Forecast validation

Forecasts stored in Zoltar are validated upon upload based on the expected structure of each forecast. Below, we document the checks and tests that are performed on all forecasts. We first list the test that is performed for every prediction, and after that, tests are broken down by the class of prediction.

Definitions

For clarity, we define specific terms that we will use below.

  • Forecast: a collection of data specific to a project > model > timezero.
  • Prediction: a group of a prediction elements(s) specific to a location and target.
  • Prediction Element: data that define a unique single prediction, specific to the class of prediction it is.
  • Prediction Class: data structures representing different types of predictions, e.g. "Point" and "Bin" (see Data Model for more detail)
  • Target Type: the classification for a specific forecast target, one of "continuous", "discrete", "nominal", "binary" or "date" (see Targets for more info)
  • Database Row(s): the entry(ies)/row(s) in the database that comprise a prediction element.

Tests for all Prediction Elements

These tests are performed when a forecast is created or updated.

  • The Prediction's class must be valid for its target's type (see Valid prediction types by target type.
  • Within a Prediction, there cannot be more than 1 Prediction Element of the same class.

Tests for Prediction Elements by Prediction Class

These tests are performed when a forecast is created or updated.

Bin Prediction Elements

  • If a Bin Prediction Element exists, it should have >=1 Database Rows.
  • |cat| = |prob|. The number of elements in the cat and prob vectors should be identical.
  • cat (i, f, t, d, b): Entries in the database rows in the cat column cannot be “”, “NA” or NULL (case does not matter). Entries in cat must be a subset of Target.cats from the target definition.
  • prob (f): [0, 1]. Entries in the database rows in the prob column must be numbers in [0, 1]. For one prediction element, the values within prob must sum to 1.0 (values within +/- 0.001 of 1 are acceptable).
  • The data format of cat should correspond or be translatable to the type as in the target definition.
  • NB: Rows for Bin predictions where prob == 0 are not stored in the database.

Named Prediction Elements

We note that Named predictions currently only support fairly simple distributions. We currently support distributions with up to three parameters. Future versions of Zoltar could support distributions with larger numbers of parameters.

  • If a Named Prediction Element exists, it should have exactly 1 Database Row.
  • family: must be one of the abbreviations shown in the table below.
  • param1, param2, param3 (f): The number of param columns with non-NULL entries count must match family definition (see note below).
  • Parameters for each distribution must be within valid ranges, which, if constraints exist, are specified in the table below.

For reference, here is the mapping between the generic parameter names and the family-specific use of them (based on predx_classes.md):

Family abbreviation param1 param2 param3
Normal norm mean sd>=0 -
LogNormal lnorm mean sd>=0 -
Gamma gamma shape>0 rate>0 -
Beta beta a>0 b>0 -
Poisson pois rate>0 - -
Neg.Binom1 nbinom r>0 0<=p<=1 -
Neg.Binom2 nbinom2 mean>0 disp>0 -

Point Prediction Elements

  • If a Point Prediction Element exists, it should have exactly 1 Database Row for all targets.
  • value (i, f, t, d, b): Entries in the database rows in the value column cannot be “”, “NA” or NULL (case does not matter).
  • The data format of value should correspond or be translatable to the type as in the target definition.

Sample Prediction Elements

  • If a Sample Prediction Element exists, it should have >=1 Database Rows.
  • sample (i, f, t, d, b): Entries in the database rows in the sample column cannot be “”, “NA” or NULL (case does not matter).
  • The data format of sample should correspond or be translatable to the type as in the target definition.

Quantile Prediction Elements

  • If a Quantile Prediction Element exists, it should have >=1 Database Rows.
  • |quantile| = |value|. The number of elements in the quantile and value vectors should be identical.
  • quantile (f): [0, 1]. Entries in the database rows in the quantile column must be numbers in [0, 1]. quantiles must be unique.
  • value (i, f, d): Entries in value must be non-decreasing as quantiles increase. Entries in value must obey existing ranges for targets.
  • The data format of value should correspond or be translatable to the type as in the target definition.

Tests for Predictions by Target Type

These tests are performed when a forecast is created or updated.

"continuous"

  • Within one prediction, there can be at most one of the following prediction elements, but not both: {Named, Bin}.

"discrete"

  • Within one prediction, there can be at most one of the following prediction elements, but not both: {Named, Bin}.

Tests for Prediction Elements by Target Type

These tests are performed when a forecast is created or updated. For all target types, only valid Prediction Types are accepted.

"continuous"

  • any values in Point or Sample Prediction Elements should be numeric
  • if range is specified, any values in Point or Sample Prediction Elements should be contained within range
  • if range is specified, any Named Prediction Element should have negligible probability density (no more than 0.001 density) outside of the range.
  • for Bin Prediction Elements, the submitted set of cat values must be a subset of the cats defined by the target
  • for Named Prediction Elements, the distribution must be one of norm, lnorm, gamma, beta

"discrete"

  • any values in Point or Sample Prediction Elements should be integers
  • if range is specified, any values in Point or Sample Prediction Elements should be contained within range
  • if range is specified, any Named Prediction Element should have negligible probability density (no more than 0.001 density) outside of the range
  • for Bin Prediction Elements, the submitted set of cat values must be a subset of the cats defined by the target
  • for Named Prediction Elements, the distribution must be one of pois, nbinom, nbinom2.

"nominal"

  • any values in Point or Sample Prediction Elements should be contained within the valid set of cats defined by the target
  • for Bin Prediction Elements, the submitted set of cat values must be a subset of the cats defined by the target

"binary"

  • any values in Point or Sample Prediction Elements should be either true or false.
  • for Bin Prediction Elements, there must be exactly two cat values labeled true and false. These are the two cats that are implied (but not allowed to be specified) by binary target types.

"date"

  • any values in Point or Sample Prediction Elements should be string that can be interpreted as a date in YYYY-MM-DD format, and these values should be contained within the set of valid responses defined by cats defined by the target.
  • for Bin Prediction Elements, the submitted set of cats must be a subset of the valid outcomes defined by the target range.

Tests for target definitions by Target Type

These tests are performed when a target is created or updated.

"continuous"

  • if both range and cats are specified, then the min(cats) must equal the lower bound and max(cats) must be less than the upper bound.
  • if range is specified it must contain two numeric values
  • range lower bound must be smaller than the upper bound.
  • if range is specified, is assumed to be inclusive on the lower bound and open on the upper bound, e.g. [a, b).

"discrete"

  • if range is specified, it must include two integers.

"nominal"

  • cats must be a character vector containing a set of unique labels of the categories for this target. The labels must not include "", NA or NULL (case does not matter).

"binary"

  • none.

"date"

  • the unit parameter is required to be one of month, week, biweek, or day
  • the date parameter must contain a list of text elements in YYYY-MM-DD format that can be interpreted as dates.

Tests for ground truth data tables

Please see zoltar-ground-truth-example.csv for an example of a valid specification of ground-truth values.

For all ground truth files

  • The columns are timezero, location, target, and value.
  • For every unique target-location-timezero combination, there should be either one or zero rows of truth data.
  • Every value of timezero, target and location must be in the list of valid values defined by the project configuration file. (Note: not every combination needs to exist for the file to be valid.)
  • The value of the truth data cannot be “”, “NA” or NULL (case does not matter).
  • The value of the truth data should be interpretable as the corresponding data_type of the specified target. E.g., for a row corresponding to a date target, the entry must contain a valid ISO-formatted date string.

Range-check for ground truth data

The following test can be applied to any target with a range. This will always apply to binary, nominal, and date targets, as these targets are required to have sets of valid values specified as part of the target definition. If the range parameter is specified for a continuous or discrete target, then the following test will be applied to that target as well.

For binary targets: - The entry in the value column for a specific target-location-timezero combination must be either true or false.

For discrete and continuous targets (if range is specified): - The entry in the value column for a specific target-location-timezero combination must be contained within the range of valid values for the target. If cats is specified but range is not, then there is an implicit range for the ground truth value, and that is between min(cats) and \infty.

For nominal and date target_types: - The entry in the value column for a specific target-location-timezero combination must be contained within the set of valid cats for the target, as defined by the project config file.