R语言自定义阶数SARIMAX模型xreg参数报错问题求助

问题背景 #

用R做带外生变量的时间序列预测，已完成训练集和测试集拆分，用auto.arima()自动选择阶数时预测正常，但手动指定ARIMA阶数后运行预测代码出现错误。

数据处理与拆分代码 #

data$date <- as.Date(data$date, format = "%Y-%m-%d")
time <- data$date
time <- tail(time, -1)

df <- data.frame(date=time,
                avi96 = as.numeric(avi96_adjseason_diff),
                 avi89 = as.numeric(avi89_adjseason_diff),
                 avi82 = as.numeric(avi82_adjseason_diff),
                 avi75 = as.numeric(avi75_adjseason_diff),
                 msci = as.numeric(msci_adjseason_diff))

train <- df[1:45, ]
test <- df[45+1:nrow(df), ]
test <- na.omit(test)

xreg <- train$msci
xreg1 <- test$msci

ts_train <- ts(train[,-1], start=c(2019,1),frequency=12)
ts_test <- ts(test[,-1],start=c(2022,10),frequency=12)

可正常运行的auto.arima代码 #

fit1 <- auto.arima(ts_train[,1], xreg = ts_train[,5])
fcast <- forecast::forecast(fit1, xreg = xreg1, h = 6)
autoplot(fcast)

出错的手动指定阶数代码及错误信息 #

fit1 <- arima(ts_train[,1], order = c(5, 0, 5), seasonal = list(order = c(1, 0, 0), period = 12), xreg = ts_train[,5])

fcast <- forecast::forecast(fit1, xreg = xreg1, h = 6)
autoplot(fcast)

错误信息：

Warning message in forecast.Arima(fit1, xreg = xreg1, h = 6):
“xreg not required by this model, ignoring the provided regressors”
Error in predict.Arima(object, n.ahead = h): 'xreg' and 'newxreg' have different numbers of columns
Traceback:

1. forecast::forecast(fit1, xreg = xreg1, h = 6)
2. forecast.Arima(fit1, xreg = xreg1, h = 6)
3. predict(object, n.ahead = h)
4. predict.Arima(object, n.ahead = h)
5. stop("'xreg' and 'newxreg' have different numbers of columns")

错误原因 #

1. 外生变量被自动剔除：arima()函数会自动检验外生变量的系数显著性，如果系数不显著，会直接将其从模型中移除，导致拟合后的模型不再包含xreg，触发第一个警告。
2. 格式不匹配：训练时传入的xreg是向量，测试集的xreg1也是向量，但arima()内部处理时可能将向量视为“无列名的单列结构”，预测时容易出现列数匹配异常；另外原代码中test <- df[45+1:nrow(df), ]的索引写法有问题，应该是(45+1):nrow(df)，否则可能导致测试集行数错误。

解决方法 #

1. 统一外生变量格式为矩阵 #

将训练和测试的外生变量都转为单列矩阵，确保列数和结构一致：

xreg <- as.matrix(train$msci)
xreg1 <- as.matrix(test$msci)

2. 修正测试集索引 #

原索引写法会生成错误的序列，改成正确的区间索引：

test <- df[(45+1):nrow(df), ]

3. 强制保留外生变量（可选） #

如果外生变量被自动剔除，可通过fixed参数强制约束其系数不为0，确保模型保留外生变量：

# 先拟合临时模型查看参数位置
fit_temp <- arima(ts_train[,1], order = c(5,0,5), seasonal = list(order = c(1,0,0), period=12), xreg = xreg)
# 构造fixed向量：NA表示允许自由估计，确保外生变量系数位置不为0
fixed_vec <- rep(NA, length(fit_temp$coef))
# 找到外生变量msci对应的系数位置（通常是最后一个）
msci_coef_pos <- grep("msci", names(fit_temp$coef))
fixed_vec[msci_coef_pos] <- NA # 设为NA允许估计，若设为固定值则强制该系数为指定值

# 重新拟合模型
fit1 <- arima(ts_train[,1], 
              order = c(5, 0, 5), 
              seasonal = list(order = c(1, 0, 0), period = 12), 
              xreg = xreg,
              fixed = fixed_vec)

4. 修正预测代码中的变量名 #

原代码中autoplot(fcast_cons_temp)的变量名错误，应该用fcast：

fcast <- forecast::forecast(fit1, xreg = xreg1, h = 6)
autoplot(fcast)

完整修正代码 #

# 数据处理
data$date <- as.Date(data$date, format = "%Y-%m-%d")
time <- tail(data$date, -1)

df <- data.frame(
  date = time,
  avi96 = as.numeric(avi96_adjseason_diff),
  avi89 = as.numeric(avi89_adjseason_diff),
  avi82 = as.numeric(avi82_adjseason_diff),
  avi75 = as.numeric(avi75_adjseason_diff),
  msci = as.numeric(msci_adjseason_diff)
)

# 修正测试集索引
train <- df[1:45, ]
test <- df[(45+1):nrow(df), ]
test <- na.omit(test)

# 统一外生变量为矩阵格式
xreg <- as.matrix(train$msci)
xreg1 <- as.matrix(test$msci)

ts_train <- ts(train[, -1], start = c(2019,1), frequency = 12)
ts_test <- ts(test[, -1], start = c(2022,10), frequency = 12)

# 手动指定阶数拟合ARIMA
fit1 <- arima(ts_train[,1], 
              order = c(5, 0, 5), 
              seasonal = list(order = c(1, 0, 0), period = 12), 
              xreg = xreg)

# 预测并绘图
fcast <- forecast::forecast(fit1, xreg = xreg1, h = 6)
autoplot(fcast)

内容的提问来源于stack exchange，提问作者Laura_777