Index Tracking Issues
Prof. Dr. Marlene M ¨uller This version: December 4, 2009
Plan
Problem Tracking Basket Index Replication Case Study ConclusionsProblem
index tracking issues
- tracking an index (e.g. to construct an ETF)
- determining a strategy to (regularly) re-adjust tracking basket
- limit transaction costs by restricting the number of adjusted weights - implement other restrictions on weights (e.g. no shortselling)
- implement budget constraints (e.g. self-financing→sell ≈buy)
this talk
- which optimization criterion to use?
- case study for comparing different optimization criteria
2
Tracking Criteria
letBdenote the tracking portfolio (”basket”) andI the index (we do not specify if these denote returns or prices for the moment)
consider B= m X i=1 wiXi =w⊤X
wherew = (w1, . . . ,wm)⊤ is a weight vector (unit: no. of stocks) and X = (X1, . . . ,Xm)⊤ the universe of stocks to analyze
notation:
- X =R(returns) or X =S(stock prices) - RI, RB (index and basket returns)
- SI, SB (index and basket values)
Tracking Error
tracking error (volatility of) return differences
TE=
q
E(RI−RB)2 variants:
tracking error volatility
TEvol =pVar(RI−RB)
volatility of portfolio value instead of return differences
TEvol
,prices = p
Var(SI−γ·SB)
(see e.g. Meucci; 2005)
residual standard deviations
TEres =σ(RB) q
1−Corr(I,B)2
(see e.g. Spremann; 2008)
mean absolute deviation
TEmad =E|RI−RB|
(Rudolf, Wolter and Zimmermann; 1999; Satchell and Hwang; 2001)
4
Correlation and Other Concepts
correlation Corr =Corr(RI,RB) = Cov(RI,RB) p Var(RI)· p Var(RB)
correlation of portfolio values
Corr=Corr(SI,SB) = Cov(SI,SB) p Var(SI)· p Var(SB) cointegration
SI andSB are cointegrated(see e.g. Alexander; 2001)
i.e. SI−γ·SB is stationary, can be tested by e.g.:
- Dickey-Fuller (DF), Augmented Dickey–Fuller (ADF) tests - Kwiatkowski-Phillips-Schmidt-Shin (KPSS) test
market capitalization weight method
Relationships
TE vs. TEvol
E{(RI−ERI)−(RB−ERB)}2
= (RI−ERI)2 +E(RB −ERB)2 −2E(RI −ERI)(RB−ERB) = Var(RI) +Var(RB)−2 Cov(RI,RB)
= Var(RI−RB)
TEres vs. TEvol
RB = α+βRI+ε
⇒ Var(ε) =Var(RB −βRI) =Var(RB)
n
1−Corr(RI,RB)2 o
TEres vs. Corr
⇒ TEres leads to a quadratic optimization problem while Corr gives
an non-convex objective function inw (see next slides)
6
Which Criterion to Use?
all criteria are ,,average deviations” between index and basket
⇒ no difference between over- and under-performance
in particular, important to note:
- a TEvol of 0 does not imply that index and basket (returns) are
proportional
- a correlation of 1 does not imply that index and basket (returns) are proportional
(as this may happen if both have constant difference)
Optimization with TE
voland Correlation
assume the following constellation:
- hold a part Bhold of the basket fixed and optimize only w.r.t. a small
stock universe X
- then:
B=Bhold +w⊤X
- optimize
min
w TEvol(w) or maxw Corr(w)
under additional restrictions onw (e.g. w ≥0, . . . )
8
TE
volOptimization
we have
TE2vol(w) = Var(I−B) = Var(I) +Var(B)−2 Cov(I,B) = Var(I) +Nw2 −2Dw with Dw = Cov(I,Bhold) + m X i=1 wiCov(Xi,I) Nw2 =Var(Bhold) +2 m X i=1 wiCov(Bhold,Xi) + m X i,j=1 wiwjCov(Xi,Xj) therefore min w TEvol(w) ⇐⇒ minw n Nw2 −2Dw o ⇐⇒ min w n (c+2d⊤w +w⊤Σw)−2(a+b⊤w)o
Correlation Optimization
we have now
Corr(w) = p Cov(I,B)
Var(I)·pVar(B) =
Cov(I,Bhold) +Cov(I,Bnew) p
Var(I)·pVar(B)
= p Dw Var(I)·Nw with again Dw = Cov(I,Bhold) + m X i=1 wiCov(Xi,I) Nw2 =Var(Bhold) +2 m X i=1 wiCov(Bhold,Xi) + m X i,j=1 wiwjCov(Xi,Xj) therefore max w Corr(w) ⇐⇒ maxw Dw Nw ⇐⇒ max w a+b⊤w √ c+2d⊤w+w⊤Σw 10
Prices or Returns?
simple scenario:(jointly) log-normal stock prices ⇐⇒ (jointly) normal returns
Rit ∼N(ri, σ2i) iid. over timet
then forreturnsof stocksi,j
ERit =ri, Var(Rit) =σi2, Cov(Rit,Rjt) = ρijσiσj
but forstock prices
ESit = Si0exp(rit+σi2t2/2) Var(Sit) = S2i0exp “ 2rit+σi2t2 ” {exp(σi2t2)−1}
Cov(Sit,Sjt) = Si0Sj0exp
(ri+rj)t+ 1 2(σ 2 i +σj2)t2 ff n
exp(ρijσiσjt2)−1o
⇒ optimization over time horizont+ model assumption
⇒ on the other hand: objective function on prices / portfolio values
Index Replication
two essential steps - selection - allocation
variety of methods (+ constraints!) - regression type
- least squares regression
- principal components regression - partial least squares
- . . . - optimization - correlation - tracking error - . . . - cointegration analysis - . . . 12
Case Study
- aiming at comparing different criteria (TE, TEvol, Corr)
- uses a constraint on the maximal numberm of re-adjustments (take
care of transaction costs)
- uses a budget constraint (self-financing)
- selection⇒simple search algorithm
step 0: find an initial basket by (positive) regression on the 8 stocks which are most correlated with the indexI
step 1: remove consecutively thosem stocks which give
largest deviations of the objective criterion from its starting value (backward model selection)
step 2: add consecutively up tom stocks which give largest
improvements to the objective criterion from its starting value (forward model selection, test all
combinations of up to mstocks)
remark: stocks removed in step 1 can be added again in step 2 (reweighting!)
DAX –
Fit: 01–12/2008, Eval: 01–06/2009, Corr Optimization 3500 4000 4500 5000 5500 6000 Eval Period (2009−01−02 to 2009−06−30) Inde xJan Mär Mai Jul
Index with Initial and Optimal Baskets
Index Initial Optimal 0.95 1.00 1.05 1.10 1.15 Eval Period (2009−01−02 to 2009−06−30) Ratio Inde x to Bask ets
Jan Mär Mai Jul
Ratio of Index to Initial and Optimal Baskets
Index Initial Optimal
- initial basket: regression (w≥0) - maximal changes: 4 StartWeights D.ALV 0.0294 D.BAS 0.2164 D.BMW 0.1183 D.DAI 0.0495 D.DBK 0.0361 D.MAN 0.0136 D.SIE 0.0911 D.TKA 0.1178 OptWeights D.ALV 0.0294 D.BAYN 0.0771 D.DAI 0.0495 D.DBK 0.0361 D.DTE 0.0427 D.HEN3 0.0290 D.TKA 0.1178 D.VOW 0.0383 Deleted Added Changed 1 D.BAS D.BAYN 2 D.BMW D.DTE 3 D.MAN D.HEN3 4 D.SIE D.VOW StartBasket OptBasket Value 20.6832 20.6832 Corr 0.8523 0.9185 TE 0.0001 0.0002 TE.vol 0.0118 0.0140 TE.res 0.0100 0.0040 TE.mad 0.0006 0.0014 DF.p 0.4162 0.1234 ADF.p 0.8102 0.4796 KPSS.p 0.0100 0.0100 14
DAX –
Fit: 01–12/2008, Eval: 01–06/2009, TE Optimization3500 4000 4500 5000 5500 6000 Eval Period (2009−01−02 to 2009−06−30) Inde x
Jan Mär Mai Jul
Index with Initial and Optimal Baskets
Index Initial Optimal 0.95 1.00 1.05 1.10 1.15 Eval Period (2009−01−02 to 2009−06−30) Ratio Inde x to Bask ets
Jan Mär Mai Jul
Ratio of Index to Initial and Optimal Baskets
Index Initial Optimal
- initial basket: regression (w≥0) - maximal changes: 4 StartWeights D.ALV 0.0294 D.BAS 0.2164 D.BMW 0.1183 D.DAI 0.0495 D.DBK 0.0361 D.MAN 0.0136 D.SIE 0.0911 D.TKA 0.1178 OptWeights D.ALV 0.0294 D.BAS 0.2402 D.BMW 0.0937 D.DAI 0.0495 D.DBK 0.0361 D.DPW 0.1302 D.DTE 0.1067 D.SIE 0.0911 Deleted Added Changed 1 D.MAN D.DPW D.BAS 2 D.TKA D.DTE D.BMW StartBasket OptBasket Value 20.6832 20.6832 Corr 0.8523 0.8596 TE 0.0001 0.0001 TE.vol 0.0118 0.0116 TE.res 0.0100 0.0099 TE.mad 0.0006 0.0002 DF.p 0.4162 0.1456 ADF.p 0.8102 0.5297 KPSS.p 0.0100 0.0100 15
DAX –
Fit: 01–12/2008, Eval: 01–06/2009, TEvol Optimization 3500 4000 4500 5000 5500 6000 Eval Period (2009−01−02 to 2009−06−30) Inde xJan Mär Mai Jul
Index with Initial and Optimal Baskets
Index Initial Optimal 0.95 1.00 1.05 1.10 1.15 Eval Period (2009−01−02 to 2009−06−30) Ratio Inde x to Bask ets
Jan Mär Mai Jul
Ratio of Index to Initial and Optimal Baskets
Index Initial Optimal
- initial basket: regression (w≥0) - maximal changes: 4 StartWeights D.ALV 0.0294 D.BAS 0.2164 D.BMW 0.1183 D.DAI 0.0495 D.DBK 0.0361 D.MAN 0.0136 D.SIE 0.0911 D.TKA 0.1178 OptWeights D.ALV 0.0294 D.BAS 0.2070 D.BMW 0.1183 D.DAI 0.0495 D.DBK 0.0361 D.DTE 0.1156 D.SIE 0.0911 D.TKA 0.0944 Deleted Added Changed 1 D.MAN D.DTE D.BAS
2 D.TKA StartBasket OptBasket Value 20.6832 20.6832 Corr 0.8523 0.8612 TE 0.0001 0.0001 TE.vol 0.0118 0.0115 TE.res 0.0100 0.0098 TE.mad 0.0006 0.0005 DF.p 0.4162 0.1142 ADF.p 0.8102 0.4703 KPSS.p 0.0100 0.0100 16
DAX –
Fit: 01–12/2008, Eval: 01–06/2009, TEres Optimization3500 4000 4500 5000 5500 6000 Eval Period (2009−01−02 to 2009−06−30) Inde x
Jan Mär Mai Jul
Index with Initial and Optimal Baskets
Index Initial Optimal 0.95 1.00 1.05 1.10 1.15 Eval Period (2009−01−02 to 2009−06−30) Ratio Inde x to Bask ets
Jan Mär Mai Jul
Ratio of Index to Initial and Optimal Baskets
Index Initial Optimal
- initial basket: regression (w≥0) - maximal changes: 4 StartWeights D.ALV 0.0294 D.BAS 0.2164 D.BMW 0.1183 D.DAI 0.0495 D.DBK 0.0361 D.MAN 0.0136 D.SIE 0.0911 D.TKA 0.1178 OptWeights D.BMW 0.1183 D.DAI 0.0495 D.FME 0.0092 D.FRE3 0.0029 D.MRK 0.0005 D.SIE 0.0911 D.TKA 0.1178 D.VOW 0.0370 Deleted Added Changed 1 D.ALV D.FME 2 D.BAS D.FRE3 3 D.DBK D.MRK 4 D.MAN D.VOW StartBasket OptBasket Value 20.6832 20.6832 Corr 0.8523 0.8882 TE 0.0001 0.0002 TE.vol 0.0118 0.0138 TE.res 0.0100 0.0051 TE.mad 0.0006 0.0010 DF.p 0.4162 0.0461 ADF.p 0.8102 0.3132 KPSS.p 0.0100 0.0100
DAX –
Fit: 01–12/2009, Eval: 01–06/2009, TEmad Optimization 3500 4000 4500 5000 5500 6000 Eval Period (2009−01−02 to 2009−06−30) Inde xJan Mär Mai Jul
Index with Initial and Optimal Baskets
Index Initial Optimal 0.95 1.00 1.05 1.10 1.15 Eval Period (2009−01−02 to 2009−06−30) Ratio Inde x to Bask ets
Jan Mär Mai Jul
Ratio of Index to Initial and Optimal Baskets
Index Initial Optimal
- initial basket: regression (w≥0) - maximal changes: 4 StartWeights D.ALV 0.0294 D.BAS 0.2164 D.BMW 0.1183 D.DAI 0.0495 D.DBK 0.0361 D.MAN 0.0136 D.SIE 0.0911 D.TKA 0.1178 OptWeights D.ALV 0.0294 D.DAI 0.0495 D.DBK 0.0361 D.EOAN 0.1176 D.MAN 0.0136 D.MEO 0.1261 D.MUV2 0.0474 D.TKA 0.1832 Deleted Added Changed 1 D.BAS D.EOAN D.TKA 2 D.BMW D.MEO 3 D.SIE D.MUV2 StartBasket OptBasket Value 20.6832 20.6832 Corr 0.8523 0.8084 TE 0.0001 0.0002 TE.vol 0.0118 0.0135 TE.res 0.0100 0.0096 TE.mad 0.0006 0.0000 DF.p 0.4162 0.1682 ADF.p 0.8102 0.4913 KPSS.p 0.0100 0.0445 18
SX5E –
Fit: 01–12/2008, Eval: 01–06/2009, Corr Optimization160 180 200 220 240 260 Eval Period (2009−01−02 to 2009−06−30) Inde x
Jan Mär Mai Jul
Index with Initial and Optimal Baskets
Index Initial Optimal 0.95 1.00 1.05 1.10 Eval Period (2009−01−02 to 2009−06−30) Ratio Inde x to Bask ets
Jan Mär Mai Jul
Ratio of Index to Initial and Optimal Baskets
Index Initial Optimal
- initial basket: regression (w≥0) - maximal changes: 4 StartWeights D.DAI 0.0547 E.BBVA 0.0310 E.SCH 0.0786 E.TEF 0.1386 F.MIDI 0.1043 F.TAL 0.0773 H.RDSA 0.1675 S.CSGN 0.0852 OptWeights D.DAI 0.0547 E.SCH 0.0786 E.TEF 0.1386 F.GSZ 0.0587 F.MIDI 0.1043 H.RDSA 0.1793 S.ABB 0.0642 S.CSGN 0.0852 Deleted Added Changed 1 E.BBVA F.GSZ H.RDSA 2 F.TAL S.ABB StartBasket OptBasket Value 14.6749 14.6749 Corr 0.9666 0.9718 TE 0.0000 0.0000 TE.vol 0.0059 0.0053 TE.res 0.0053 0.0052 TE.mad 0.0003 0.0006 DF.p 0.0100 0.0402 ADF.p 0.0369 0.0632 KPSS.p 0.0100 0.0100 19
SX5E –
Fit: 01–12/2008, Eval: 01–06/2009, TE Optimization 160 180 200 220 240 Eval Period (2009−01−02 to 2009−06−30) Inde xJan Mär Mai Jul
Index with Initial and Optimal Baskets
Index Initial Optimal 0.95 1.00 1.05 1.10 Eval Period (2009−01−02 to 2009−06−30) Ratio Inde x to Bask ets
Jan Mär Mai Jul
Ratio of Index to Initial and Optimal Baskets
Index Initial Optimal
- initial basket: regression (w≥0) - maximal changes: 4 StartWeights D.DAI 0.0547 E.BBVA 0.0310 E.SCH 0.0786 E.TEF 0.1386 F.MIDI 0.1043 F.TAL 0.0773 H.RDSA 0.1675 S.CSGN 0.0852 OptWeights D.DAI 0.0547 E.SCH 0.0786 E.TEF 0.1386 F.MIDI 0.0895 F.TAL 0.0773 H.RDSA 0.2061 S.ABB 0.0728 S.UBSN 0.0739 Deleted Added Changed 1 E.BBVA S.ABB F.MIDI 2 S.CSGN S.UBSN H.RDSA StartBasket OptBasket Value 14.6749 14.6749 Corr 0.9666 0.9729 TE 0.0000 0.0000 TE.vol 0.0059 0.0052 TE.res 0.0053 0.0052 TE.mad 0.0003 0.0009 DF.p 0.0100 0.3859 ADF.p 0.0369 0.0946 KPSS.p 0.0100 0.0100 20
SX5E –
Fit: 01–12/2008, Eval: 01–06/2009, TEmadOptimization160 180 200 220 240 260 Eval Period (2009−01−02 to 2009−06−30) Inde x
Jan Mär Mai Jul
Index with Initial and Optimal Baskets
Index Initial Optimal 0.95 1.00 1.05 1.10 1.15 1.20 Eval Period (2009−01−02 to 2009−06−30) Ratio Inde x to Bask ets
Jan Mär Mai Jul
Ratio of Index to Initial and Optimal Baskets
Index Initial Optimal
- initial basket: regression (w≥0) - maximal changes: 4 StartWeights D.DAI 0.0547 E.BBVA 0.0310 E.SCH 0.0786 E.TEF 0.1386 F.MIDI 0.1043 F.TAL 0.0773 H.RDSA 0.1675 S.CSGN 0.0852 OptWeights D.EON 0.0705 E.BBVA 0.0310 E.SCH 0.0786 E.TEF 0.1386 F.MIDI 0.1555 F.TAL 0.0773 H.RDSA 0.1675 M.NOKP 0.0966 Deleted Added Changed 1 D.DAI D.EON F.MIDI 2 S.CSGN M.NOKP StartBasket OptBasket Value 14.6749 14.6749 Corr 0.9666 0.9600 TE 0.0000 0.0000 TE.vol 0.0059 0.0063 TE.res 0.0053 0.0062 TE.mad 0.0003 0.0000 DF.p 0.0100 0.1427 ADF.p 0.0369 0.5016 KPSS.p 0.0100 0.0100
Conclusions
(discrete-time) index tracking is closely related to Markowitzportfolio optimization, requires similar algorithms (quadratic optimization), see also Roll (1992)
the optimization task gets more involved as soon as there areconstraints like e.g. the number of stocks to be reweigted when re-adjusting the basket
criteria like TE, TEvol, Corr seem to be similar on a first glance, buttheir choice may be sensitive for the optimization results . . . more analyses needed
other concepts are cointegration approaches and marketcapitalization weight methods
22
References
Alexander, C. (2001). Market Models: A Guide to Financial Data Analysis, Wiley.
Alighanbari, M. and Mougeot, N. (2009). On optimal tracking error,Working paper, Deutsche Bank, Quantitative Research.
Meucci, A. (2005). Risk and Asset Allocation, Springer-Verlag, Berlin, Heidelberg.
Roll, R. (1992). A mean-variance analysis of tracking error,Portfolio Managementpp. 13–22. Rudolf, M., Wolter, H.-J. and Zimmermann, H. (1999). A linear model for tracking error
minimization,Journal of Banking & Finance23(1): 85–103.
Satchell, S. and Hwang, S. (2001). Tracking error: Ex-ante versus ex-post measures,Working Paper wp01-15, Warwick Business School, Financial Econometrics Research Centre. Spremann, K. (2008). Portfoliomanagement, Oldenbourg, M ¨unchen. 4. ¨uberarb. Aufl.