tbl.now

• seasons count time from 1970-01-01 in the data's units, so a date gets the same sine and cosine values whatever the first date of the series. Fourier terms for report dates are still counted in the event units, as before. A numeric (non-date) axis has no calendar and still counts from its first event.

Column names are unchanged, but their values change, so models refitted with

these effects code them differently. A model's fit with Fourier seasons

alone is unaffected in principle -- shifting the phase rotates the sine and

cosine pair onto the same span -- while models with week_of_year or

month_of_year now use January / epiweek 1, rather than the first period in

the data, as the reference level. tbl_now_to_epinowcast() derives the

effects on its completed grid from the calendar too, so its values match

compute_temporal_effects() wherever the grid starts.

add_temporal_effects.data.frame() no longer needs numeric_col for seasons on

a date column (an explicit numeric_col still takes precedence) and gains

season_units, the units the Fourier periods are counted in.

nowcast_ensemble() can take the median (#100)

nowcast_ensemble(type = "median") combines the members' predictive quantiles

by their median, level by level, as the COVID-19 and FluSight forecast hubs'

quantile median ensembles do. One outlying member cannot drag it the way it

drags the average (type = "quantile"). The weights still apply: any weights

other than "equal", including "inverse_score" and "optim" from a

backtest, give a weighted median. With two equally weighted members the median

equals the average, so it needs three or more to differ.

epidist comes from CRAN

epidist in Suggests is now >= 1.0.0, its first CRAN release, so it

installs without an extra repository. tidy.epidist_fit() uses the interface

that epidist 0.5.0 introduced, delay_parameter_draws() and

add_summaries(), in place of the removed predict_delay_parameters() and

add_mean_sd(), which had made tidy() fail on every fit and R CMD check

warn about missing objects. Because the new draws carry every column of the

model data, all numeric, the reported terms are taken from the fit's family

(mu and sigma for a lognormal, say) plus mean and sd, rather than from

whatever numeric column is left after dropping a list of bookkeeping names.

baselinenowcast in Suggests is relaxed to >= 0.2.0, the version on CRAN.

diagnose() reports fractional delays under units / delay again

1.1.0 accidentally moved the fractional-.delay finding of diagnose() and

validate_tbl_now() from scope "delay" to "report". It is back under

"delay".