Ytukyg

I have followed the instructions in this StMoMo package document to fit Lee Carter to mortality data for Canada.

The next step in my project is to measure the forecast accuracy of the Lee Carter model when fit to this Canadian data.

To do this, I tried to use accuracy() but ran into an error since my Lee Carter fit is of class "fitStMoMo" and not of class "forecast" or a time series.

Is there an alternative forecast accuracy function that I can use on "fitStMoMo" objects that will calculate Mean Error, Root Mean Squared Error, Mean Absolute Error, Mean Percentage Error, Mean Absolute Percentage Error and Mean Absolute Scaled Error for me?

Reprex

Reprex created using EWMaleData as used in the StMoMo document to flag the error specifically:

library("StMoMo")

library("demography")

library("forecast")



constLC <- function(ax, bx, kt, b0x, gc, wxt, ages){

    c1 <- mean(kt[1, ], na.rm = TRUE)

    c2 <- sum(bx[, 1], na.rm = TRUE)

    list(ax = ax + c1 * bx, bx = bx / c2, kt = c2 * (kt - c1))

}

LC <- StMoMo(link = "logit", staticAgeFun = TRUE, periodAgeFun = "NP",

constFun = constLC)

LC <- lc(link = "logit")

LC$gnmFormula

#> [1] "D/E ~ -1 + offset(o) + factor(x) + Mult(factor(x), factor(t), inst = 1)"



EWMaleData

#> Mortality data for England and Wales

#>     Series:  male

#>     Years: 1961 - 2011

#>     Ages:  0 - 100

#>     Exposure:  central



EWMaleIniData <- central2initial(EWMaleData)

ages.fit <- 55:89

wxt <- genWeightMat(ages = ages.fit, years = EWMaleIniData$years,

clip = 3)

LCfit <- fit(LC, data = EWMaleIniData, ages.fit = ages.fit, wxt = wxt)

#> StMoMo: The following cohorts have been zero weigthed: 

#>   1872 1873 1874 1954 1955 1956 

#> StMoMo: Start fitting with gnm

#> Initialising

#> Running start-up iterations..

#> Running main iterations.....

#> Done

#> StMoMo: Finish fitting with gnm



LCfor <- forecast(LCfit, h = 50)

class(LCfit)

#> [1] "fitStMoMo"

class(LCfor)

#> [1] "forStMoMo"

accuracy(LCfit)

#> Error in accuracy.default(LCfit): First argument should be a forecast object 

#>   or a time series.

accuracy(LCfor)

#> Error in accuracy.default(LCfor): First argument should be a forecast object

#>   or a time series.

I'm not entirely sure about how accuracy() from forecast works, but in some way it must be comparing real and predicted values, and returning metrics on how much they differ. This can, in a broad sense, be regarded as a form of cross-validation. As accuracy() doesn't work for StMoMo objects, we might as well develop a cross-validation routine ourself.

For a short primer on this form of cross-validation, I'd recommend Rob Hyndman's notes on tsCV() from forecast. It would have been nice if we could use tsCV() here, but it only works for univariate time series, and mortality data is essentially multivariate time series.

I should also mention that before today I had never heard about Mortality Modeling, so the model theory part of this I'm very fuzzy on.

This first bit is identical to what was already posted

library(StMoMo)

library(demography)

library(forecast)



data(EWMaleData)



constLC <- function(ax, bx, kt, b0x, gc, wxt, ages){

    c1 <- mean(kt[1, ], na.rm = TRUE)

    c2 <- sum(bx[, 1], na.rm = TRUE)

    list(ax = ax + c1 * bx, bx = bx / c2, kt = c2 * (kt - c1))

}



LC <- StMoMo(link="logit", staticAgeFun=TRUE, periodAgeFun="NP", constFun=constLC)

LC <- lc(link="logit")



EWMaleIniData <- central2initial(EWMaleData)

Then things get a little different. The central point of performing CV on a time series is to do a prediction on data we actually have, but we pretend we don't. Therefore we'll have to subset our data so that the chunk we want to predict on isn't part of the model. In this concrete example we'll use the first 30 years, and then predict the next 10

ages.fit <- 55:89

years.fit <- EWMaleIniData$years[1]:(EWMaleIniData$years[1] + 30)

years.for <- 10



wxt <- genWeightMat(ages=ages.fit, years=years.fit, clip=3)

LCfit <- fit(LC, data=EWMaleIniData, ages.fit=ages.fit,

  years.fit=years.fit, wxt=wxt)

LCfor <- forecast(LCfit, h=years.for)

Now that we have a ten-year forecast, we can compare those years to our actual data and use what ever error measure we want to see how accurate the prediction is.

cvy <- LCfor$years  # years used in forecast

cva <- LCfor$ages   # ages used in forecast



pred <- LCfor$rates # predicted mortality rates



# actual mortality rates subset to the same ages and years as forecast

actual <- EWMaleIniData$Dxt/EWMaleIniData$Ext

actual <- actual[rownames(actual) %in% cva,

                 colnames(actual) %in% cvy]



# A collection of error measures. plenty of others can be devised

err <- pred - actual

Q <- pred/actual

rmse <- sqrt(rowMeans(err^2))

mae <- rowMeans(abs(err))

smape <- 100 * (rowMeans(exp(abs(log(Q)))) - 1)

This bit is purely for displaying the results

par(mfrow=c(3, 1), mar=c(3, 3, 1, 1), mgp=c(2, 0.8, 0), oma=c(0, 0, 2, 0))

plot(as.numeric(names(rmse)), rmse, type="h", xlab="")

plot(as.numeric(names(mae)), mae, type="h", xlab="")

plot(as.numeric(names(smape)), smape, type="h", xlab="Ages")

mtext(paste("Forecast accuracy for the years", 

  paste(cvy[c(1, years.for)], collapse=" - ")), 

  3, outer=TRUE)

As seen in Hyndman's notes, to do this properly we'd have to do this comparison using forecasts at several points in our time series, and the average the scores.