RDP 2021-07: Macroprudential Limits on Mortgage Products: The Australian Experience 3. Data and Variables

This section details the data, sample and variables used for the empirical analysis. Section 3.1 describes the data sources and the sample choices. Section 3.2 defines the regression variables and provides summary statistics.

3.1 Datasets and sample

The two main data sources comprise banks' confidential data reporting to APRA, and banks' advertised mortgage rates from the subscription-based Canstar database. We supplement these with housing price data from the subscription-based CoreLogic database and publicly available national accounts data from the Australian Bureau of Statistics (ABS).

The sample is quarterly, from 2008:Q1 to 2019:Q3, and covers banks with at least $1 billion in total loan assets. The regression samples each exclude banks that were not in the sample one year prior to the policy announcement, leaving around 28 banks in most cases. Sample banks are classified as either large or mid-sized. The large banks comprise the 4 major banks – ANZ, CBA, NAB and Westpac – which are each similar sized, and together hold around 80 per cent of total mortgage credit in Australia. The remaining banks are much smaller and classified as mid-sized. There were 3 mergers between sample banks during the sample period, so for the quarters prior to the merger, we combine the 2 pre-merged banks into 1 consolidated entity.

To analyse new housing lending, we use data on mortgage commitments. A mortgage commitment takes place after the borrower receives an accepted and signed offer of credit from the lender, typically after the borrower has signed the contract for purchase of the property.^[6] Commitments are categorised by either the borrower type (occupier or investor), or by the repayment structure (P&I or IO), or by both. Commitments are a better measure of mortgage origination than changes in credit outstanding, for 2 reasons. First, following the policy announcement in December 2014, banks raised interest rates on investor mortgages (Section 5.2), and many customers updated their bank about their classification as an investor or occupier. This heavily affected credit quantities outstanding, despite not generating mortgage originations, but only rarely would have affected commitments. These reclassifications are discussed further in Section 7.3. Second, credit outstanding can move for reasons unrelated to new lending, such as refinancing, early repayments, or drawing down on excess balances. Nonetheless, commitments do not perfectly map to new mortgage lending (i.e. originations), which was the target of the limits. Commitments typically lead originations by a few weeks, and in some cases result in no origination.

The Canstar data on advertised mortgage rates cover various mortgage subtypes and require some discretionary alignment with the 4 categories of approvals (occupier P&I, occupier IO, investor P&I, investor IO).^[7] It is important to note that the advertised rates do not cleanly represent rates charged. Borrowers often receive a discount on the advertised rate, which is more common among large banks than mid-sized banks. Our analysis of advertised rates could therefore miss policy effects on discounts, which did occur to some extent (e.g. ACCC 2018b). However, when analysing targeted mortgages, this would more likely understate than overstate the policy effect estimates. That is, given that banks were using rates to reduce demand for targeted mortgages (Sections 5.2 and 6.2), discounts were more likely to be lowered rather than raised, in which case the rises in advertised rates would understate the rises in offered rates. For non-targeted mortgages, the opposite holds true, so it is possible that those estimates of policy effects on interest rates are overstated, particularly for large banks.

3.2 Regression variables

The 2 main dependent variables are commitments growth $\left(CommitGrt{h}_{b,t}^m\right)$ and mortgage rates $\left(IntRat{e}_{b,t}^m\right)$ . $CommitGrt{h}_{b,t}^m$ measures the percentage change in the total dollar value of mortgage commitments from quarter t–1 to quarter t, for mortgage type m $\in$ {investor, occupier, IO, P&I, Total} and by bank b. It is expressed as a decimal, so, for example, 10 per cent growth is 0.1. All infinite values are excluded (i.e. when a bank grows commitments from zero) and remaining outliers above 300 per cent are truncated to 3.

The dependent variable $IntRat{e}_{b,t}^m$ measures bank b ′s advertised mortgage rate at the end of quarter t, minus the cash rate, in percentage points. It is first differenced in most regressions. When analysing the investor policy, ‘investor rates’ are from investor P&I mortgages, and ‘occupier rates’ are from occupier P&I mortgages. When analysing the IO policy, ‘IO rates’ are from investor IO mortgages and ‘P&I rates’ are from investor P&I mortgages. This approach varies the mortgage-type dimension affected by the policy while holding the other dimension constant. It also avoids relying on rates for occupier IO mortgages, which are less common than other mortgage types.

The 2 key explanatory variables are a bank-invariant indicator variable for the quarter a policy is announced (𝕀(policy)_t) and a measure of how far bank b is from the policy limit ( Treatment_b,t ). 𝕀(policy)_t is one in the quarter t the policy is announced, and zero in other quarters. Most regressions use its first to fourth lags, which capture policy effects in the 4 quarters after the announcement. Treatment_b,t takes 2 alternative forms: linear and binary. For the investor policy, the linear measure is the bank's year-ended growth in the total value of investor mortgages outstanding, minus 10 per cent, measured in percentage points.^[8] For the IO policy, the linear measure is the proportion of the bank's quarterly housing commitments that are IO commitments, minus 30 per cent, measured in percentage points. As noted in Section 3.1, the policy referred to the proportion of new mortgage lending, but we do not have data on originations, so we use commitments as a close proxy. The binary form of Treatment_b,t is an indicator variable for whether bank b is above the limit in quarter t, or in other words, for whether the linear form of Treatment_b,t is positive.

The regressions also include sets of macroeconomic and bank-level control variables. The (bank invariant) macroeconomic controls include quarterly growth in national GDP (from the ABS), quarterly growth in national housing prices (from CoreLogic), and the first difference in the quarter-end ‘cash rate’, which is Australia's key monetary policy rate (from the RBA).^[9] These are all expressed in percentages. The macroeconomic controls also include a set of seasonal indicator variables that control for the 4 quarters each year – for example, a ‘Q2’ indicator, which is 1 in Q2 each year and 0 in other quarters – because commitments tend to be highly seasonal (evident in Figure 1 below). In the regression formulas, the vector of macroeconomic controls is represented by MacroControls_t. The bank-level controls are the tier 1 capital ratio and the ratio of deposit liabilities over total liabilities, both expressed in percentage points, sourced from the confidential APRA data. These are represented in the regression formulas by BankControls_b,t.

The regression variables are summarised in Table 1. Most of the regression samples separate large from mid-sized banks. The size difference between large and mid-sized banks is clear from the commitments values. Large banks' mean value of quarterly total commitments is $14.8 billion, with a $4.5 billion standard deviation, while mid-sized banks' mean is $0.9 billion and standard deviation is also $0.9 billion. Commitments growth rates are similar for large and mid-sized banks, but vary more across mortgage types for mid-sized banks. Large banks tend to advertise higher spreads than mid-sized banks, but this likely reflects that large banks more commonly offer discounts off the advertised rate. (The regressions mostly analyse changes in spreads, so any time-invariant influence of discounting would be removed.) Mid-sized banks rely on deposit funding more so than large banks.

Figure 1 displays the aggregate behaviour of some key variables. Most commitments are for occupier P&I mortgages, with the remaining commitments roughly equally split across the other 3 mortgage types, until around 2013 when investor IO commitments pick up (top panel). Commitments have clear seasonality in a 4-quarter cycle. Average mortgage rates across banks are initially very close for different mortgage types, until after the first policy announcement in late 2014 (middle panel). Comparing the middle and bottom panels makes clear that mortgage rates move fairly closely with the cash rate. The most noteworthy behaviour in the macroeconomic variables is in 2015, when the investor policy was being implemented (bottom panel). In the first half of 2015, the cash rate is cut twice – by 25 basis points in both February and May – but it remained constant for the rest of 2015, and throughout 2017 and 2018. Also in the second half of 2015, housing price growth declined from around 5 per cent per quarter to around zero. This is discussed briefly in Section 5.4.2.

Footnotes

Technically, our data are on approvals rather than commitments, but the 2 concepts are very similar in definition and empirically, and ‘commitments’ is now the more commonly used term. An approval occurs once the borrower receives the signed and accepted offer of credit from the lender; a commitment occurs once the borrower signs that offer and hands it back to the lender. [6]

Details of the alignment are as follows. Most banks offer multiple mortgage packages within some categories, such as one ‘basic’ occupier P&I package and another high LVR occupier P&I package. In most cases, there is a single ‘standard variable rate’ package for each of the occupier P&I, investor P&I and investor IO categories, and we use these wherever possible. This helps for consistent alignment across banks. Where this alignment is not possible, we align mortgage packages with the commitments categories based on what alignment gives the least series jumps from discontinuance of particular packages. Prior to the policies, many banks did not operationally distinguish the targeted loan type from the non-targeted type, including in the interest rate charged. In these cases, interest rates are reported for only one category, which we apply to both mortgage types in our sample. [7]

One of the large banks reported an infeasible jump in investor credit outstanding in 2014:Q4 ($24 billion or 45 per cent), accompanied by a very large decline in occupier credit outstanding ($13 billion). This jump appears to be a reclassification of pre-existing credit from occupier to investor, so we adjust the data by adding $13 billion to that bank's investor credit in all quarters before 2014:Q4. [8]

We use non-farm, chain volume, seasonally adjusted GDP. [9]