Testing for changing persistence in US Treasury on/off spreads under weighted-symmetric estimation

Abstract

The debt management policy changes of 1998–2001 and subsequent reversal of the US government's fiscal position have prompted research on the dynamics of the US Treasury bond market. The recursive break test procedure of Leybourne et al. (2003) is extended by using weighted-symmetric estimation to detect a single change in persistence in US Treasury on/off spreads. It is found that a significant change from I(0) to I(1) occurred in the late 1990s, which appears to be linked to changes in the US Treasury's debt management policy. Monte Carlo evidence shows that correcting for conditional heteroscedasticity in the data can successfully deal with the tests being oversized, albeit at a considerable loss in power for smaller sample sizes and large short-run variation in volatility. It is therefore advisable mainly for large sample sizes.

Keywords:

• persistence
• unit root test
• break point
• Monte Carlo simulation
• US Treasury bonds
• liquidity

Acknowledgements

We would like to thank Victoria Ilyassova for excellent data assistance. The authors would also like to thank the coeditor and an anonymous referee for their helpful and constructive comments. The usual disclaimer applies.

Notes

1. Changes included: January 1998, when the 3-year note was discontinued; May 1998 and February 2000, when the 5-year note and 1-year bill's auction frequencies were reduced from monthly to quarterly; and February and October 2001, when the 1-year bill and 30-year bond were discontinued. The Treasury also increased issue sizes, leading to more liquidity through lower inventory costs. The reduction in the supply of traded securities as a result of the budget surpluses is particularly noteworthy in the case of the 5-year note. Although awarded amounts per auction increased, maturing 5-year notes decreased from $162 billion in 2001 to $61 billion in 2005 based on US General Accounting Office figures. The~issuance frequency of the 5-year note was restored to monthly in mid-2003.

2. On related power gains see Leybourne et al. (2005) and Pantula et al.(1994).

3. Note that if a trend is included in the regression the denominator for becomes .

4. We select Wednesday observations from the daily data to address day-of-the-week effects. Inflation-indexed and callable bond issues are excluded, as are holidays and observations more than 30 basis points, reflecting posting errors. Data source: GovPX.

5. A trend term is not included in the regressions on market efficiency grounds.

6. We let τ vary between 0.15 and 0.85 in 0.01 increments. Note that LKSN use GLS-detrending and trim at 0.20 employing the usual ADF statistics.

7. Following a referee's suggestion we also performed a standard Chow test for structural breaks on the first-difference of the two series. The stability of the regression coefficients was rejected for both maturities using the break dates given in Table 2. These results, however, should be treated with caution as such a test is conducted conditional on equal variances in the pre- and post-break samples.

8. Seo (1999) provides some results demonstrating that DF t-tests have lower power under conditional heteroscedasticity, but is not concerned with the issue of White-correction.

9. The unconditional distribution for in the GARCH model with conditionally normal errors has heavier tails than the normal, but these do not adequately capture the excess kurtosis observed in many financial time series.

10. He and Teräsvirta (1999) show that the unconditional fourth moment of , where , is an iid sequence and , with , exists in GARCH(1, 1) models iff and .

11. LKSN also found this to be the case for the GLS-detrended DF tests.

12. Detailed results for these tests are available upon request.