Question

我已经建立了一种创建纠错模型（ECM）的方法，纠错模型是多个ECM的平均值。为此，我利用R中的lm()函数来创建多个表示ECM的lm对象。我平均每个对象的系数以获得最终模型。 lm对象表示ECM的方式是，在对数据运行lm()之前，我将数据转换为ECM所需的格式。

我还使用AIC向后选择来消除不需要的变量。在创建ECM时，该过程似乎运行良好。但是，当我创建一个具有与模型中的系数相匹配的列名的数据框时，我收到一条错误消息，指出数据中缺少必要的变量。但是，在最终模型中未包含此变量，因此不必进行预测。那么，predict()为什么要寻找该变量？我在做什么错了？

#Load data
library(ecm)
data(Wilshire)
trn <- Wilshire[Wilshire$date<='2015-12-01',]
y <- trn$Wilshire5000
xeq <- xtr <- trn[c('CorpProfits', 'FedFundsRate', 'UnempRate')]

#Function to split data into k partitions and build k models, each on a (k-1)/k subset of the data
avelm <- function(formula, data, k = 5, seed = 5, ...) {
  lmall <- lm(formula, data, ...)
  modellist <- 1:k
  set.seed(seed)
  models <- lapply(modellist, function(i) {
    tstIdx <- sample(nrow(data), 1/k * nrow(data))
    trn <- data[-tstIdx, ]
    lm(as.formula(formula), data = trn)
  })
  lmnames <- names(lmall$coefficients)
  lmall$coefficients <- rowMeans(as.data.frame(sapply(models, function(m) coef(m))))
  names(lmall$coefficients) <- lmnames
  lmall$fitted.values <- predict(lmall, data)
  target <- trimws(gsub("~.*$", "", formula))
  lmall$residuals <- data[, target] - lmall$fitted.values
  return(lmall)
}

#Function to create an ECM using backwards selection based on AIC (leveraged avelm function above)
aveecmback <- function (y, xeq, xtr, k = 5, seed = 5, ...) {
  xeqnames <- names(xeq)
  xeqnames <- paste0(xeqnames, "Lag1")
  xeq <- as.data.frame(xeq)
  xeq <- rbind(rep(NA, ncol(xeq)), xeq[1:(nrow(xeq) - 1), ])

  xtrnames <- names(xtr)
  xtrnames <- paste0("delta", xtrnames)
  xtr <- as.data.frame(xtr)
  xtr <- data.frame(apply(xtr, 2, diff, 1))
  yLag1 <- y[1:(length(y) - 1)]
  x <- cbind(xtr, xeq[complete.cases(xeq), ])
  x <- cbind(x, yLag1)
  names(x) <- c(xtrnames, xeqnames, "yLag1")
  x$dy <- diff(y, 1)
  formula <- "dy ~ ."

  model <- avelm(formula, data = x, k = k, seed = seed, ...)
  fullAIC <- partialAIC <- AIC(model)
  while (partialAIC <= fullAIC) {
    todrop <- rownames(drop1(model))[-grep("none|yLag1", rownames(drop1(model)))][which.min(drop1(model)$AIC[-grep("none|yLag1", rownames(drop1(model)))])]
    formula <- paste0(formula, " - ", todrop)
    model <- avelm(formula, data = x, seed = seed, ...)
    partialAIC <- AIC(model)
    if (partialAIC < fullAIC & length(rownames(drop1(model))) > 2) {
      fullAIC <- partialAIC
    }
  }

  return(model)
}

finalmodel <- aveecmback(y, xeq, xtr)
print(finalmodel)

Call:
lm(formula = formula, data = data)

Coefficients:
     (Intercept)  deltaCorpProfits    deltaUnempRate   CorpProfitsLag1             yLag1  
       -0.177771          0.012733         -1.204489          0.002046         -0.024294  

#Create data frame to predict on
set.seed(2)
df <- data.frame(deltaCorpProfits=rnorm(5), deltaUnempRate=rnorm(5), CorpProfitsLag1=rnorm(5), yLag1=rnorm(5))

predict(finalmodel, df)
Error in eval(predvars, data, env) : object 'deltaFedFundsRate' not found

Answer 1

我知道了。问题出在aveecmback()函数的一部分中，我在while循环内修改了formula。相反，如果我修改x来删除变量，问题就解决了。这是因为即使在公式中将其删除，这样的数据在数据框中仍然需要disp：

data(mtcars)
model <- lm(mpg~.-disp, mtcars)
predict(model, mtcars[-which(names(mtcars) %in% 'disp')])
Error in eval(predvars, data, env) : object 'disp' not found

但是，类似这样的操作将允许predict()在没有disp的数据帧上工作：

data(mtcars)
model <- lm(mpg~., mtcars[-which(names(mtcars) %in% 'disp')])
predict(model, mtcars[-which(names(mtcars) %in% 'disp')])
          Mazda RX4       Mazda RX4 Wag          Datsun 710      Hornet 4 Drive   Hornet Sportabout 
           22.37587            22.07853            26.58631            20.82285            17.26052 
            Valiant          Duster 360           Merc 240D            Merc 230            Merc 280 
           20.46572            14.04956            22.38273            24.20323            18.97756 
          Merc 280C          Merc 450SE          Merc 450SL         Merc 450SLC  Cadillac Fleetwood 
           19.37670            15.10244            16.12864            16.26339            11.31787 
Lincoln Continental   Chrysler Imperial            Fiat 128         Honda Civic      Toyota Corolla 
           10.68985            10.65062            28.03687            29.29545            29.42472 
      Toyota Corona    Dodge Challenger         AMC Javelin          Camaro Z28    Pontiac Firebird 
           23.72382            16.91215            17.78366            13.53713            16.15156 
          Fiat X1-9       Porsche 914-2        Lotus Europa      Ford Pantera L        Ferrari Dino 
           28.35383            26.31886            27.36155            18.86561            19.75073 
      Maserati Bora          Volvo 142E 
           13.86302            24.78865

R预测期望变量不在lm对象中

1 个答案: