NEWS.md

In epiforecasts/scoringutils: Utilities for Scoring and Assessing Predictions

scoringutils 1.2.2.9000

This update represents a major rewrite of the package and introduces breaking changes. If you want to keep using the older version, you can download it using remotes::install_github("epiforecasts/scoringutils@v1.2").

The update aims to make the package more modular and customisable and overall cleaner and easier to work with. In particular, we aimed to make the suggested workflows for evaluating forecasts more explicit and easier to follow (see visualisation below). To do that, we clarified input formats and made them consistent across all functions. We refactord many functions to S3-methods and introduced forecast objects with separate classes for different types of forecasts. A new set of as_forecast_<type>() functions was introduced to validate the data and convert inputs into a forecast object (a data.table with a forecast class and an additional class corresponding to the forecast type (see below)). Another major update is the possibility for users to pass in their own scoring functions into score(). We updated and improved all function documentation and added new vignettes to guide users through the package. Internally, we refactored the code, improved input checks, updated notifications (which now use the cli package) and increased test coverage.

The most comprehensive documentation for the new package after the rewrite is the revised version of our original scoringutils paper.

Package updates

score()

• The main function of the package is still the function score(). However, we reworked the function and updated and clarified its input requirements.
• The previous columns "true_value" and "prediction" were renamed. score() now requires columns called "observed" and "predicted" and "model". The column quantile was renamed to quantile_level and sample was renamed to sample_id
• score() is now a generic. It has S3 methods for the classes forecast_point, forecast_binary, forecast_quantile, forecast_sample, and forecast_nominal, which correspond to the different forecast types that can be scored with scoringutils.
• score() now calls na.omit() on the data, instead of only removing rows with missing values in the columns observed and predicted. This is because NA values in other columns can also mess up e.g. grouping of forecasts according to the unit of a single forecast.
• score() and many other functions now require a validated forecast object. forecast objects can be created using the functions as_forecast_point(), as_forecast_binary(), as_forecast_quantile(), and as_forecast_sample() (which replace the previous check_forecast()). A forecast object is a data.table with class forecast and an additional class corresponding to the forecast type (e.g. forecast_quantile). score() now returns objects of class scores with a stored attribute metrics that holds the names of the scoring rules that were used. Users can call get_metrics() to access the names of those scoring rules.
• score() now returns one score per forecast, instead of one score per sample or quantile. For binary and point forecasts, the columns "observed" and "predicted" are now removed for consistency with the other forecast types.
• Users can now also use their own scoring rules (making use of the metrics argument, which takes in a named list of functions). Default scoring rules can be accessed using the functions metrics_point(), metrics_sample(), metrics_quantile(), metrics_binary(), and metrics_nominal(), which return a named list of scoring rules suitable for the respective forecast type. Column names of scores in the output of score() correspond to the names of the scoring rules (i.e. the names of the functions in the list of metrics).
• Instead of supplying arguments to score() to manipulate individual scoring rules users should now manipulate the metric list being supplied using purrr::partial() and select_metric(). See ?score() for more information.
• the CRPS is now reported as decomposition into dispersion, overprediction and underprediction.

Creating a forecast object

• The as_forecast_<type>() functions create a forecast object and validates it. They also allow users to rename/specify required columns and specify the forecast unit in a single step, taking over the functionality of set_forecast_unit() in most cases. See ?as_forecast() for more information.
• Some as_forecast_<type>() functions like e.g. as_forecast_point() and as_forecast_quantile() have S3 methods for converting from another forecast type to the respective forecast type. For example, as_forecast_quantile() has a method for converting from a forecast_sample object to a forecast_quantile object by estimating quantiles from the samples.

Updated workflows

• An example workflow for scoring a forecast now looks like this: forecast_quantile <- as_forecast_quantile( example_quantile, observed = "observed", predicted = "predicted", model = "model", quantile_level = "quantile_level", forecast_unit = c("model", "location", "target_end_date", "forecast_date", "target_type") ) scores <- score(forecast_quantile)
• Overall, we updated the suggested workflows for how users should work with the package. The following gives an overview (see the updated paper for more details).

Input formats

• We standardised input formats both for score() as well as for the scoring rules exported by scoreingutils. The following plot gives a overview of the expected input formats for the different forecast types in score().

• Support for the interval format was mostly dropped (see PR #525 by @nikosbosse and reviewed by @seabbs). The co-existence of the quantile and interval format let to a confusing user experience with many duplicated functions providing the same functionality. We decided to simplify the interface by focusing exclusively on the quantile format.

  • The function bias_range() was removed (users should now use bias_quantile() instead)
  • The function interval_score() was made an internal function rather than being exported to users. We recommend using wis() instead.

(Re-)Validating forecast objects

• To create and validate a new forecast object, users can use as_forecast_<type>(). To revalidate an existing forecast object users can call assert_forecast() (which validates the input and returns invisible(NULL). assert_forecast() is a generic with methods for the different forecast types. Alternatively, validate_forecast() can be used (which calls assert_forecast()), which returns the input and is useful in a pipe. Lastly, users can simply print the object to obtain additional information.
• Users can test whether an object is of class forecast_*() using the function is_forecast(). Users can also test for a specific forecast_* class using the appropriate is_forecast.forecast_* method. For example, to check whether an object is of class forecast_quantile, you would use you would use scoringutils:::is_forecast.forecast_quantile().

Pairwise comparisons and relative skill

• The functionality for computing pairwise comparisons was now split from summarise_scores(). Instead of doing pairwise comparisons as part of summarising scores, a new function, add_relative_skill(), was introduced that takes summarised scores as an input and adds columns with relative skill scores and scaled relative skill scores.
• The function pairwise_comparison() was renamed to get_pairwise_comparisons(), in line with other get_-functions. Analogously, plot_pairwise_comparison() was renamed to plot_pairwise_comparisons().
• Output columns for pairwise comparisons have been renamed to contain the name of the metric used for comparing.
• Replaced warnings with errors in get_pairwise_comparison to avoid returning NULL

Computing coverage values

• add_coverage() was replaced by a new function, get_coverage(). This function comes with an updated workflow where coverage values are computed directly based on the original data and can then be visualised using plot_interval_coverage() or plot_quantile_coverage(). An example workflow would be example_quantile |> as_forecast_quantile() |> get_coverage(by = "model") |> plot_interval_coverage().

Obtaining and plotting forecast counts

• The function avail_forecasts() was renamed to get_forecast_counts(). This represents a change in the naming convention where we aim to name functions that provide the user with additional useful information about the data with a prefix "get_". Sees Issue #403 and #521 and PR #511 by @nikosbosse and reviewed by @seabbs for details.
• For clarity, the output column in get_forecast_counts() was renamed from "Number forecasts" to "count".
• get_forecast_counts() now also displays combinations where there are 0 forecasts, instead of silently dropping corresponding rows.
• plot_avail_forecasts() was renamed plot_forecast_counts() in line with the change in the function name. The x argument no longer has a default value, as the value will depend on the data provided by the user.

Renamed functions

• The function find_duplicates() was renamed to get_duplicate_forecasts().
• Renamed interval_coverage_quantile() and interval_coverage_dev_quantile() to interval_coverage() and interval_coverage_deviation(), respectively.
• "range" was consistently renamed to "interval_range" in the code. The "range"-format (which was mostly used internally) was renamed to "interval"-format
• Renamed correlation() to get_correlations() and plot_correlation() to plot_correlations()
• pit() was renamed to get_pit().

Deleted functions

• Removed abs_error and squared_error from the package in favour of Metrics::ae and Metrics::se.get_duplicate_forecasts() now sorts outputs according to the forecast unit, making it easier to spot duplicates. In addition, there is a counts option that allows the user to display the number of duplicates for each forecast unit, rather than the raw duplicated rows.
• Deleted the function plot_ranges(). If you want to continue using the functionality, you can find the function code here or in the Deprecated-visualisations Vignette.
• Removed the function plot_predictions(), as well as its helper function make_NA(), in favour of a dedicated Vignette that shows different ways of visualising predictions. For future reference, the function code can be found here (Issue #659) or in the Deprecated-visualisations Vignette.
• Removed the function plot_score_table(). You can find the code in the Deprecated-visualisations Vignette.
• Removed the function merge_pred_and_obs() that was used to merge two separate data frames with forecasts and observations. We moved its contents to a new "Deprecated functions"-vignette.
• Removed interval_coverage_sample() as users are now expected to convert to a quantile format first before scoring.
• Function set_forecast_unit() was deleted. Instead there is now a forecast_unit argument in as_forecast_<type>() as well as in get_duplicate_forecasts().

Function changes

• bias_quantile() changed the way it handles forecasts where the median is missing: The median is now imputed by linear interpolation between the innermost quantiles. Previously, we imputed the median by simply taking the mean of the innermost quantiles.
• In contrast to the previous correlation function, get_correlations doesn't round correlations by default. Instead, plot_correlations now has a digits argument that allows users to round correlations before plotting them. Alternatively, using dplyr, you could call something like mutate(correlations, across(where(is.numeric), \(x) signif(x, digits = 2))) on the output of get_correlations.

Internal package updates

• The deprecated ..density.. was replaced with after_stat(density) in ggplot calls.
• Files ending in ".Rda" were renamed to ".rds" where appropriate when used together with saveRDS() or readRDS().
• Added a subsetting [ operator for scores, so that the score name attribute gets preserved when subsetting.
• Switched to using cli for condition handling and signalling, and added tests for all the check_*() and test_*() functions. See #583 by @jamesmbaazam and reviewed by @nikosbosse and @seabbs.
• scoringutils now requires R >= 4.0

Documentation and testing

• Updates documentation for most functions and made sure all functions have documented return statements
• Documentation pkgdown pages are now created both for the stable and dev versions.
• Added unit tests for many functions

scoringutils 1.2.2

Package updates

• scoringutils now depends on R 3.6. The change was made since packages testthat and lifecycle, which are used in scoringutils now require R 3.6. We also updated the Github action CI check to work with R 3.6 now.
• Added a new PR template with a checklist of things to be included in PRs to facilitate the development and review process

Bug fixes

• Fixes a bug with set_forecast_unit() where the function only worked with a data.table, but not a data.frame as an input.
• The metrics table in the vignette Details on the metrics implemented in scoringutils had duplicated entries. This was fixed by removing the duplicated rows.

scoringutils 1.2.1

Package updates

• This minor update fixes a few issues related to gh actions and the vignettes displayed at epiforecasts.io/scoringutils. It
• Gets rid of the preferably package in _pkgdown.yml. The theme had a toggle between light and dark theme that didn't work properly
• Updates the gh pages deploy action to v4 and also cleans up files when triggered
• Introduces a gh action to automatically render the Readme from Readme.Rmd
• Removes links to vignettes that have been renamed

scoringutils 1.2.0

This major release contains a range of new features and bug fixes that have been introduced in minor releases since 1.1.0. The most important changes are:

• Documentation updated to reflect changes since version 1.1.0, including new transform and workflow functions.
• New set_forecast_unit() function allows manual setting of forecast unit.
• summarise_scores() gains new across argument for summarizing across variables.
• New transform_forecasts() and log_shift() functions allow forecast transformations. See the documentation for transform_forecasts() for more details and an example use case.
• Input checks and test coverage improved for bias functions.
• Bug fix in get_prediction_type() for integer matrix input.
• Links to scoringutils paper and citation updates.
• Warning added in interval_score() for small interval ranges.
• Linting updates and improvements.

Thanks to @nikosbosse, @seabbs, and @sbfnk for code and review contributions. Thanks to @bisaloo for the suggestion to use a linting GitHub Action that only triggers on changes, and @adrian-lison for the suggestion to add a warning to interval_score() if the interval range is between 0 and 1.

Package updates

• The documentation was updated to reflect the recent changes since scoringutils 1.1.0. In particular, usage of the functions set_forecast_unit(), check_forecasts() and transform_forecasts() are now documented in the Vignettes. The introduction of these functions enhances the overall workflow and help to make the code more readable. All functions are designed to be used together with the pipe operator. For example, one can now use something like the following:

example_quantile |> 
  set_forecast_unit(c("model", "location", "forecast_date", "horizon", "target_type")) |> 
  check_forecasts() |> 
  score()

Documentation for the transform_forecasts() has also been extended. This functions allows the user to easily add transformations of forecasts, as suggested in the paper "Scoring epidemiological forecasts on transformed scales". In an epidemiological context, for example, it may make sense to apply the natural logarithm first before scoring forecasts, in order to obtain scores that reflect how well models are able to predict exponential growth rates, rather than absolute values. Users can now do something like the following to score a transformed version of the data in addition to the original one:

data <- example_quantile[true_value > 0, ]
data |>
  transform_forecasts(fun = log_shift, offset = 1) |> 
  score() |> 
  summarise_scores(by = c("model", "scale"))

Here we use the log_shift() function to apply a logarithmic transformation to the forecasts. This function was introduced in scoringutils 1.1.2 as a helper function that acts just like log(), but has an additional argument offset that can add a number to every prediction and observed value before applying the log transformation.

Feature updates

• Made check_forecasts() and score() pipeable (see issue #290). This means that users can now directly use the output of check_forecasts() as input for score(). As score() otherwise runs check_forecasts() internally anyway this simply makes the step explicit and helps writing clearer code.

scoringutils 1.1.7

Release by @seabbs in #305. Reviewed by @nikosbosse and @sbfnk.

Breaking changes

• The prediction_type argument of get_forecast_unit() has been changed dropped. Instead a new internal function prediction_is_quantile() is used to detect if a quantile variable is present. Whilst this is an internal function it may impact some users as it is accessible via `find_duplicates().

Package updates

• Made imputation of the median in bias_range() and bias_quantile() more obvious to the user as this may cause unexpected behaviour.
• Simplified bias_range() so that it uses bias_quantile() internally.
• Added additional input checks to bias_range(), bias_quantile(), and check_predictions() to make sure that the input is valid.
• Improve the coverage of unit tests for bias_range(), bias_quantile(), and bias_sample().
• Updated pairwise comparison unit tests to use more realistic data.

Bug fixes

• Fixed a bug in get_prediction_type() which led to it being unable to correctly detect integer (instead categorising them as continuous) forecasts when the input was a matrix. This issue impacted bias_sample() and also score() when used with integer forecasts resulting in lower bias scores than expected.

scoringutils 1.1.6

Feature updates

• Added a new argument, across, to summarise_scores(). This argument allows the user to summarise scores across different forecast units as an alternative to specifying by. See the documentation for summarise_scores() for more details and an example use case.

scoringutils 1.1.5

Feature updates

• Added a new function, set_forecast_unit() that allows the user to set the forecast unit manually. The function removes all columns that are not relevant for uniquely identifying a single forecast. If not done manually, scoringutils attempts to determine the unit of a single automatically by simply assuming that all column names are relevant to determine the forecast unit. This can lead to unexpected behaviour, so setting the forecast unit explicitly can help make the code easier to debug and easier to read (see issue #268). When used as part of a workflow, set_forecast_unit() can be directly piped into check_forecasts() to check everything is in order.

scoringutils 1.1.4

Package updates

• Added links to the scoringutils paper Evaluating Forecasts with scoringutils in R to the package.
• Updated citation formatting to comply with newer standards.

scoringutils 1.1.3

Package updates

• Added a warning to interval_score() if the interval range is between 0 and 1. Thanks to @adrian-lison (see #277) for the suggestion.

Package updates

• Switched to a linting GitHub Action that only triggers on changes. Inspired by @bisaloo recent contribution to the epinowcast package.
• Updated package linters to be more extensive. Inspired by @bisaloo recent contribution to the epinowcast package.
• Resolved all flagged linting issues across the package.

scoringutils 1.1.2

Feature updates

• Added a new function, transform_forecasts() to make it easy to transform forecasts before scoring them, as suggested in Bosse et al. (2023), https://www.medrxiv.org/content/10.1101/2023.01.23.23284722v1.
• Added a function, log_shift() that implements the default transformation function. The function allows to add an offset before applying the logarithm.

scoringutils 1.1.1

• Added a small change to interval_score() which explicitly converts the logical vector to a numeric one. This should happen implicitly anyway, but is now done explicitly in order to avoid issues that may come up if the input vector has a type that doesn't allow the implicit conversion.

scoringutils 1.1.0

A minor update to the package with some bug fixes and minor changes.

Feature updates

Package updates

• Removed the on attach message which warned of breaking changes in 1.0.0.
• Renamed the metric argument of summarise_scores() to relative_skill_metric. This argument is now deprecated and will be removed in a future version of the package. Please use the new argument instead.
• Updated the documentation for score() and related functions to make the soft requirement for a model column in the input data more explicit.
• Updated the documentation for score(), pairwise_comparison() and summarise_scores() to make it clearer what the unit of a single forecast is that is required for computations
• Simplified the function plot_pairwise_comparison() which now only supports plotting mean score ratios or p-values and removed the hybrid option to print both at the same time.

Bug fixes

• Missing baseline forecasts in pairwise_comparison() now trigger an explicit and informative error message.
• The requirements table in the getting started vignette is now correct.
• Added support for an optional sample column when using a quantile forecast format. Previously this resulted in an error.

scoringutils 1.0.0

Major update to the package and most package functions with lots of breaking changes.

Feature updates

• New and updated Readme and vignette.
• The proposed scoring workflow was reworked. Functions were changed so they can easily be piped and have simplified arguments and outputs.

New functions and function changes

• The function eval_forecasts() was replaced by a function score() with a much reduced set of function arguments.
• Functionality to summarise scores and to add relative skill scores was moved to a function summarise_scores()
• New function check_forecasts() to analyse input data before scoring
• New function correlation() to compute correlations between different metrics
• New function add_coverage() to add coverage for specific central prediction intervals.
• New function avail_forecasts() allows to visualise the number of available forecasts.
• New function find_duplicates() to find duplicate forecasts which cause an error.
• All plotting functions were renamed to begin with plot_. Arguments were simplified.
• The function pit() now works based on data.frames. The old pit function was renamed to pit_sample(). PIT p-values were removed entirely.
• The function plot_pit() now works directly with input as produced by pit()
• Many data-handling functions were removed and input types for score() were restricted to sample-based, quantile-based or binary forecasts.
• The function brier_score() now returns all brier scores, rather than taking the mean before returning an output.
• crps(), dss() and logs() were renamed to crps_sample(), dss_sample(), and logs_sample()

Bug fixes

• Testing was expanded
• Minor bugs were fixed, for example a bug in the as_forecast_quantile() function (https://github.com/epiforecasts/scoringutils/pull/223)

Package data updated

• Package data is now based on forecasts submitted to the European Forecast Hub (https://covid19forecasthub.eu/).
• All example data files were renamed to begin with example_.
• A new data set, summary_metrics was included that contains a summary of the metrics implemented in scoringutils.

Other breaking changes

• The 'sharpness' component of the weighted interval score was renamed to dispersion. This was done to make it more clear what the component represents and to maintain consistency with what is used in other places.

scoringutils 0.1.8

Feature updates

• Added a function check_forecasts() that runs some basic checks on the input data and provides feedback.

scoringutils 0.1.7.2

Package updates

• Minor bug fixes (previously, 'interval_score' needed to be among the selected metrics).
• All data.tables are now returned as table[] rather than as table, such that they don't have to be called twice to display the contents.

scoringutils 0.1.7

Feature updates

• Added a function, pairwise_comparison() that runs pairwise comparisons between models on the output of eval_forecasts()
• Added functionality to compute relative skill within eval_forecasts().
• Added a function to visualise pairwise comparisons.

Package updates

• The WIS definition change introduced in version 0.1.5 was partly corrected such that the difference in weighting is only introduced when summarising over scores from different interval ranges.
• "sharpness" was renamed to 'mad' in the output of [score()] for sample-based forecasts.

scoringutils 0.1.

Feature updates

• eval_forecasts() can now handle a separate forecast and truth data set as as input.
• eval_forecasts() now supports scoring point forecasts along side quantiles in a quantile-based format. Currently the only metric used is the absolute error.

Package updates

• Many functions, especially eval_forecasts() got a major rewrite. While functionality should be unchanged, the code should now be easier to maintain
• Some of the data-handling functions got renamed, but old names are supported as well for now.

scoringutils 0.1.5

Package updates

• Changed the default definition of the weighted interval score. Previously, the median prediction was counted twice, but is no only counted once. If you want to go back to the old behaviour, you can call the interval_score function with the argument count_median_twice = FALSE.

scoringutils 0.1.4

Feature updates

• Added basic plotting functionality to visualise scores. You can now easily obtain diagnostic plots based on scores as produced by score.
• correlation_plot() shows correlation between metrics.
• plot_ranges() shows contribution of different prediction intervals to some chosen metric.
• plot_heatmap() visualises scores as heatmap.
• plot_score_table() shows a coloured summary table of scores.

package updates

• Renamed "calibration" to "coverage".
• Renamed "true_values" to "true_value" in data.frames.
• Renamed "predictions" to "prediction" in data.frames.
• Renamed "is_overprediction" to "overprediction".
• Renamed "is_underprediction" to "underprediction".

scoringutils 0.1.3

(Potentially) Breaking changes

• The by argument in score now has a slightly changed meaning. It now denotes the lowest possible grouping unit, i.e. the unit of one observation and needs to be specified explicitly. The default is now NULL. The reason for this change is that most metrics need scoring on the observation level and this the most consistent implementation of this principle. The pit function receives its grouping now from summarise_by. In a similar spirit, summarise_by has to be specified explicitly and e.g. doesn't assume anymore that you want 'range' to be included.
• For the interval score, weigh = TRUE is now the default option.
• Renamed true_values to true_value and predictions to prediction.

Feature updates

• Updated quantile evaluation metrics in score. Bias as well as calibration now take all quantiles into account.
• Included option to summarise scores according to a summarise_by argument in score() The summary can return the mean, the standard deviation as well as an arbitrary set of quantiles.
• score() can now return pit histograms.
• Switched to ggplot2 for plotting.

scoringutils 0.1.2

(Potentially) Breaking changes

• All scores in score were consistently renamed to lower case. Interval_score is now interval_score, CRPS is now crps etc.

Feature updates

• Included support for grouping scores according to a vector of column names in score().
• Included support for passing down arguments to lower-level functions in score()
• Included support for three new metrics to score quantiles with score(): bias, sharpness and calibration

Package updates

• Example data now has a horizon column to illustrate the use of grouping.
• Documentation updated to explain the above listed changes.

scoringutils 0.1.1

Feature updates

• Included support for a long as well as wide input formats for quantile forecasts that are scored with score().

Package updates

• Updated documentation for the score().
• Added badges to the README.

epiforecasts/scoringutils documentation built on Sept. 16, 2024, 5:20 a.m.