使用fable包在Validation_set上计算准确率指标报错求助

无法让accuracy()函数仅指向验证集validation_set：当该函数传入完整数据集data_tsibble时代码运行正常，但指定validation_set时，会收到“参数必须可转换为非负整数”的错误提示。

Messaggi di avvertimento:
1: 2 errors (1 unique) encountered
[2] l'argomento dev'essere coercibile in un intero non-negativo

2: 2 errors (1 unique) encountered
[2] l'argomento dev'essere coercibile in un intero non-negativo

3: 2 errors (1 unique) encountered
[2] l'argomento dev'essere coercibile in un intero non-negativo

4: 2 errors (1 unique) encountered
[2] l'argomento dev'essere coercibile in un intero non-negativo

5: 2 errors (1 unique) encountered
[2] l'argomento dev'essere coercibile in un intero non-negativo

（注：错误提示为意大利语，翻译为“参数必须可转换为非负整数”）

# 清空工作区所有变量
rm(list = ls())

# 加载所需包
library(tidyverse)
library(fpp3)
library(readxl)
library(distributional)

# 读取数据集
data <- read_excel("D:/TEST.xlsx")

# 转换为tsibble格式（时间序列数据框）
data_tsibble <- data |>
  mutate(Mesi = yearmonth(Mesi)) |>
  as_tsibble(index = Mesi)

dput(data_tsibble)

# 划分训练集与验证集
train_set <- data_tsibble |> filter_index(~ "12/01/2023")
validation_set <- data_tsibble |> filter_index("01/01/2024" ~ .)


# 拟合多个模型
fit_models <- train_set |>
  model(Ets = ETS(Intermediato),
        Arima = ARIMA(Intermediato, stepwise=FALSE, approximation=FALSE),
        tslm1 = TSLM(Intermediato ~ trend() + season()),
        tslm2 = TSLM(Intermediato ~ trend() + I(trend()^2) + season()),
        tslm3 = TSLM(Intermediato ~ trend() + I(trend()^2) + season()))
glance(fit_models) |> arrange(AICc) |> select(.model:BIC)


# 生成预测
fc_models <- fit_models |>
  forecast(h=nrow(validation_set))

# 传入完整数据集时accuracy函数运行正常
fc_models |>
  accuracy(
    data = validation_set,
    measures = list(
      crps = CRPS,
      rmse = RMSE,
      mase = MASE,
      ss_crps = skill_score(CRPS),
      ss_rmse = skill_score(RMSE)
    ),
    n_quantiles = 100
  ) |>
  group_by(.model) %>%
  summarise(
    ss_crps = mean(ss_crps) * 100,
    ss_rmse = mean(ss_rmse) * 100,
    mase = mean(mase)
  )


# 传入完整数据集时accuracy函数运行正常
fc_models |>
  accuracy(validation_set,
           list(rmse=RMSE,
                mae=MAE,
                mape=MAPE,
                mase=MASE,
                rmsse=RMSSE,
                acf1=ACF1,
                crps=CRPS,
                skill_crps=skill_score(CRPS),
                skill_rmse=skill_score(RMSE)))

问题根源在于MASE和RMSSE指标需要依赖训练集数据计算基准误差，仅传入验证集时，函数无法获取训练集信息，导致参数转换错误。提供两种解决方式：

方法一：指定训练集参数 #

在accuracy()中通过train参数传入训练集，让函数能正常计算MASE/RMSSE：

fc_models |>
  accuracy(
    data = validation_set,
    train = train_set,
    measures = list(
      crps = CRPS,
      rmse = RMSE,
      mase = MASE,
      ss_crps = skill_score(CRPS),
      ss_rmse = skill_score(RMSE)
    ),
    n_quantiles = 100
  ) |>
  group_by(.model) %>%
  summarise(
    ss_crps = mean(ss_crps) * 100,
    ss_rmse = mean(ss_rmse) * 100,
    mase = mean(mase)
  )

方法二：移除依赖训练集的指标 #

如果不需要MASE/RMSSE，可从指标列表中删除这两项，直接用验证集计算剩余指标：

fc_models |>
  accuracy(
    data = validation_set,
    measures = list(
      crps = CRPS,
      rmse = RMSE,
      ss_crps = skill_score(CRPS),
      ss_rmse = skill_score(RMSE)
    ),
    n_quantiles = 100
  ) |>
  group_by(.model) %>%
  summarise(
    ss_crps = mean(ss_crps) * 100,
    ss_rmse = mean(ss_rmse) * 100
  )

额外提示：代码中tslm2和tslm3的公式完全一致，属于冗余模型，可删除其中一个以优化代码。

内容的提问来源于stack exchange，提问作者Mark_