RDP 2015-01: Stress Testing the Australian Household Sector Using the HILDA Survey 6. Limitations and Future Work

As with all stress-testing models, the model described in this paper has some limitations that are critical to its interpretation. In addition to the limitations already discussed above, some other notable limitations are:

  • The one-period nature of the model means that the impact on ADIs cannot be compared with that predicted by APRA in its stress tests using scenarios with a time dimension. For example, a 5 percentage point increase in the unemployment rate in the model means that 5 per cent of individuals in the labour force (on top of those already unemployed) instantly become unemployed. Within this extra 5 per cent, any household whose financial margin falls below zero is assumed to default instantly. By contrast, in a real-world downturn involving a multi-year period of high unemployment, a certain proportion of the individuals that become unemployed would find jobs prior to defaulting. Additionally, loan losses would be spread over time rather than occurring instantaneously.
  • Household surveys may not adequately identify households with negative financial margins (for instance, because households tend to understate their debt and income).[16] In addition, although efforts are made to ensure that the HILDA Survey sample is representative, households with precarious finances often do not disclose their financial position, while higher-income households are less likely to remain in the survey over time. Furthermore, household surveys such as this are generally only available around 12 to 15 months after fieldwork has been completed, reducing their usefulness as a real-time stress-testing tool.
  • The predictive ability of household micro-simulations is relatively untested. While these models have been established in a number of countries, none of these countries have had recent crises that emanated from the household sector. The US household sector could be a useful case study to test this. Household surveys, such as the Federal Reserve's Survey of Consumer Finances, contain many of the required variables to run such an experiment.

A number of adjustments could be made to potentially improve the model:

  • Asset price variability: asset prices are currently assumed to fall by a set percentage for all households. However, it might be more realistic for asset prices to fall by differing amounts for each household; for example, housing prices could fall based on characteristics such as the property's location. Shocks to asset prices could also be geographically correlated with unemployment shocks. Preliminary exercises indicate that allowing for variability in asset price changes – so that some households experience very large price falls – can substantially affect loan losses.
  • Property possession costs: the baseline model assumes that there are no costs involved in selling the collateral securing a defaulted loan. However, the default process may be costly, including costs such as property depreciation while the property is unoccupied, lost interest income, fees paid to sales intermediaries, legal fees and increased labour costs in collections departments. Preliminary investigation suggests that including estimates of these other costs has a relatively small (but non-trivial) effect.
  • Multiple periods: the model assumes that households ‘jump to default’ in a single period. However, households with negative financial margins could gradually draw down on liquid assets, possibly sell less-liquid assets, such as property, and unemployed households could return to employment. Including multiple periods and other dynamics could potentially increase or decrease losses.

Preliminary analysis suggests that including liquid assets directly in households' financial margins does not affect the DAR results. For instance, assuming that unemployed households can draw down on their assets for three months to avoid default reduces the pre-stress share of households with negative financial margins in 2010 to about 1¾ per cent (down from 8 per cent) and in the hypothetical scenario to about 2¼ per cent (down from 10 per cent). However, there is a negligible change in DAR, since households with negative financial margins have less assets (having drawn down on them) and thus higher LGD.


For example, Watson and Wooden (2004) demonstrate that the population-weighted sum of debt reported in the 2002 HILDA Survey was about 20 per cent below aggregate measures. [16]