Target Definition

Predictive models need something to predict. Extracted data is used to set (or confirm) the target definition, if not cast in stone. This is one of the most crucial steps of the process, which may not be straightforward. (1) Overview—i) continuous versus binary targets, and distinctions between probability of default (PD), exposure-at-default (EAD) and loss given default (LGD); ii) definition requirements—relevance, focus, transparency, adequacy and data quality; iii) performance components—automated counters/statuses and manual statuses; iv) code cross-checks—to determine whether statuses are properly understood. (2) Definition strictness—i) status nodes—define treatment {out-of-scope, exclusion, trivial balance, Bad/Indeterminate/Good}; ii) roll-rates—used to define delinquency levels; iii) trivial balances—avoid penalization of minor infractions; iv) closed accounts—possible treatments. (3) Integrity checks—i) consistency—period-on-period changes in distribution; ii) characteristics—ensuring they have the intended influence within the definition; iii) swap-set—assessing new versus old OR alternative definitions.