RDP 2020-01: Credit Spreads, Monetary Policy and the Price Puzzle 6. Robustness to Model Misspecification and the GFC Episode

I find my results to be robust along several dimensions. In the following, I will highlight two main findings relating to model misspecification and the sensitivity of my results to the GFC episode. I discuss further robustness exercises in Appendix D. Throughout, I will focus on the effects on inflation and the unemployment rate since the response of real GDP is unchanged.

6.1 Model Misspecification – The Role of Financial Variables for Policy Transmission

An alternative explanation for the emergence of the price puzzle is that the SVAR model is misspecified. One possibility highlighted by Caldara and Herbst (2019) and Gertler and Karadi (2015) is that credit spreads may not only be an important omitted variable in the central bank's reaction function, but also an important channel for the transmission of monetary policy typically omitted from the VAR model. To test this explanation, I add the domestic money market and large business lending spreads to the SVAR model. Here, I compare the results from a VAR including the original, anticipated RR shock used by BT to one using my new preferred, unanticipated shock series. I order both variables after the policy shock so that they may respond instantaneously to a change in the cash rate. For the BT shock series this also retains the assumption that the cash rate does not respond to credit conditions instantaneously, while the new preferred shock series is already purged of this contemporaneous response. While adding domestic money and credit spreads to the baseline SVAR including the BT shock series removes the price puzzle in the long run, it cannot explain the initial emergence of the price or unemployment puzzles (Figure 10, left panels). In contrast, the results using the new preferred, unanticipated shock are little changed (right panels).

As discussed in Appendix D, I find my results to also be robust to a range of other SVAR specifications, and when estimating the effects of a cash rate change on inflation and unemployment using the univariate LP framework by Jordà (2005).

Figure 10: Monetary Policy Effects – Augmented VAR
Cumulative quarterly responses to 100 basis point cash rate shock, 1994:Q1–2018:Q4
Figure 10: Monetary Policy Effects – Augmented VAR

Notes: Responses from baseline SVAR (dark solid and dashed lines, see Figure 1 for further notes) and SVAR augmented by money market risk spread and large business lending rate spread (lighter solid lines)
(a) Original, anticipated policy shock used by Bishop and Tulip ((2017), baseline SVAR responses as shown in left panels of Figure 7)
(b) New, unanticipated policy shock purged of the response to credit spreads (baseline SVAR responses as shown in right panels of Figure 9)

6.2 Excluding the GFC and Sub-sample Evidence

Since most of the variation in both cash rate changes and credit spreads occurred during the GFC, a natural question to ask is if my results are driven purely by this episode. This is not the case.

First, the cash rate response is largely robust to including a GFC dummy that takes the value of 1 for the four quarters from 2008:Q3–2009:Q2 or dropping this episode from the sample (Table 7). The only exception to this is the response to US corporate bond spreads which is no longer significant. This suggests that the cash rate responded to US financial market conditions only, or more strongly, during the GFC, possibly picking up the sharp spike in Australian corporate bond spreads not included in the model. The policy reaction function also appears largely robust over various sub-samples. Here, the only the exception is the response to large business lending spreads which appears to have increased (in the admittedly short sub-sample) since the GFC.

Table 7: Estimated Policy Rules with Credit Spreads – Sub-sample Evidence
Variable GFC dummy Excluding GFC Pre-GFC Post-GFC
c s t MM −0.49*** −0.48*** −0.54*** −0.47
c s t LB −0.09* −0.08* −0.15* −0.33**
c s t USBAA −0.04 −0.05 −0.09* 0.02
D t GFC −0.26**      
Observations 100 96 62 38
R 2 0.575 0.447 0.635 0.414
Adjusted R 2 0.527*** 0.390*** 0.572*** 0.225*
Notes: See notes for Table 1; all models include a constant and the set of forecasts as in Table 1; ‘Excluding GFC’ excludes 2008:Q3–2009:Q2, ‘Pre-GFC’ uses 1994:Q1–2008:Q2, ‘Post-GFC’ uses 2009:Q3–2018:Q4; statistical significance of the difference in model fit is assessed against the BT benchmark model over the same sub-samples (not shown)

Second, the predictive ability of domestic money and credit market spreads for the Bank's inflation forecast errors is also not exclusively driven by the GFC episode (Table 8). However, since the GFC the predictive information in credit spreads over and above what is already captured in the Bank's forecasts appears to have declined. This may be driven by two explanations. First, the sample is admittedly very small, and hence this result may reflect the lack of power of the forecast efficiency test. Second, it may also reflect learning by the Bank. As the GFC highlighted the importance of a smooth functioning of credit markets for the real economy, credit spreads arguably received stronger attention in the aftermath of the crisis. This may be reflected either in better forecasts by the Bank or a stronger response of the cash rate to offset any macroeconomic effects of changes in domestic credit market conditions as shown in Table 7.

Table 8: Credit Market Conditions and Inflation Forecast Errors – Sub-samples
One-year-ahead forecast, 1994:Q1–2018:Q4
Predictor GFC dummy Excluding GFC Pre-GFC Post-GFC
Constant 0.47* 0.39 1.20*** −0.76
π t1 0.37 0.55 0.11 −0.37
Δu r t1 0.02 −0.09 0.06 0.40
Δgd p t1 0.10 0.12 0.07 −0.24**
c s t MM −0.73 −0.73* −0.18 0.15
c s t LB −0.30 −0.33* −0.53 0.29
D t GFC 0.14      
Observations 91 80 51 29
R 2 0.200*** 0.226*** 0.344*** 0.531
F-statistic 2.01** 2.27** 2.39** 2.39*
Notes: ‘Excluding GFC’ excludes 2008:Q3–2009:Q2, ‘Pre-GFC’ uses 1994:Q1–2008:Q2, ‘Post-GFC’ uses 2009:Q3–2018:Q4; see Table 4 for further notes; statistical significance of the marginal predictive power relative to the baseline model (not shown) is assessed over the same sub-samples

As a result, the estimated effects of monetary policy are also largely robust to excluding the GFC episode (Figure 11). However, while inflation shows the expected response throughout, the unemployment rate appears to fall in the long run. This is likely due to the considerable estimation uncertainty associated with long-run estimates from the LP framework used here. I exclude the GFC from both the first-step estimation of the RR regression and from the second-step LP regression. See Appendix D for further specifications of the LP regression.

Figure 11: Monetary Policy Effects – Excluding the GFC
Cumulative quarterly responses to 100 basis point cash rate shock
Figure 11: Monetary Policy Effects – Excluding the GFC

Notes: Dashed lines show 90% heteroskedasticity and autocorrelation robust confidence intervals; 1994:Q1–2018:Q4, excluding 2008:Q3–2009:Q2
(a) Original, anticipated policy shock used by Bishop and Tulip (2017)
(b) New, preferred unanticipated policy shock