House prices, credit and the effect of monetary policy in Norway: evidence from structural VAR models

Abstract

This paper investigates the responses of house prices and household credit to monetary policy shocks in Norway, using Bayesian structural VAR models. The analysis indicates that the effect of a monetary policy shock on house prices is large, while the effect on household credit is muted. This is consistent with a relatively small refinancing rate (refinancing rate refers to the share of outstanding mortgages that are refinanced each period due to changes in, for example, house prices or interest rate) of the mortgage stock each quarter. Using monetary policy to guard against financial instability by mitigating property-price movements may prove effective, but trying to mitigate household credit may prove costly in terms of GDP and inflation variation.

Appendices

Appendix 1: Data

• Nominal interest rate Norway: Three-month money market rate (NIBOR). Source: Norges Bank

• Prices Norway: Seasonally adjusted consumer price index adjusted for tax changes and excluding energy products (CPI-ATE). Sources: Statistics Norway and Norges Bank

• Domestic Prices Norway: Seasonally adjusted consumer price index domestic sources adjusted for tax changes and excluding energy products. Sources: Statistics Norway and Norges Bank

• Prices Abroad: Trade-weighted consumer price index for 25 trading partners. Sources: Ecowin and Norges Bank

• Real house prices: Seasonally adjusted nominal house prices deflated by the CPI-ATE. Sources: Statistics Norway, Eiendomsmeglerforetakenes forening (EFF), Finn.no, Eiendomsverdi and Norges Bank

• Real household credit: C2 for households chained and break-adjusted deflated by the CPI-ATE and adjusted for population growth. Sources: Statistics Norway and Norges Bank

• GDP mainland Norway: Seasonally adjusted GDP mainland Norway (volumes) from national accounts adjusted for population growth. Source: Statistics Norway

• Household consumption: Seasonally adjusted household final consumption (volumes) from national accounts adjusted for population growth. Source: Statistics Norway

• Residential investment: Seasonally adjusted gross investment in housing (volumes) from national accounts adjusted for population growth. Source: Statistics Norway

• Real exchange rate: Trade-weighted nominal exchange rate index (I-44) for 44 trading partners adjusted for relative prices in Norway and abroad. Sources: Thomson Reuters, Statistics Norway, Ecowin and Norges Bank

• Hours worked: Total hours worked in mainland Norway from national accounts adjusted for population growth. Source: Statistics Norway

• Population: Population from 16 to 74. Source: Statistics Norway

• Output gap: The percentage deviation between GDP for mainland Norway and projected potential GDP for mainland Norway estimated by Norges Bank. See Hagelund and Sturød (2012) for a more detailed description of the output gap estimation. Source: Norges Bank

• Interest rates abroad: Trade-weighted 3-month nominal money market interest rate for four trading partners (SWE, USA, EUR and GBR) Sources: Thomson Reuters and Norges Bank

• Population USA: Working-age population. Source: Bureau of Labor Statistics

• Prices USA: Consumer price index (CPI). Sources: OECD, Main Economic Indicators (database)

• Real house prices USA: Residential property prices, existing houses adjusted for CPI. Source: Federal Housing Finance Agency (FHFA)

• Real household credit USA: Credit to households and NPISHs provided by all sectors adjusted for CPI and population growth. Source: BIS

• Oil price: Brent dated. Source: Thomson Reuters

Appendix 2: Bayesian estimation and priors

I use an uninformative version of the natural conjugate priors described in Gary and Dimitris (2009). For simplicity, assume Eq. 1 is rewritten in the following form:

$$\begin{aligned} Y=X A + E \end{aligned}$$

(3)

where X now includes all regressors in Eq. 1, i.e. lagged endogenous, the constant and the time trend and E has a variance–covariance matrix $\varSigma$. Since $A = (C_0 \quad C_1 \quad A_1 \quad A_2)^{\prime }$, Eq. 3 can be written in the following form:

$$\begin{aligned} y=(I_n\otimes {X}) \alpha + \epsilon \end{aligned}$$

(4)

where n is the number of endogenous variables in the VAR and $\alpha = \mathbf {(}A)$. The natural conjugate prior has the following form:

$$\begin{aligned} \alpha |\varSigma \sim \mathcal {N}(\underline{\alpha },\varSigma \otimes {\underline{V}}) \end{aligned}$$

(5)

and

$$\begin{aligned} \varSigma ^{-1}\sim \mathcal {W}(\underline{S}^{-1},\underline{\nu }) \end{aligned}$$

(6)

where $\underline{\alpha }$, $\underline{V}$, $\underline{\nu }$ and $\underline{S}$ are hyperparameters. Noninformativeness is then achieved by setting $\underline{\nu }=\underline{S}=\underline{V}^{-1}=cI$ and letting $c\rightarrow {0}$. With this prior the posterior becomes:

$$\begin{aligned} \alpha |\varSigma ,y\sim \mathcal {N}(\overline{\alpha },\varSigma \otimes {\overline{V}}) \end{aligned}$$

(7)

and

$$\begin{aligned} (\varSigma ^{-1}|y)\sim \mathcal {W}(\overline{S}^{-1},\overline{\nu }) \end{aligned}$$

(8)

where

$$\begin{aligned}&\displaystyle \overline{V}=(\underline{V}^{-1}+ X^{\prime }X)^{-1}, \end{aligned}$$

(9)

$$\begin{aligned}&\displaystyle \overline{A}=\overline{V}(\underline{V}^{-1}\underline{A}+ X^{\prime }X{\hat{A}}),\end{aligned}$$

(10)

$$\begin{aligned}&\displaystyle \overline{\alpha }=\mathbf {(}\overline{A}),\end{aligned}$$

(11)

$$\begin{aligned}&\displaystyle \overline{S}=S+\underline{S}+{\hat{A}}^{\prime }X^{\prime }X{\hat{A}}+\underline{A}^{\prime }\underline{V}^{-1}\underline{A}-\overline{A}^{\prime }(\underline{V}^{-1}+X^{\prime }X)\overline{A} \end{aligned}$$

(12)

and

$$\begin{aligned} \overline{\nu }=T+\underline{\nu } \end{aligned}$$

(13)

where T is the number of observations and ${\hat{A}} = (X^{\prime }X)^{-1}X^{\prime }Y$ is the OLS estimate of A.

Appendix 3: Robustness

In order to arrive at the best alternative reduced form specifications, I estimate a large number of candidate VARs and compare their out-of-sample forecasting performance. The number of variables in each VAR is 6 or 7²⁰ with a lag order of two or three. A general requirement is that every individual VAR includes the main variables of interest, namely the nominal interest rate, real house price growth and real household credit growth. The candidate reduced form VARs are then generated by adding all possible three or four-variable subsets from a larger set of potentially relevant variables. The list of “relevant” variables includes GDP growth (mainland Norway), output gap, change in output gap, change in hours worked, consumption growth, core inflation, domestic inflation, residential investment growth, change in the real exchange rate, foreign interest rates and real foreign interest rates. This amounts to approximately 400 reduced form models.

All the reduced form models are estimated from 1994 Q1 to 2003 Q4 and than recursively estimated from 2004 Q1 to 2013 Q4.²¹ In the recursive estimation period, the out-of-sample point forecast error is stored. The average mean root squared forecast error (MRSE) for interest rate, house prices and household credit 1–8 quarters ahead is then calculated. The MRSE is adjusted for the standard deviation of the variables when the MRSEs for the three variables of interest are weighted together. The ten models with the lowest MRSE are reported in Table 3.²² Given the potential importance of the oil price for the Norwegian economy, I also estimate the benchmark model with oil price (measured in USD) as an exogenous variable.

Table 4 Alternative models

Monetary policy shocks are identified for all the models in Table 4 and the oil price model using sign restrictions. I impose the restriction that all economic activity²³ and inflation²⁴ measures do not on impact increase after a contractionary policy shock and that the real exchange rate appreciates. I place no restrictions on the response of foreign interest rates or the oil price. Figure 13 shows the median response of real house prices and real household credit to a monetary policy shock for all the models in Table 3 and the model with oil price. This robustness exercise supports the claim that the effect of a monetary policy shock is relatively large for house prices and small for household credit.

Appendix 4: Figures

See Figs. 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12 and 13.

Fig. 1

Core inflation Norway (in this figure, 4-period log difference of prices is used since this is how Norges Bank presents inflation in MPRs. In the models in this paper, log first differences are used)

Fig. 2

Credit and house prices

Fig. 3

Band-pass filtered credit gap. The band-pass filter applied removes growth cycles longer than 8 years. The first difference of this series is used in the models

Fig. 4

Robustness: filtering of the data. Monetary policy shock(sign restrictions) with alternative data transformations. Impulse responses from a monetary policy shock. Identification is achieved through sign restrictions on impact. Solid black lines are median estimates from the baseline model, while dotted lines are the 16th and 84th percentile probability bands, also from the baseline model. The blue dotted lines are the median response when credit is differenced twice. The red dotted lines show the median response when a band-pass filter is applied to all growth series, while the green circled lines are the median response when a HP filter with $\hbox {lambda} = 1600$ is used to filter the credit growth series. (Color figure online)

Fig. 5

Monetary policy shock: sign restrictions. Impulse responses from a monetary policy shock. Identification is achieved through sign restrictions on impact. Solid lines are median estimates, while dotted lines are the 16th and 84th percentile probability bands

Fig. 6

Monetary policy shock: Choleski. Monetary policy ordered last. Impulse responses from a monetary policy shock. Identification is achieved through a Choleski factorisation of the variance–covariance matrix with interest rate ordered last. Solid lines are median estimates, while dotted lines are the 16th and 84th percentile probability bands

Fig. 7

Monetary policy shock: Choleski. House prices ordered last. Impulse responses from a monetary policy shock. Identification is achieved through a Choleski factorisation of the variance–covariance matrix with house prices ordered last and interest rate ordered penultimate. Solid lines are median estimates, while dotted lines are the 16th and 84th percentile probability bands

Fig. 8

Monetary policy shock: long- and short-run restrictions. Impulse responses from a monetary policy shock. Identification is achieved through a combination of zero and long- run restrictions. Solid lines are median estimates, while dotted lines are the 16th and 84th percentile probability bands

Fig. 9

Monetary policy shock: real credit over GDP. Impulse responses from a monetary policy shock on real household debt/GDP using the four identification schemes described above. Solid lines are median estimates, while dotted lines are the 16th and 84th percentile probability bands

Fig. 10

Real house prices and real household credit per capita

Fig. 11

Variance decomposition. Contributions from monetary policy shock. Share of total forecast error variance. The reported shares are median over the posterior distribution for the different identification assumptions

Fig. 12

Monetary policy shock(sign restrictions) in different subsamples. Impulse responses from a monetary policy shock for three different subsamples and the full sample. Identification is achieved through sign restrictions on impact. Solid black lines are median estimates from the full sample, while dotted black lines are the 16th and 84th percentile probability bands, also from the full sample. The coloured dotted lines are median estimates from the different subsamples. (Color figure online)

Fig. 13

Robustness: reduced form model. Monetary policy shock(sign restrictions). Impulse responses from a monetary policy shock for all the models in Table 3. Identification is achieved through sign restrictions on impact. The lines are median estimates from the different models