Caught in the Debt Trap

Abstract

New Zealand households entering the FinCap charitable-support network — households unable to meet routine debt obligations and seeking mentor-based financial intervention — disproportionately concentrate in two demographic intersections: younger working-age borrowers (25–34) and high-deprivation regions (NZ Deprivation Index 7–10). The third-tier-lender industry — non-bank, non-mainstream consumer-credit providers operating outside the prudential perimeter — has expanded its branch network in precisely these regions over the past decade.

The behavioural-finance construct most directly relevant is **present-biased borrowing under salience**: third-tier lenders advertise short-term financial relief with high salience and low immediate cost, which present-biased borrowers over-weight relative to the future debt-servicing burden. Standard household-finance models with time-consistent preferences predict that rational borrowers should not over-borrow at third-tier rates; the empirical evidence that such over-borrowing occurs at scale is what motivates the present-bias-plus-salience account.

The empirical setting is the FinCap NZ client caseload (155,482 records) — actual client-mentoring records linked to RBNZ household-debt indicators and Stats NZ regional-deprivation data. This is not a generic household survey; it is the lived caseload of households in active financial distress. The existing v6 draft applies the classical identification stack: OLS regression, difference-in-differences around regional third-tier-lender entry events, panel fixed-effects, probit estimation for binary debt-distress indicators, and Oaxaca-Blinder decomposition quantifying the high-vs-low-deprivation debt-distress gap. The existing evidence documents third-tier-lender concentration in high-deprivation regions and elevated debt-servicing distress among younger borrowers.

We propose extending the existing analysis with two ML layers: a geographic Causal Forest that would recover the conditional third-tier-lender-penetration effect by region × age band, and a planned NLP topic-model on FinCap caseworker case notes that would extract behavioural reasons-for-debt the structured survey items miss. The NLP layer is the methodological centrepiece: FinCap caseworkers record qualitative observations on each client's path-to-distress, and these notes contain present-bias-signal evidence (impulse-purchase language, deferred-future framing, salience-of-short-term-relief language) that structured questionnaires suppress. The findings sharpen the case for region-targeted prudential supervision of third-tier lending.

Data & Methods