Nowcasting with Chain-Ladder Method.

nowcastr

R package for nowcasting with non-cumulative chain-ladder method.

• 1 main nowcast function
  • nowcast_cl() returns object with all intermediary results
• 4 plots
  • plot_nc_input(option = "triangle") / plot(which = "data", option = "triangle")
  • plot_nc_input(option = "millipede") / plot(which = "data", option = "millipede")
  • plot_delays() / plot(which = "delays")
  • plot_nowcast() / plot(which = "results")
• 3 utility functions
  • calculate_retro_score(): Calculate retro-scores for all groups
  • rm_repeated_values(): Remove duplicated reported values in reporting matrix
  • fill_future_reported_values(): Fill future reported values with last known values
• Accuracy Evaluation
  • nowcast_eval(): perform evaluation
  • plot_nowcast_eval(): plot main eval results
  • plot_nowcast_eval_by_delay(): plot eval results by delay
  • plot_nowcast_eval_detail(): plot detailed eval results

Installation

## install from GitHub
pak::pak("whocov/nowcastr") # recommended, more up to date versions

## install from CRAN
install.packages("nowcastr")

Quick Start

library(nowcastr)

## Get your data
nc_data <- nowcast_demo

## Plot input data
nc_data %>%
  plot_nc_input(
    option = "triangle", # or "millipede"
    col_date_occurrence = date_occurrence,
    col_date_reporting = date_report,
    col_value = value,
    group_cols = "group"
  )

## Run nowcast with built-in demo data
nc_obj <- nc_data %>% 
  nowcast_cl(
    max_delay = 5, # optional
    max_reportunits = 8, # optional
    col_date_occurrence = date_occurrence,
    col_date_reporting = date_report,
    col_value = value,
    group_cols = "group",
    time_units = "weeks",
    do_model_fitting = TRUE
  )

## Plot nowcasted time series
plot(nc_obj, which = "results")
print(nc_obj@results) # inspect data frame

## Plot delay distribution
plot(nc_obj, which = "delays")
print(nc_obj@delays) # inspect data frame

More detailed examples are available in the Getting started vignette.

Data Requirements

Dataset with at least 2 date columns and a value column. The dataset can also have multiple group-by columns for batch processing.

Note that the delays (difference between the 2 dates) should have constant intervals, i.e., multiples of 1 day or 7 days.

dplyr::glimpse(nowcast_demo, 70)

# Rows: 1,624
# Columns: 4
# $ value           <dbl> 251563, 219818, 219815, 253451, 253454, 3116…
# $ date_occurrence <date> 2024-12-16, 2024-12-23, 2024-12-23, 2024-12…
# $ date_report     <date> 2025-05-26, 2025-05-26, 2025-06-02, 2025-05…
# $ group           <chr> "Syndromic ARI", "Syndromic ARI", "Syndromic…

Output Object

nowcast_cl() returns an S7 object of class nowcast_results with the following slots (access with @):

Methods Summary

1. Input Data: Ensure three core columns: `observed_value` / `date_of_reporting` / `date_of_occurrence` (e.g. date_of_event / date_of_onset)
2. Calculate the `reporting_delay` (= `date_of_reporting` - `date_of_occurrence`)
3. Compute the `completeness` (= `observed_value` / `true_value` (approximated by `last_reported_value`))
4. Aggregate the `avg_completeness` for each `reporting_delay`
5. Optional: Fit a curve through that
6. Apply Nowcast: `nowcast` = `observed_value` / `avg_completeness`