42
Threshold Cointegration and Credit Dynamics Nov 2003, Ames, Iowa Pin Chung

2003 Ames.Tc&Cd

Embed Size (px)

DESCRIPTION

Threshold Cointegration and Credit Dynamics

Citation preview

Page 1: 2003 Ames.Tc&Cd

Threshold Cointegration and Credit Dynamics

Nov 2003, Ames, Iowa

Pin Chung

Page 2: 2003 Ames.Tc&Cd

2

Introduction

Kao (2000)–high-yield bond debacle (Junk bonds).–large derivative losses (LTCM, 1998).–a global credit/liquidity crisis (Argentina, 1999 to

2001).

Credit Risk Pricing Models–1959 to 1992: 5 papers.–1993 to 1999: more than 10 papers.–1998 & 1999: more than 30 papers.

Page 3: 2003 Ames.Tc&Cd

3

Wilson & Jones (1990), Chang & Huang (1990), Adrangi & Ghazanfari (1996/1997):-seasonality, e.g., the January effect and the weekday effect;-defy the market efficiency hypothesis.

Collin-Dufresne, Goldstein & Martin (1999):-hedge funds are sensitive to changes in the credit spread;-a common factor to explain the variation;-gain from studying individual bonds is limited.

Pedrosa & Roll (1998):-credit spread risks are non-diversifiable;-credit spreads of indices are affected by some common factors.

Introduction (continued)

Page 4: 2003 Ames.Tc&Cd

4

The Corporate Debt Pricing Models

Segmentation Model:

Fisher (1959), Silvers (1973), Boardman & McEnally (1981).

Market Yield Premium Model:

Fons (1987), Altman & Bencivenga (1995).

Yield Spread Model:

Fridson & Jonsson (1995), Garman & Fridson (1996).

Yield Premium and Yield Spread model:

Barnhill, Joutz, & Maxwell (2000).

Page 5: 2003 Ames.Tc&Cd

5

Neal, Rolph and Morris (2000, NRM)

• Johansen’s (1988, 1991) cointegration approach.

• U.S. government rates are cointegrated with corporate rates.

• Time horizon dictates the dynamic relationships between credit spreads and Treasury rates.

• Asymmetric Results:– Short-run: Treasury rate ↑è credit spread to narrow.

– Long-run: Treasury rate ↑è credit spreads to widen.

– Corporate bonds more sensitive to interest rate movements.

– Time varying correlation between credit spreads and interest rates.

Page 6: 2003 Ames.Tc&Cd

6

Discontinuous Nonlinear Asymmetric Adjustment Process

Main idea:Asymmetric è specification error.

Why use threshold cointegration methods:Capture the asymmetric behaviors.

We have conducted the tests:Q1: Are interest rates cointegrated? Q2: If rates are cointegrated, then:

1. Test same speeds of adjustment in a two-regime environment.2. In a three-regime environment:

• a random walk process when rates are inside the band;• a mean-reverting process to the equilibrium band, possibly with

different adjustment speeds.

Page 7: 2003 Ames.Tc&Cd

7

OUR GOAL

•Offer non-linear, discontinuous, asymmetric alternatives to the traditional linear, continuous and symmetric approach.

•Examine credit spread dynamics for different maturities and different investment grades.

•Identify the equilibrium adjustment process.

•Perform out-of-sample forecast performance evaluations: symmetric vs. asymmetric.

Page 8: 2003 Ames.Tc&Cd

8

Economic implications

• to study real economic activities.

• to price portfolios.

• to find the value of a firm.

• to price credit derivatives, improve credit management quality.

• to apply to risk management and hedging activities.

• to price different financial instruments.

• to improve the adequacy of reserves held by banks and insurance companies.

• to improve the profitability of trading, the accuracy of current asset pricing and option pricing models.

Page 9: 2003 Ames.Tc&Cd

9

Empirical studies: Threshold Cointegration

• Ghosh (1993), Brenner and Kroner (1995):-futures and spot prices.

• Martens, Kofman and Vorst (1998):-index-futures trading strategies.

• Balke and Wohar (1998):-covered interest parity.

• Goodwin and Holt (1999):-price linkages of U.S. beef market.

Page 10: 2003 Ames.Tc&Cd

10

Models in this paper:

Asymmetric Threshold Cointegration Models:

-Lo and Zivot (2001)

-Hansen and Seo (2002)

-Enders and Siklos (2001)

Symmetric Cointegration Model:

-Engle and Granger (1987)

-Neal, Rolph and Morris (2000)

Page 11: 2003 Ames.Tc&Cd

11

DATA

• Monthly data, 1/1960 to 12/1997, 456 observations.

– 10-year constant maturity Treasury note (Tsy).

– Ibbotson Bond Index for 20-year Treasury bond (Ibb).

– Moody’s Aaa Bond Index (Aaa).

– Moody’s Baa Bond Index (Baa).

• Table 4.1: Tsy, Ibb, Aaa, Baa.

• Table 4.2: high autocorrelations è (near) unit root process.

Page 12: 2003 Ames.Tc&Cd

12

DATA (continued)

Tsy-Ibb-Aaa-Baa

0

3

6

9

12

15

18

21

Jan-

60

Jan-

63

Jan-

66

Jan-

69

Jan-

72

Jan-

75

Jan-

78

Jan-

81

Jan-

84

Jan-

87

Jan-

90

Jan-

93

Jan-

96

%Tsy

Ibb

Aaa

Baa

Page 13: 2003 Ames.Tc&Cd

13

DATA (continued)

• Four yield spreads:– (Aaa-Tsy), (Aaa-Ibb), (Baa-Tsy), (Baa-Ibb).

• 5 unit root tests (Tables 4.3 to 4.6):– Dickey-Fuller, TAR, M-TAR, C-TAR, and M-C TAR.

• Dickey-Fuller test:– reject the null of a unit root process at 5% ( for all pairs).

• Symmetric Adjustment Speeds???:– Aaa-Tsy: M-TAR, C-TAR @ 10%; M-C TAR @ 5%.– Aaa-Ibb: M-C TAR @ 5%.– Baa-Tsy: C-TAR @ 10%; M-TAR and M-C TAR @ 5%.– Baa-Ibb: M-C TAR @ 21%.

Page 14: 2003 Ames.Tc&Cd

14

Results from Lo-Zivot Model

•Tables 5.1.1 5.1.2: TVECM(3) with lag = 1, 2 against the null of Linear VECM.

•The null of no threshold effects:–cannot be rejected for (Aaa, Ibb) and (Baa, Ibb)

pairs for either lag length.

•The null of no threshold effects:–can be rejected for (Aaa, Tsy) and (Baa, Tsy)

pairs with lag = 2.

Page 15: 2003 Ames.Tc&Cd

15

Results from Lo-Zivot Model (continued)

• Lag = 1 in levels:– All long rate eqs (Aaa or Baa) in the 3rd regime have the

expected signs (negative) for their error-correction terms;

– not so noteworthy for the short rate (Tsy or Ibb) equations.

• Lag = 2 in levels:– all long rate eqs (Aaa or Baa) in the 3rd regime have the

expected signs (negative) for their error-correction terms;

– for the short rate (Tsy or Ibb) equations only one coefficient violates the negative sign expectation.

Page 16: 2003 Ames.Tc&Cd

16

Results from Hansen-Seo Model

•Four pairs of interest rates.

•Lag = 1, 2 in levels.

•β = 1 or β is estimated from the model.

•(Aaa, Tsy): strong TC relationship.

•(Baa, Tsy): TC effect if β is estimated.

•(Aaa, Ibb) and (Baa, Ibb): less TC effect.

Page 17: 2003 Ames.Tc&Cd

17

Results from Hansen-Seo Model (continued)

• If β is estimated from the model:– six of eight β estimates are greater than unity, only two are

less than unity (Aaa, Ibb pair).

– in contrast to conventional assumption.

• 1st (2nd) regime is “typical” or “extreme” regime:– depends on interest rate pairs, lag length and how we specify

the β value.

• Sensitivity Tests:– stable estimation results.

– 100 vs. 300 grid points è AIC criterion improves.

– 1000 vs. 5000 replications è moderate changes of p-values.

Page 18: 2003 Ames.Tc&Cd

18

Results from Enders-Siklos model

• Relax the restriction: [1,-1].

• Tables 5.3.1 to 5.3.4 report 5 cointegration tests:– Linear: Engle-Granger.– Nonlinear: TAR, M-TAR, C-TAR, M-C TAR.

• Engle-Granger test:– reject the null of unit root process.– cointegration relationship with symmetric adjustment.

• Asymmetric adjustment @ 5% significance level:– (Aaa, Ibb): M-C TAR.– (Aaa, Tsy): C-TAR, M-C TAR.– (Baa, Ibb): C-TAR.– (Baa, Tsy): M-C TAR.

Page 19: 2003 Ames.Tc&Cd

19

Results from Enders-Siklos model (continued)

•M-C TAR provides stronger evidence of asymmetric behavior than the M-TAR.

•Similar observation for C-TAR vs. TAR.

•Expect smaller AIC and BIC under M-C TAR:–(Aaa, Tsy) and (Aaa, Ibb): Yes.

–(Baa, Tsy) and (Baa, Ibb): TAR.

Page 20: 2003 Ames.Tc&Cd

20

Results from Enders-Siklos model (continued)

Error-Correction model: (Aaa, Tsy)

• Enders-Siklos M-C TVECM:1. +1 unit deviation: Tsy 0.28%, Aaa 3.91%.

2. -1 unit deviation: Tsy 0.99%, Aaa 14.31%.

3. Stronger adjustment when Aaa rates wander away under the negative change environment.

• Engle-Granger Linear VECM:1. Symmetric adjustment.

2. Tsy 0.46%, Aaa 6.60%.

Page 21: 2003 Ames.Tc&Cd

21

Forecasting Performance Evaluation

• One-step-ahead to six-step-ahead forecasts.

• (Aaa, Tsy) and (Baa, Tsy): 7 competing models– (2 Lo-Zivot, 2 Hansen-Seo, Enders-Siklos, Engle-Granger, NRM)

• (Aaa, Ibb) and (Baa, Ibb): 6 competing models– (2 Lo-Zivot, 2 Hansen-Seo, Enders-Siklos, Engle-Granger)

• Forecast period: 01/1998 to 12/2002 (60 months)

• Six accuracy measures:– ME, EV, RMSE, RMSPE, MAE, MAPE.

Page 22: 2003 Ames.Tc&Cd

22

Forecasting Performance Evaluation (continued)

(Aaa, Tsy):• 7 under-estimate Aaa: 1.23 to 5.40 bp (me > 0).

• 6 over-estimate Tsy: 2.43 to 23.84 bp (me < 0).– Lo-Zivot (lag=2) under-estimate Tsy 40.40 bp.

• 1-step-ahead:– Enders-Siklos M-C TAR (rmse, rmpse, mae, mape).– Under-estimate Aaa 4.63 bp.– Over-estimate Tsy 2.51 bp.

• 6-step-ahead:– NRM (ev, rmse), Engle-Granger (rmpse, mae), Hansen-

Seo (β=1) (mape).

Page 23: 2003 Ames.Tc&Cd

23

Forecasting Performance Evaluation (continued)

(Baa, Tsy):• 7 under-estimate Baa: 2.66 to 5.44 bp.

• 5 over-estimate Tsy: 1.83 to 4.59 bp.– 2 L-Z under-estimate Tsy 1.08 to 10.59 bp.

• 1-step-ahead:– Hansen-Seo (β=1) (rmse, rmpse, mae, mape).– Under-estimate Baa 3.60 bp.– Over-estimate Tsy 1.83 bp.

• 6-step-ahead:– Hansen-Seo (β=1) (ev, rmse, rmpse, mae, mape)– Under-estimate Baa 15.69 bp.– Over-estimate Tsy 21.38 bp.

Page 24: 2003 Ames.Tc&Cd

24

Forecasting Performance Evaluation (continued)

(Aaa, Ibb):• 6 under-estimate Aaa: 1.77 to 14.21 bp.

• 4 over-estimate Ibb: 3.82 to 7.06 bp.– 2 L-Z under-estimate Ibb 22.48 to 9.16 bp.

• 1-step-ahead:– Hansen-Seo (β estimated) (ev, rmse, rmpse, mae, mape).– Under-estimate Aaa 3.69 bp.– Over-estimate Ibb 3.82 bp.

• 6-step-ahead:– Hansen-Seo (β estimated) (ev, rmse, rmpse, mae, mape).– Under-estimate Aaa 7.54 bp.– Over-estimate Ibb 20.13 bp.

Page 25: 2003 Ames.Tc&Cd

25

Forecasting Performance Evaluation (continued)

(Baa, Ibb):• 5 under-estimate Baa: 1.06 to 5.07 bp.

– Lo-Zivot (lag=1) over-estimate Baa 1.30 bp.

• 4 over-estimate Ibb: 0.72 to 4.89 bp.– 2 L-Z under-estimate Ibb 3.07 to 0.31 bp.

• 1-step-ahead:– Lo-Zivot (lag=1) (ev, rmse, rmpse, mae, mape).– Over-estimate Baa 1.30 bp.– Under-estimate Ibb 3.07 bp.

• 6-step-ahead:– Hansen-Seo (β=1) (ev, rmse, rmpse, mae, mape)– Under-estimate Baa 12.95 bp.– Over-estimate Ibb 6.22 bp.

Page 26: 2003 Ames.Tc&Cd

26

Forecasting Performance Evaluation (continued)

Lo-Zivot with non-unity cointegrating vector:

• (Aaa, Tsy): [1,-1.028](NRW), [1,-1.039](Hansen-Seo)

• (Baa, Tsy): [1,-1.178](NRW), [1,-1.108](Hansen-Seo)

• (Aaa, Ibb): [1,-0.981] (Hansen-Seo)

• (Baa, Ibb): [1,-1.385] (Hansen-Seo)

• 1-step-ahead forecast:– same leading models

• 6-step-ahead forecast:– (Baa, Tsy) [1,-1.178] (NRW) (rmse, rmpse, mae, mape)

Page 27: 2003 Ames.Tc&Cd

27

Enders-Siklos lag=2 Aaa vs Tsy

0

2

4

6

8

10

Jan-98 Jan-99 Jan-00 Jan-01 Jan-02

%

Aaa-Actual

1-step-ahead Aaa

Tsy-Actual

1-step-ahead Tsy

Page 28: 2003 Ames.Tc&Cd

28

Hansen-Seo Fxd1 Baa vs Tsy

0

2

4

6

8

10

Jan-98 Jan-99 Jan-00 Jan-01 Jan-02

%

Baa-Actual

1-step-ahead Baa

Tsy-Actual

1-step-ahead Tsy

Page 29: 2003 Ames.Tc&Cd

29

Hansen-Seo CI1 Aaa vs Ibb

0

2

4

6

8

10

Jan-98 Jan-99 Jan-00 Jan-01 Jan-02

%

Aaa-Actual

1-step-ahead Aaa

Ibb-Actual

1-step-ahead Ibb

Page 30: 2003 Ames.Tc&Cd

30

Lo-Zivot lag=1 Baa vs Ibb

0

2

4

6

8

10

Jan-98 Jan-99 Jan-00 Jan-01 Jan-02

%

Baa-Actual

1-step-ahead Baa

Ibb-Actual

1-step-ahead Ibb

Page 31: 2003 Ames.Tc&Cd

31

Forecasting Performance Evaluation (continued)

• 1-step-ahead forecast: – under-estimate long rate (Aaa, Baa).

– over-estimate short rate (Tsy, Ibb).

– threshold cointegration models perform better than linear cointegration models.

• None of the threshold cointegration models dictates the overall performance.

• Some gains by incorporating the non-zero cointegrating vector into the Lo-Zivot specification.

Page 32: 2003 Ames.Tc&Cd

32

Forecast 2003: Aaa vs. Tsy (E-S)

-

1

2

3

4

5

6

7

8

9

10

J-03 F-03 M-03 A-03 M-03 J-03 J-03 A-03 S-03 O-03

%

Aaa

Tsy

Aaa-For

Tsy-For

Page 33: 2003 Ames.Tc&Cd

33

Forecast 2003: Baa vs. Tsy (H-S)

-

1

2

3

4

5

6

7

8

9

10

J-03 F-03 M-03 A-03 M-03 J-03 J-03 A-03 S-03 O-03

%

Baa

Tsy

Baa-For

Tsy-For

Page 34: 2003 Ames.Tc&Cd

34

Forecast 2003: Aaa vs. Ibb (H-S)

-

1

2

3

4

5

6

7

8

9

10

J-03 F-03 M-03 A-03 M-03 J-03 J-03 A-03 S-03 O-03

%

Aaa

Ibb

Aaa-For

Ibb-For

Page 35: 2003 Ames.Tc&Cd

35

Forecast 2003: Baa vs. Ibb (L-Z)

-

1

2

3

4

5

6

7

8

9

10

J-03 F-03 M-03 A-03 M-03 J-03 J-03 A-03 S-03 O-03

%

Baa

Ibb

Baa-For

Ibb-For

Page 36: 2003 Ames.Tc&Cd

36

Conclusions

•Major findings:–exist long run equilibrium relationships.

–all interest rates pairs follow the threshold cointegration behavior.

–spreads are stationary.

–the speeds of adjustment are asymmetric.

–the threshold estimates are asymmetric in a three-regime environment.

Page 37: 2003 Ames.Tc&Cd

37

Conclusions (continued)

•Major findings (continued):–1% ↑ in Treasury rates (Tsy or Ibb)

è more than 1% ↑ in corporate bond indices.

–the Baa bond index is more sensitive.

–above findings are coherent with NRW(2000) but inconsistent with the view that increased credit risk will make corporate bonds less interest rate sensitive.

Page 38: 2003 Ames.Tc&Cd

38

Conclusions (continued)

•Major findings (continued):–no one particular threshold cointegration model

dictates the overall forecasting accuracy.

–for different interest rates pairs, different threshold cointegration model offers a better fit.

–linear cointegration models perform relatively less accurate than the threshold cointegration models.

Page 39: 2003 Ames.Tc&Cd

39

Conclusions (continued)

•Future work (theoretical):–allow for two threshold variables in the Enders-

Siklos model (a three-regime setting).

–allow for estimating cointegrating vector, delay variable and threshold variables simultaneously.

–allow for at least three variables in the threshold cointegration model (i.e., multiple cointegrating vectors.)

–develop a distribution theory for the parameter estimates.

Page 40: 2003 Ames.Tc&Cd

40

Conclusions (continued)

•Future work (empirical):–extend the model to have multiple corporate

bond indices.

–include some other macroeconomic variables in the setting to control for economic evolution.

–incorporate some other variables like, liquidity risk, default risk, the expected loss in the event of default to model the yield on risky debts.

Page 41: 2003 Ames.Tc&Cd

41

ACKNOWLEDGEMENTS

•Dr. Barry Falk

•Dr. Harvey Lapan

•Dr. John Schroeter

•Dr. Dermot Hayes

•Dr. Rick Dark

Page 42: 2003 Ames.Tc&Cd

42

Appendix