Sie sind auf Seite 1von 18

2012/9/18

Intro
Last presentation: Apply Nearest Neighborhood method on 4 time-series to identify macro conditions Make use of more time-series?
Challenge: Curse of dimensionality (especially troublesome for NN method)

Dimension reduction through PCA


132 macro time-series are analyzed in this paper

The authors found the first few PCs can predict future annual bond excess return with 40% R^2
2012/9/18 2

Main Theme of The Paper


Predict bond excess return with macro factors backed out from PCA Analysis

Where
= Excess bonding holding return = PCA Components = Cochrane-Piazzesis Tent-shape factor
2012/9/18 3

Bond Excess Return


Bond (of maturity n at t)s excess return over the period of (t, t+1) is the difference between:
Realized holding return: log(Price of the bond with maturity n years at t/Price of the SAME bond with maturity n-1 years at t+1) Yield of 1-year bond (Risk-free): log(1/Price of zero-coupon bond with maturity 1 year at t)

In this paper, n = 2,3,4,5


2012/9/18 4

Macro Factors
PCA components of 132 time-series
Monthly frequency from 1964 to 2003 Real ouput/income, employment, retail, sales, international trade, consumer spending, housing starts, inventory, inflation indices, interest rates, credit spreads, equity/FX indices Important: need to be standardized Two approaches:
Full-sample: use data in the future Expanding-window: more appropriate for trading
2012/9/18 5

Macro Factors Top 8

2012/9/18

Full-Sample Monthly Regression

2012/9/18

Full-Sample Monthly Regression

2012/9/18

Expanding Window Regression


Use data (-inf, t) to perform PCA Then form predictor vector Use data (-inf, t) to fit linear coeffs Use to forecast bond excess return from t to t+1

2012/9/18

Expanding Window Regression

2012/9/18

10

Factor Interpretation: real factor

2012/9/18

11

Factor Interpretation: rates factor

2012/9/18

12

Factor Interpretation: inflation factor

2012/9/18

13

Factor Interpretation: inflation factor

2012/9/18

14

Factor Interpretation: stock market factor

2012/9/18

15

Realized Factor Values Over History

2012/9/18

16

Realized Factor Values Over History

2012/9/18

17

Discussion
Data issue:
They do not consider the gap between end of month and the announcement day Not many independent responses (annual return is used)

Correlation:
Many of the time-series are highly correlated (subcomponents of a single time-series)

PCA stability and interpretation


How stable are factor loadings

Higher frequency data


Initial jobless claims? Commodities prices? Others?
2012/9/18 18

Das könnte Ihnen auch gefallen