Accuracy at Past Elections

Compares trends and polling firms with the results of the elections in x$elections: how far was each trend, and the last poll of each firm, from the result shortly before the election?

Usage

popAccuracy(
  x,
  trends = list(kalman = list(type = "kalman")),
  days = 1,
  firms = TRUE,
  window = 30
)

Arguments

x

A popPolls object.

trends

Named list of trends to evaluate, each a list of arguments for popAddTrend(): type, args, interpolations and houseEffects.

days

Days before the election that are evaluated.

firms

Whether the polling firms should be evaluated as well.

window

Only the polls of a firm published in this many days before the election are taken into account.

Value

A data.table with one row per election, trend or firm and party, and the columns election (date), kind ("trend" or "firm"), source (name of the trend or firm), party, estimate, result and error (estimate - result).

Details

For every election the trends are calculated from scratch, using only the polls published at least days before the election and the earlier elections. The comparison is therefore out of sample: even smoothed trends do not know the result, which makes it possible to compare trend types and their arguments, e.g. different values of sd.

Examples

if (FALSE) { # \dontrun{
de = popRead('DE-parliament')
accuracy = popAccuracy(de, trends = list(
    'sd 0.001' = list(type = 'kalman', args = list(sd = 0.001)),
    'sd 0.003' = list(type = 'kalman', args = list(sd = 0.003)),
    'mean 30d' = list(type = 'weightedMeanLastDays')
))
# mean absolute error, in percentage points
accuracy[, .(elections = uniqueN(election), mae = 100*mean(abs(error))),
         by = .(kind, source)][order(mae)]
} # }