New insights into the non-linearity of the ECB Taylor Rule

ABSTRACT

We apply a non-parametric procedure, Multiple Adaptive Regression Splines (MARS), to endogenously select the best multiple threshold model for the European Central Bank (ECB) Taylor Rule (TR). MARS offers the advantage of not excluding simpler models if they better fit the data. On monthly data from 1/2000 to 9/2016, the TR exhibits thresholds for both the output gap and inflation and uncover interesting information about the different sensitivities of the ECB to each variable. We conclude that the ECB cares more about the economic activity than officially declared and reacts to inflation deviations only if they sufficiently exceed the official 2% target. We also detect an overall shift in the monetary policy during the 2008 financial crisis that pushed the system towards the Zero Lower Bound (ZLB). We reach a more complete description of the ECB TR than published so far.

Abbreviations: TR, ECB,MARS

KEYWORDS:

• Monetary policy
• nonlinearity
• thresholds
• Taylor Rule
• MARS methodology

JEL CLASSIFICATION:

• E42
• E58
• C14

JEL CLASSIFICATION:

• C14
• E42
• E58

Disclosure statement

No potential conflict of interest was reported by the authors.

Notes

¹ We use data of monthly frequency because MARS needs an important amount of data to select and estimate the best model. Our sample period starts in 1/2000 for the following reasons: a) Eurostat Industrial Production Index are estimated values for 1999 in contrast with the normal values from 2000 onwards; (b) computation of the inflation rates of 1999 needs the monthly HCPI of 1998, a year in which the euro was not operating yet. Even so, we also estimated the model with data from 1/1999 to 9/2016, and obtained no significant differences neither between coefficients nor between thresholds.

² Working with EONIA interest rates presents the advantage of not being constrained by the Zero Lower Bound region as would be the case with official policy rates; this avoids the econometric problems associated with censored data.

³ We have also estimated different models moving the Crisis dummy from 6 months before to 6 months after the proposed date, with no changes either in the detected thresholds and no significant changes in the estimated coefficients.

⁴ We checked for possible alterations of the monetary policy decisions as a result of the enlargement of the Eurozone after May 2004, by including a dummy variable that activates from that date onwards; no structural break was detected before and after the enlargement process.

⁵ We also introduced dummy variables to distinguish between Duisenberg’s, Trichet’s and Draghi’s chairmanships, but they were non-significant.

⁶ As usual for MARS estimations, we discard detection of possible thresholds in the 10% of extreme values of the sample.

⁷ Our output gap threshold is close to the threshold detected by Kulikauskas (2014), who separates the economy into two states according to an output gap value estimated at −0.0135. However, he obtained either non-significant or negative coefficients for ECB response to inflation rates, while we obtain a more complete map of ECB behaviour.