R手册(Time Series)--forecast and prophet

时间序列分析(Time-Series Analysis)是指将原来的数据分解为四部分来看——趋势、周期、时期和不稳定因素， 然后综合这些因素， 提出预测。

forecast : for Time Series and Linear Models

时间序列分析

tsclean(x, replace.missing = TRUE, lambda = NULL)识别和替换异常值和缺失值（lambda给出Box-Cox变换参数的数值）
ndiffs, nsdiffs固定系列所需的差异数
seasonal(object)提取季节分量
trendcycle(object)提取趋势周期分量
remainder(object)提取余数分量
findfrequency查找时间序列的主频
ma(x, order, centre = TRUE)计算更平滑的移动平均

模型

arfima: FitARFIMAmodel
Arima, auto.arima : FitARIMAmodel
ets(y,model=”ZZZ”)指数预测模型
baggedETS, bats, tbats: FitbaggedETS/BATS/TBATSmodel
nnetar神经网络时间序列预测

forecastHybrid: 组合模型

hybridModel(y, models = "aefnst",                         #模型组a(auto.arima),e(ets),f(thetam),n(nnetar),s(stlm),t(tbats)
  weights = c("equal", "insample.errors", "cv.errors"),    #模型加权方法
  parallel = FALSE,                                        #是否并行运算
  num.cores = 2L)                                          #并行内核数

example: 
hybridModel(wineind, models = "aet", weights = "equal")%>%
  forecast(hm1, h = 48)%>%plot()

预测

forecast(object, h = ifelse(frequency(object) > 1, 2*frequency(object), 10),level=c(80,95))

参数：
h：预测数
level：置信区间

ggplot2扩展

Acf, Pacf, Ccf, taperedacf, taperedpacf自相关和相关函数估计
autoplot(object, …)通用制图函数

模型评估

checkresiduals(object, lag, df = NULL, test, plot = TRUE, ...)检查残差
accuracy(f, x)准确率
CV, CVar, dsCV交叉验证
dm.testDiebold-Mariano测试的预测准确性

example: WWWusage %>%auto.arima %>%forecast(h=20) %>%autoplot()

prophet

模型组成：Y(t)=Trend(t)+Seasonal(t)+Holiday(t)+Irregular(t)

构建模型

prophet(df = df,              #data.frame:ds(date type)+ y,cap+floor指定饱和最大值和最小值
  growth = "linear",           #linearorlogistictrend
  changepoints = NULL,          #包含潜在变化点的日期向量
  n.changepoints = 25,          #潜在变化点数
  changepoint_prior_scale=0.05,   #调整trend灵活性
  yearly.seasonality = "auto",     #适合年度seasonality;'auto',TRUE,orFALSE
  weekly.seasonality = "auto",     #适合周度
  holidays = NULL,                 #data.frame:holiday(character)+ds(datetype),lower_window+upper_window(可选，指定假日周围的天数)
  seasonality.prior.scale = 10,     #调整季节性模型的强度
  holidays.prior.scale = 10,        #调整假期组件模型的强度
  mcmc.samples = 0, 
  interval.width = 0.8,             #trend间隔不确定性
  uncertainty.samples = 1000,       #season的不确定习性
  fit = TRUE)

模型预测

furture<-make_future_dataframe(m,           #Prophet model object|
  periods,                                  #要预测的数量
  freq = "d",                               #day','week','month','quarter',or'year'
  include_history = TRUE)                   #历史日期是否包含在预测中

predict(object,df = NULL)

参数：
object：Prophet modelo bject
df：NULL or future

可视化

plot(x, fcst)
prophet_plot_components(m, fcst) ggplot2组件，将预测细分为趋势，每周季节性和年度季节

交叉验证

cross_validation(m, horizon ,initial,units = 'days')

参数horizon, initial, units：初始日期，截止日期，间隔