R手册(Time Series)--zoo,forecast and prophet

R手册(Time Series)–zoo,forecast and prophet

---

zoo

Part 1 基础对象

zoo(x = NULL, order.by = index(x), frequency = NULL)有序的时间序列对象
zooreg(data, start = 1, end = numeric(), frequency = 1)规则的的时间序列对象，继承zoo对象
as.zoo(x)把一个对象转型为zoo类型，泛型函数
is.regular(x, strict = FALSE)检查是否是规则的序列

Part 2 ggplot2扩展

autoplot(object, geom = "line", facets, ...)
fortify(model, data, melt = FALSE, …)数据操作

Part 3 数据清洗

read.zoo(file, format = "", index.column = 1, drop = TRUE)
write.zoo(x, file = "", index.name = "Index", row.names = FALSE, col.names = NULL)
coredata(x)提取/替换zoo数据部分
index(x)提取/替换zoo索引部分
window(x, index. = index(x), start = NULL, end = NULL)按时间筛选数据
merge()合并多个zoo对象
aggregate(x, by, FUN = sum)分类计算
lag(x, k = 1, na.pad = FALSE, ...)计算步长
diff(x, lag = 1, differences = 1)计算分差
rollapply(data, width, FUN)对zoo数据的滚动处理

rollmean, rollmax, rollmedian, rollsum,etc

MATCH(x, table)值匹配
ORDER(x)值排序，输出索引

Part 4 缺失值处理

na.fill(object, fill)NA值的填充
na.locf(object, na.rm = TRUE, fromLast=FALSE)最近值替换NA
na.aggregate(object,by = 1,FUN = mean,na.rm = FALSE)计算统计值替换NA
na.approx(object)计算插值替换NA
na.StructTS(object,na.rm=FALSE)计算seasonalKalmanfilter替换NA
na.trim(object)过滤有NA的记录

Part 5 显示控制

yearqtr以年季度显示时间
yearmon以年月显示时间

forecast : for Time Series and Linear Models

Part 1 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

Part 2 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()

Part 3 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

Part 4 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)

Part 1 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)

Part 2 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

Part 3 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

Part 4 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