store['CompetitionOpenSinceMonth'].fillna(store['CompetitionOpenSinceMonth'].median(),inplace=True)
store['CompetitionOpenSinceYear'].fillna(store['CompetitionOpenSinceYear'].median(),inplace=True)
Promo2SinceWeek/year、PromoInterval 缺失值处理
- Promo2SinceWeek/year、PromoInterval缺失值分析：
- Promo2SinceWeek/year、PromoInterval的空值数据打印可知，这部分数据是缺失了导致的空值，同时Promo2SinceWeek/year、PromoInterval的空值均为544
这544个数据为空值的同时，Promo2也均为0也就是没做持续促销，由于Promo2与三个特征强相关，所以全部按0来填充
- Promo2SinceWeek/year、PromoInterval的空值数据打印可知，这部分数据是缺失了导致的空值，同时Promo2SinceWeek/year、PromoInterval的空值均为544
#3.5.4 Promo2SinceWeek/year、PromoInterval 缺失值处理
null_Promo2SinceWeek = store[pd.isnull(store['Promo2SinceWeek'])]
print(f'null_Promo2SinceWeek\n{null_Promo2SinceWeek.head()}')
null_Promo2SinceYear = store[pd.isnull(store['Promo2SinceYear'])]
print(f'null_Promo2SinceWeek\n{null_Promo2SinceYear.head()}')
null_PromoInterval = store[pd.isnull(store['PromoInterval'])]
print(f'null_Promo2SinceWeek\n{null_PromoInterval.head()}')
#3.5.4.1 检查空值情况下，Promo2不为0的情况
Promo2_is_1_null_Promo2SinceWeek =null_Promo2SinceWeek[null_Promo2SinceWeek['Promo2'] != 0]
Promo2_is_1_null_Promo2SinceYear =null_Promo2SinceYear[null_Promo2SinceYear['Promo2'] != 0]
Promo2_is_1_null_PromoInterval =null_PromoInterval[null_PromoInterval['Promo2'] != 0]
print(f'Promo2_is_1_null_Promo2SinceWeek\n{Promo2_is_1_null_Promo2SinceWeek.shape}')
print(f'Promo2_is_1_null_Promo2SinceYear\n{Promo2_is_1_null_Promo2SinceYear.shape}')
print(f'Promo2_is_1_null_PromoInterval\n{Promo2_is_1_null_PromoInterval.shape}')
#3.5.4.2 0填充
store['Promo2SinceWeek'].fillna(0,inplace=True)
store['Promo2SinceYear'].fillna(0,inplace=True)
store['PromoInterval'].fillna(0,inplace=True)
#查看处理完后的store数据集
null_1 = store.isnull().sum()
print(f'store空值预览：\n{null_1}')
null_Promo2SinceWeek
Store StoreType Assortment CompetitionDistance CompetitionOpenSinceMonth CompetitionOpenSinceYear Promo2 Promo2SinceWeek Promo2SinceYear PromoInterval
0 1 c a 1270.0 9.0 2008.0 0 NaN NaN NaN
3 4 c c 620.0 9.0 2009.0 0 NaN NaN NaN
4 5 a a 29910.0 4.0 2015.0 0 NaN NaN NaN
5 6 a a 310.0 12.0 2013.0 0 NaN NaN NaN
6 7 a c 24000.0 4.0 2013.0 0 NaN NaN NaN
null_Promo2SinceWeek
Store StoreType Assortment CompetitionDistance CompetitionOpenSinceMonth CompetitionOpenSinceYear Promo2 Promo2SinceWeek Promo2SinceYear PromoInterval
0 1 c a 1270.0 9.0 2008.0 0 NaN NaN NaN
3 4 c c 620.0 9.0 2009.0 0 NaN NaN NaN
4 5 a a 29910.0 4.0 2015.0 0 NaN NaN NaN
5 6 a a 310.0 12.0 2013.0 0 NaN NaN NaN
6 7 a c 24000.0 4.0 2013.0 0 NaN NaN NaN
null_Promo2SinceWeek
Store StoreType Assortment CompetitionDistance CompetitionOpenSinceMonth CompetitionOpenSinceYear Promo2 Promo2SinceWeek Promo2SinceYear PromoInterval
0 1 c a 1270.0 9.0 2008.0 0 NaN NaN NaN
3 4 c c 620.0 9.0 2009.0 0 NaN NaN NaN
4 5 a a 29910.0 4.0 2015.0 0 NaN NaN NaN
5 6 a a 310.0 12.0 2013.0 0 NaN NaN NaN
6 7 a c 24000.0 4.0 2013.0 0 NaN NaN NaN
Promo2_is_1_null_Promo2SinceWeek
(0, 10)
Promo2_is_1_null_Promo2SinceYear
(0, 10)
Promo2_is_1_null_PromoInterval
(0, 10)
train、store数据合并
合并后train_store.shape是(844338, 22)代表两个表数据全部都匹配上了
print('以store列为索引合并train和store数据集')
train_store = pd.merge(train,store,how='inner',on='Store')
print(f'train_store尺寸：{train_store.shape}')
print(train_store.head())
特征分析
Store types:透视分析商店类型与销售额关系
- 将StoreType按Sales透视发现，b类型商店的平均销售额最高；但是对Sales和Customers进行求和透视发现，a和d类型商店分别在总销售
和顾客数排前两名，商店b在销售和顾客规模排名最后 - 判断原因：商店b的数量远低于其他店铺（只是a店铺规模的3.2%)
StoreType_Sales = train_store.groupby('StoreType')['Sales'].describe()
print(f'StoreType_Sales\n{StoreType_Sales}')
StoreType_Sales_cust_sum = train_store.groupby('StoreType')[['Sales','Customers']].sum()
print(f'StoreType_Sales_cust_sum\n{StoreType_Sales_cust_sum}')
count mean std min 25% 50% 75% max
StoreType
a 457042.0 6925.697986 3277.351589 46.0 4695.25 6285.0 8406.00 41551.0
b 15560.0 10233.380141 5155.729868 1252.0 6345.75 9130.0 13184.25 38722.0
c 112968.0 6933.126425 2896.958579 133.0 4916.00 6408.0 8349.25 31448.0
d 258768.0 6822.300064 2556.401455 538.0 5050.00 6395.0 8123.25 38037.0
StoreType_Sales_cust_sum
Sales Customers
StoreType
a 3165334859 363541431
b 159231395 31465616
c 783221426 92129705
d 1765392943 156904995
trends:画图趋势分析
- 商店类型和促销对sales趋势影响：
- 商店类型和促销不会影响整体的销售额趋势，但是有促销能提升销售额规模
- 同时12月份圣诞节销售额在不同店铺都会很高，后面要专门进行时间序列分析季节性和趋势
- 星期几和商店类型对sales趋势影响：
- 周一不同店铺的销售额会高于其他日子
- c类型店铺周日不开业，d类店铺11月的周日不开业
- 商店类型和促销对Customers趋势影响：商店类型和促销不会影响整体的客流趋势，但是有促销能提升客流规模
- 商店类型和促销对SalePerCustomer趋势影响：同样不会客单销售额有趋势影响，但是促销能提升客单销售额规模，其中店铺d的客单销售额最高，无促销10元，有促销12元
#4.2.1 sales trends:商店类型和促销对sales趋势影响
sns.catplot(data=train_store,
x='Month',
y='Sales',
col='StoreType',
row='Promo',
palette='plasma',
hue='StoreType',
kind='point')
#4.2.2 sales trends:星期几和商店类型对sales趋势影响
sns.catplot(data=train_store,
x='Month',
y='Sales',
col='DayOfWeek',
row='StoreType',
palette='plasma',
hue='StoreType',
kind='point')
#4.2.3 Customers trends:商店类型和促销对Customers趋势影响
sns.catplot(data=train_store,
x='Month',
y='Customers',
col='StoreType',
row='Promo',
palette='plasma',
hue='StoreType',
kind='point')
#4.2.4 sale per customer trends:商店类型和促销对SalePerCustomer趋势影响
sns.catplot(data=train_store,
x='Month',
y='SalePerCustomer',
col='StoreType',
row='Promo',
palette='plasma',
hue='StoreType',
kind='point')
plt.tight_layout()
plt.show()
竞争和促销分析 competition Promo2
- 销售额和客流规模最大的a类型商店，在持续促销时间远低于最高的b,在竞争对手开业时间上也低于b
- b类型商店的日均销售额和客流最高，同时在持续促销时间和竞争也是最高的
#竞争店铺开业时间（按月）
train_store['CompetitionOpen'] = 12*(train_store['Year'] - train_store['CompetitionOpenSinceYear']) + \
(train_store['Month'] - train_store['CompetitionOpenSinceMonth'])
#Promo2持续促销持续时间
train_store['Promo2Open'] = 12*(train_store['Year'] - train_store['Promo2SinceYear']) + \
(train_store['WeekOfYear'] - train_store['Promo2SinceWeek'])/4
Promo2_Compet = train_store.loc[:,['StoreType','Sales','Customers','Promo2Open','CompetitionOpen']].groupby('StoreType').mean()
print(Promo2_Compet)
Sales Customers Promo2Open CompetitionOpen
StoreType
a 6925.697986 795.422370 12918.492198 56.215394
b 10233.380141 2022.211825 17199.328069 58.759512
c 6933.126425 815.538073 12158.636107 57.506506
d 6822.300064 606.353935 10421.916846 51.576795
相关性分析:数值型数据
- 剔除open标签和4个分类标签，计算特征相关性矩阵
- 正相关性
- Sales和Customers：客流量和销售额强正相关
- Promo与Sales、Customers：促销与销售额和客流正相关
- 负相关性
- Promo2与Sales、Customers：持续促销会导致销售额和客流下降
corr_all = train_store.drop(['Open','StateHoliday','StoreType','Assortment','PromoInterval'],axis=1).corr()
mask = np.zeros_like(corr_all,dtype=np.bool_)
mask[np.triu_indices_from(mask)] = True
f,ax = plt.subplots(figsize=(11,9))
sns.heatmap(corr_all,mask=mask,square=True,linewidths=.5,ax=ax,cmap='BuPu')
plt.tight_layout()
plt.show()
Promo和Promo2对销售额影响
- 当完全没有促销的时候（Promo、Promo2都为0），销售额在周日达到峰值，前面分析可知c类店铺周日不开门，所以这里贡献销售额主要是abd店铺
- 在有促销但是没有持续促销Promo2的时候，销售额峰值出现在周一（同时有促销和持续促销也呈现这个趋势）
- 当只有持续促销Promo2的时候，整体销量低于其他三种情况，对销量提升影响不明显，这个在热度图里也有显示
sns.catplot(data=train_store,
x='DayOfWeek',
y='Sales',
col='Promo',
row='Promo2',
hue='Promo2',
palette='RdPu',
kind='point')
plt.tight_layout()
plt.show()
总结
- a类型店铺销售总额和客流总量最高
- d类店铺的客均销售额最高，顾客购买额具有优势，其中在有促销但没持续促销的时候，客户购买额更高；公司可以考虑给d类店铺提供更广的商品多样性
- b类店铺的客均销售额和销售总额规模最低，但是平均每日的销售额和客流是最高的，说明客户购买的物品价值较低，销售规模低原因之一是因为店铺数量较少，
b店铺在持续促销和竞争上都是做的最多和竞争最大的，在客流上和购买转化上具有潜力和优势 - 促销获得有利于提升销售额和客流，在有促销的时候顾客会在周一买的更多，如果没有任何促销会在周日买的更多
- 如果只有持续促销，并不能有效提升销售额
Seasonality：不同类型店铺的销售额时间序列趋势
- 画图可知acd店铺销售趋势较为一致，b店铺由于销售额整体较小，波动更明显但整体趋势相差不大
- 这种方法无法分离趋势（trend）、季节性（seasonal）和残差（residual），只能观察到表面趋势
-深入分析长期趋势或周期性规律，seasonal_decompose 将时间序列分解为趋势 (trend)、季节性 (seasonal)、残差 (residual)三部分
可以清晰看到长期变化趋势和季节性变动
train_store['year_week'] = train_store['Date'].dt.to_period('W').apply(lambda x:x.start_time)
StoreType_sale_trends = train_store.groupby(['StoreType','year_week']).agg({'Sales':'sum'}).reset_index()
StoreType_sale_trends.columns = ['StoreType','year_week','Sales_sum']
store_types = StoreType_sale_trends['StoreType'].unique()
fig,ax = plt.subplots(4,1,figsize=(10,6))
for i,type in enumerate(store_types):
subset = StoreType_sale_trends[StoreType_sale_trends['StoreType'] == type]
ax[i].plot(subset['year_week'],subset['Sales_sum'],label=f'店铺类型{type}')
ax[i].set_title(f'店铺类型{type}的销售额趋势')
ax[i].set_xlabel('日期')
ax[i].set_ylabel('销售额')
ax[i].legend()#图例
plt.tight_layout()
plt.show()
seasonal_decompose:深入分析时间序列趋势
- 用seasonal_decompose进行时间序列分解，发现b店铺整体销售额呈现上升趋势，c店铺销售额2014年7月跌倒谷底后目前上升接近历史最高点,店铺a和d均出现销售额降低，其中a下降幅度最大
StoreType_sale_trends['Sales_sum'] = StoreType_sale_trends['Sales_sum']*1.0#转换维浮点数
#提取不同类型店铺销售额数据
sale_a = StoreType_sale_trends[StoreType_sale_trends['StoreType']=='a'].set_index('year_week')['Sales_sum']
sale_b = StoreType_sale_trends[StoreType_sale_trends['StoreType']=='b'].set_index('year_week')['Sales_sum']
sale_c = StoreType_sale_trends[StoreType_sale_trends['StoreType']=='c'].set_index('year_week')['Sales_sum']
sale_d = StoreType_sale_trends[StoreType_sale_trends['StoreType']=='d'].set_index('year_week')['Sales_sum']
print(sale_a.head())
c = 'blue'
f,(ax1,ax2,ax3,ax4) = plt.subplots(4,1,figsize=(12,13))
decomposition_a = seasonal_decompose(sale_a,model='additive',period=52)#一年52周
decomposition_a.trend.plot(color=c,ax=ax1)#trend是提取分解函数里面的趋势，seasonal、resid就是提取季节性和残差
ax1.set_title('店铺类型 a 的销售额趋势')
decomposition_b = seasonal_decompose(sale_b,model='additive',period=52)
decomposition_b.trend.plot(color=c,ax=ax2)
ax2.set_title('店铺类型 b 的销售额趋势')
decomposition_c = seasonal_decompose(sale_c,model='additive',period=52)
decomposition_c.trend.plot(color=c,ax=ax3)
ax3.set_title('店铺类型 c 的销售额趋势')
decomposition_d = seasonal_decompose(sale_d,model='additive',period=52)
decomposition_d.trend.plot(color=c,ax=ax4)
ax4.set_title('店铺类型 d 的销售额趋势')
plt.tight_layout()
plt.show()
自相关和偏自相关：Autocorrelation Function (ACF) Partial Autocorrelation Function (PACF)
- 每个图表都呈现的2个特点：时间序列具有非随机性（当前数据和滞后期的数据之间存在显著的相关性），滞后 1 阶的相关性较高
- a、b、d店铺：都呈现出季节性特征，对于a类型店铺呈现出周的趋势,在8、15、22、29、36、43、50都出现正的峰值；
b和d类型店铺类似也出现周的趋势 - c类型店铺较为复杂，看起来每个观测值都与其相邻的观测值存在相关性，整体滞后正值峰值出现在13、24、36、48，并且收敛
StoreType_sale_dayTrends = train_store.groupby(['StoreType','Date']).agg({'Sales':'sum'}).reset_index()
StoreType_sale_dayTrends.columns = ['StoreType','Date','Sales_sum']
print(StoreType_sale_dayTrends.head())
sale_a = StoreType_sale_dayTrends[StoreType_sale_dayTrends['StoreType']=='a'].set_index('Date')['Sales_sum']
sale_b = StoreType_sale_dayTrends[StoreType_sale_dayTrends['StoreType']=='b'].set_index('Date')['Sales_sum']
sale_c = StoreType_sale_dayTrends[StoreType_sale_dayTrends['StoreType']=='c'].set_index('Date')['Sales_sum']
sale_d = StoreType_sale_dayTrends[StoreType_sale_dayTrends['StoreType']=='d'].set_index('Date')['Sales_sum']
c = 'blue'
plt.figure(figsize=(12,8))
#ACF PACF分析和可视化
plt.subplot(421)
plot_acf(sale_a,lags=50,ax=plt.gca(),color=c)
plt.title('a店铺的ACF')
plt.subplot(422)
plot_pacf(sale_a,lags=50,ax=plt.gca(),color=c)
plt.title('a店铺的PACF')
plt.subplot(423)
plot_acf(sale_b,lags=50,ax=plt.gca(),color=c)
plt.title('b店铺的ACF')
plt.subplot(424)
plot_pacf(sale_b,lags=50,ax=plt.gca(),color=c)
plt.title('b店铺的PACF')
plt.subplot(425)
plot_acf(sale_c,lags=50,ax=plt.gca(),color=c)
plt.title('c店铺的ACF')
plt.subplot(426)
plot_pacf(sale_c,lags=50,ax=plt.gca(),color=c)
plt.title('c店铺的PACF')
plt.subplot(427)
plot_acf(sale_d,lags=50,ax=plt.gca(),color=c)
plt.title('d店铺的ACF')
plt.subplot(428)
plot_pacf(sale_d,lags=50,ax=plt.gca(),color=c)
plt.title('d店铺的PACF')
plt.tight_layout()
plt.show()
用Prophet进行时间序列分析和预测
SARIMA算法介绍
- Prophet 是由 Facebook 开发的一个时间序列预测模型，专为处理具有强季节性、节假日效应和缺失值的数据而设计。它对实际业务场景中的时间序列预测非常有效，尤其适用于零售销售预测这类数据。
- 优点：
- 适应性强： Prophet 对数据中可能存在的缺失值、异常值（如促销活动）以及节假日影响非常容忍，自动进行修正。
- 灵活性： Prophet 支持用户在模型中加入节假日效应，能够捕捉到不同时间段对销量的特殊影响（如节假日促销、季节性促销等）。
- 易于使用： Prophet 的 API 接口简单，用户可以很方便地调节季节性（如年、周、日季节性）和趋势（如线性或对数趋势）。
- 处理缺失值： 对于数据缺失（例如有些日期没有销售记录），Prophet 可以处理缺失数据并生成合理的预测。
- 扩展性： 可以通过 add_seasonality 添加自定义季节性（如促销周期），从而提高模型的拟合能力。
- 趋势建模： Prophet 可以处理更复杂的趋势（如非线性趋势），并且能够自动进行趋势的切换（如增长变化）。
- 缺点：
- 不适用于非常复杂的季节性和自相关模式： 对于具有复杂自相关关系或需要精细控制的季节性模式，Prophet 可能不如 SARIMA 精确。
- 预测不如统计模型精细： 在某些情况下，Prophet 可能无法捕捉到短期波动，尤其在数据没有显著季节性时，可能预测效果不如传统统计模型。
- 需要大数据量： 如果数据量较小，Prophet 可能没有传统统计方法（如 ARIMA/SARIMA）准确。
- 预测不透明： Prophet 模型虽然可以调节趋势和季节性，但其背后的数学和假设可能不如 ARIMA 模型那样直观易懂。
挑选规模最大的店铺进行预测
- b类店铺262销售额最大，后面针对262店铺进行预测
#挑选销售规模最大的店铺
choose_store = train_store.groupby('Store')['Sales'].sum()
choose_store = choose_store.sort_values(ascending=False)
#选择262店铺的数据
sales = train_store[train_store['Store']==262].loc[:,['Date','Sales']]
sales = sales.sort_index(ascending=False)
sales = sales.rename(columns = {'Date':'ds',
'Sales':'y'})
节日建模
state = train_store[(train_store['StateHoliday']=='a') | (train_store['StateHoliday']=='b') | \
(train_store['StateHoliday']=='c')].loc[:,'Date'].values
#将属于公共节日的日期以或的关系提取，用.values转换为Numpy数组方便后续分析
school= train_store[train_store['SchoolHoliday']==1].loc[:,'Date'].values
state_holiday = pd.DataFrame({'ds':pd.to_datetime(state),
'holiday':'state_holiday'})
school_holiday = pd.DataFrame({'ds':pd.to_datetime(school),
'holiday':'school_holiday'})
holidays = pd.concat((state_holiday,school_holiday))
print(holidays.shape)#(164355, 2)
#节日列存在重复日期，会影响模型学习，剔除掉重复日期保留第一次出现的日期
holidays = holidays.drop_duplicates(subset='ds')
模型训练与销售预测
- d 表示差分阶数，目的是使数据平稳。如果数据的时间序列呈现趋势或季节性模式，通常我们需要对数据进行差分，使其变为平稳序列
- 结果显示数据平稳不需要差分，d设置为0
model = Prophet(interval_width=0.95,holidays=holidays)
model.fit(sales)
#3.1预测未来6周销售额
future_dates = model.make_future_dataframe(periods=6*7)#包含历史数据
future_dates_only = future_dates[future_dates['ds'] > max(sales['ds'])]#8.1-9.11日
print('第一周预测的日期')
print(future_dates_only.head(7))
prediction = model.predict(future_dates)#传入带历史日期的记录，预测时会考虑历史数据的季节性和趋势
#3.2 预测262店铺最后一周销售额
store262_predict = prediction[['ds','yhat','yhat_lower','yhat_upper']].tail(42)
print(f'262店铺未来6周销售额预测{store262_predict}')
store262_predict = prediction[['ds','yhat']].rename(columns={'Date':'ds','Forecast':'yhat'})
#3.2.1 画出预测图：蓝线是预测线，阴影是预测范围，黑点是实际数据点
model.plot(prediction)
#3.2.2 可视化预测的组成部分
model.plot_components(prediction)
plt.legend()
plt.show()
262店铺未来6周销售额预测 ds yhat yhat_lower yhat_upper
942 2015-08-01 17234.190257 12208.043786 22416.554313
943 2015-08-02 28063.683535 23192.079313 32952.759110
944 2015-08-03 19420.934658 14500.894748 24361.886658
945 2015-08-04 17864.460690 13094.773149 22417.402116
946 2015-08-05 17618.124573 12946.205424 22725.681036
947 2015-08-06 17936.488957 12787.462575 22635.539444
948 2015-08-07 19694.229881 15078.890684 24540.804405
949 2015-08-08 16753.085213 12126.091182 21549.504269
950 2015-08-09 27641.550635 22810.874869 32483.543202
951 2015-08-10 19069.334748 14605.754465 23702.843189
952 2015-08-11 17592.885347 12375.850373 22271.527735
953 2015-08-12 17433.706635 12735.700850 22418.450359
954 2015-08-13 17843.782212 13058.806161 22830.535524
955 2015-08-14 19695.064781 15218.925616 24593.917689
956 2015-08-15 16846.505233 12004.838855 21431.927248
957 2015-08-16 27823.831554 23054.821879 32269.041341
958 2015-08-17 19334.086654 14500.173574 24475.998952
959 2015-08-18 17931.234731 13094.005791 23042.451292
960 2015-08-19 17834.553546 13168.850093 22555.407250
961 2015-08-20 18294.125192 13589.728137 23138.315408
962 2015-08-21 20180.384252 15477.955809 25293.907834
963 2015-08-22 17351.194233 12230.919415 22601.939881
964 2015-08-23 28331.661260 23820.201379 32803.272704
965 2015-08-24 19828.693032 14910.595117 24897.845018
966 2015-08-25 18396.611955 13656.036864 23117.414699
967 2015-08-26 18255.539550 13516.145729 22887.190717
968 2015-08-27 18656.864879 13785.888786 23438.105489
969 2015-08-28 20472.756464 15787.114970 25235.704079
970 2015-08-29 17563.189569 12429.179493 22111.591806
971 2015-08-30 28455.699229 23571.438049 33355.319867
972 2015-08-31 19859.868726 14934.777428 24613.809618
973 2015-09-01 18332.864480 13566.166810 23067.041980
974 2015-09-02 18097.736687 13356.508499 22894.738611
975 2015-09-03 18408.800899 13614.450580 23111.019541
976 2015-09-04 20141.062542 15345.483022 25441.971188
977 2015-09-05 17157.158838 11966.498300 22054.083587
978 2015-09-06 27987.030986 23412.701507 32742.957406
979 2015-09-07 19342.338753 14559.944652 24269.852628
980 2015-09-08 17781.930613 12958.583042 22464.722412
981 2015-09-09 17530.090732 12313.025153 22549.582363
982 2015-09-10 17841.879638 13012.276702 22931.404952
983 2015-09-11 19592.531822 14665.422950 24282.799868
扫一扫
1万+
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
