【时间序列预测实战】【规则与baseline】

前言

规则的好处是简单高效且可解释性强，有时候样本量不足时，使用有用的规则比建模更有效，面对问题的时候不能只想着建模，而是根据问题去寻找解决的方案有哪些，互相比较，找到最好的解决方案

时间序列预测的规则

简单的统计量

• 中位数：很稳健
• 均值：分布较符合正态分布时适用
• 临近数据：指与待预测数据较近的数据，例如用8点的数据预测9点效果

周期因子

基于周的周期因子预测

周期通常有一周、一月、一年，支付行为、客流量、交通数据等时间序列都具有明显的周期性，缺点是不考虑特定日子和节假日。

若以周为周期，则如何计算周期因子?
第一种方式：除以周均值，然后按列取中位数

第二种方式：先获得列的均值（周内每日），再除以整体均值（整体均值就是所有值的平均值）

第二步就是预测，得到周期因子，就可以预测第四周的客流量，使用第一种方式，中位数直接乘以base，这里的base使用了第三周的均值。

base除了用最后一周的均值，我们可以使用去周期后取平均，即直接用第三周的原始客流量直接去除以中位数，中位数可以一定程度代表算数平均数。

基于月的周期因子预测

1. 对每日的申购总量和赎回总量都求均值
2. 统计周一到周日的频次
3. 使用基于周的周期因子作为加权均值
   
4. 根据因子和每日均值去计算预测值

规则应用到实战中

以星期为周期的中位数预测

train = data[(data['date']>=datetime.datetime(2014,3,31))&(data['date']<=datetime.datetime(2014,8,3))]

test = data[(data['date']>datetime.datetime(2014,8,3))&(data['date']<=datetime.datetime(2014,8,31))]

pred = pd.DataFrame(columns=['date','purchase','redeem'])
s = []
for i in range(4,32):
    a = datetime.datetime(2014,8,i)
    s.append(a)
    
pred['date'] = s

train.reset_index(inplace=True)
test.reset_index(inplace=True)

del train['index']
del test['index']


pred['purchase'] = test['total_purchase_amt']
pred['redeem'] = test['total_redeem_amt']
pred['date'] = pd.to_datetime(pred['date'])
pred['weekday'] = pred['date'].dt.weekday + 1

df_purchase = pd.DataFrame(columns=['weekday1','weekday2','weekday3','weekday4','weekday5','weekday6','weekday7'])

count = 0

for i in range(len(train)):
    df_purchase.loc[count,'weekday'+str((train.loc[i,'weekday'])+1)] = train.loc[i,'total_purchase_amt']
    if train.loc[i,'weekday'] == 6 :
        count = count + 1

df_redeem = pd.DataFrame(columns=['weekday1','weekday2','weekday3','weekday4','weekday5','weekday6','weekday7'])

count = 0

for i in range(len(train)):
    df_redeem.loc[count,'weekday'+str((train.loc[i,'weekday'])+1)] = train.loc[i,'total_redeem_amt']
   
    if train.loc[i,'weekday'] == 6 :
        count = count + 1

def predictByWeekdayFactor(df_,weekId):  #周期因子第一种计算方式
    
    df_['week_mean'] = df_.mean(axis=1)
    
    for i in range(7):
        df_['weekday' + str(i+1) + '_coefficient'] = df_['weekday' + str(i+1)]/ df_['week_mean'] 
    coef_list = ['weekday1_coefficient','weekday2_coefficient','weekday3_coefficient','weekday4_coefficient','weekday5_coefficient','weekday6_coefficient','weekday7_coefficient']
        
    weekday_rate = df_[coef_list].median()
        
    base = np.array(df_.loc[weekId,['weekday1','weekday2','weekday3','weekday4','weekday5','weekday6','weekday7']])  * 1/np.array(df_[coef_list].median())
    
    Pre = np.mean(base) * np.array(weekday_rate)
    return Pre 

以星期为周期的流入流出预测

trade_weekday = train.groupby(['weekday'])['total_purchase_amt','total_redeem_amt'].mean()    # 周内每日的均值
trade_weekday.reset_index(inplace=True)

purchase_mean = np.mean(train['total_purchase_amt'])    # 整体均值
redeem_mean = np.mean(train['total_redeem_amt'])

#周期因子
purchase_weekday_rate_2 = trade_weekday['total_purchase_amt']/purchase_mean
redeem_weekday_rate_2 = trade_weekday['total_redeem_amt']/ redeem_mean

basePurchase_2 = np.mean(df_purchase_copy2.loc[len(df_purchase_copy2)-1,['weekday1','weekday2','weekday3','weekday4','weekday5','weekday6','weekday7']]*1/np.array(purchase_weekday_rate_2))

purchasePre_2 = basePurchase_2 * purchase_weekday_rate_2

baseRedeem_2 = np.mean(df_redeem_copy2.loc[len(df_redeem_copy2)-1,['weekday1','weekday2','weekday3','weekday4','weekday5','weekday6','weekday7']]*1/np.array(redeem_weekday_rate_2))

redeemPre_2 = baseRedeem_2 * redeem_weekday_rate_2

以月份为周期的流入流出预测

def predictByMonthDayFactor(df_,purchase_weekday_rate,redeem_weekday_rate):
    
    # 获得每日均值
    trade_day = df_.groupby(['day'])['total_purchase_amt','total_redeem_amt'].mean()
    trade_day.reset_index(inplace=True)
    
    #初始化星期的列和周期因子列
    for i in range(7):
        trade_day['weekday'+str(i+1)] = 0
    trade_day['purchase_day_rate'] = 0
    trade_day['redeem_day_rate'] = 0
    
    #对每一天都判断，不知道啥操作，好像是和星期有关
    for i in range(31):
        for j in range(7):
            trade_day.loc[i,'weekday'+str(j+1)] = sum((df_['day'==i+1])&(df_['weekday']==j+1))
            
    for i in range(31):
        trade_day.loc[i,'purchase_day_rate'] = sum(np.array(purchase_weekday_rate)*np.array(trade_day.loc[i,['weekday1','weekday2','weekday3','weekday4','weekday5','weekday6','weekday7']])) / sum(trade_day.loc[i,['weekday1','weekday2','weekday3','weekday4','weekday5','weekday6','weekday7']])
        trade_day.loc[i,'redeem_day_rate'] = sum(np.array(redeem_weekday_rate)*np.array(trade_day.loc[i,['weekday1','weekday2','weekday3','weekday4','weekday5','weekday6','weekday7']])) / sum(trade_day.loc[i,['weekday1','weekday2','weekday3','weekday4','weekday5','weekday6','weekday7']])
            
    trade_day['purchasePre'] = trade_day['total_purchase_amt']/trade_day['purchase_day_rate']
    
    trade_day['redeemPre'] = trade_day['total_redeem_amt']/trade_day['redeem_day_rate']
    
    return trade_day
        
trade_day= predictByMonthDayFactor(train,purchase_weekday_rate_2,redeem_weekday_rate_2)

pred['day'] = pred['date'].dt.day
pred = pd.merge(pred,trade_day[['day','purchasePre','redeemPre']])

for i in range(len(pred)):
    pred.loc[i,'purchasePre'] = pred.loc[i,'purchasePre'] * purchase_weekday_rate_2[pred,loc[i,'weekday']-1]
    
    pred.loc[i,'redeemPre'] = pred.loc[i,'redeemPre'] * redeem_weekday_rate_2[pred,loc[i,'weekday']-1]