Analysing Global Fixed Income Markets with Tensors
11 Analysing Global Fixed Income Markets with Tensors
Bruno Scalzo Dees
Abstract —Global fixed income returns span across multiplematurities and economies, that is, they naturally reside on multi-dimensional data structures referred to as tensors . In contrastto standard “flat-view” multivariate models that are agnostic todata structure and only describe linear pairwise relationships,we introduce a tensor-valued approach to model the global risksshared by multiple interest rate curves. In this way, the estimatedrisk factors can be analytically decomposed into maturity-domain and country-domain constituents, which allows the investor todevise rigorous and tractable global portfolio management andhedging strategies tailored to each risk domain. An empiricalanalysis confirms the existence of global risk factors shared byeight developed economies, and demonstrates their ability tocompactly describe the global macroeconomic environment.
I. I
NTRODUCTION
Market participants have long recognized the importanceof identifying the common factors that affect the returns ofsecurities within asset classes. In such a task it is critical todistinguish the common risks that have general impact on thereturns of most securities from the idiosyncratic risks thatinfluence securities individually. For instance, following theseminal work in [1], a significant portion of the fixed incomeliterature has been devoted to the technique of principalcomponent analysis (PCA) [2] to provide a parsimoniousinterpretation to the dynamics of the term structure of inter-est rates. Empirical results suggest that three latent factors,referred to as level , slope and curvature , are required toalmost fully reflect the behaviour of the entire term structure.Moreover, the principal components are frequently identifiedwith economically meaningful events. As such, the degree ofrobustness in these findings has made PCA a fundamentalbuilding block for characterizing single-economy interest ratecurves. Notable benefits of the principal components approachinclude: (i) its analyticity and mathematical tractability; (ii)its ability to parsimoniously describe economic factors; (iii)its applications for stress-testing and scenario analysis; (iv) itsdirect applicability for hedging portfolios.However, the growing interconnectedness of the interna-tional markets presents a major challenge for risk manage-ment of fixed-income securities, owing to the high correla-tion between interest rates across maturities and countries.Financial institutions routinely invest globally using strategieswith limited avenues for diversification. This is largely dueto the legacy analytics which employ “flat-view” multivariatemethods, i.e. PCA, whereby trades are typically hedged byoffsetting their domestic-curve principal components. Thisleaves such strategies unprotected to cross-country risk arisingfrom global macroeconomic events. The most recent creditcrisis, for instance, is exemplar of how macroeconomic shockscan be crucially transmitted across interest rate curves. Forthis reason, a parsimonious model to describe the co-variationof interest rates at the relevant maturities and in the relevantcountries appears necessary for global fixed income investorsto adequately identify and manage risk. B. Scalzo Dees is with the Department of Electrical Engineering, ImperialCollege London, London SW7 2AZ, U.K. (e-mail: [email protected]).
This has naturally motivated various developments of PCAfor joint term structure analysis, however, a wide range ofsolutions have been proposed as a consequence of the ambi-guity in the problem formulation. We find a substantial lackof agreement in the literature, not only regarding the datapreparation and estimation procedure, but also regarding thenumber of latent factors that are required to explain the jointdynamics of multiple yield curves, and the nature of the globaland domestic factors obtained.One approach is to apply PCA to vectorised data obtainedfrom several term structures [3, 4, 5], however, this methodignores the multi-curve structure and leads to factors whichare difficult to interpret and can still reflect idiosyncratic anddomestic behaviour. Another approach is common PCA [6],which extracts the eigenvectors that span an identical spaceacross all countries, however, this method simultaneously di-agonalizes multiple covariance matrices, which is non-analyticfor more than two matrices, and neglects the co-variationbetween assets across countries [7]. Alternatively, inter-batteryfactor analysis [8] captures all common factors across do-mestic term structures [9], however, the solution method isalso computationally prohibitive and implicitly assumes thatidiosyncratic co-variations can only occur domestically, whichis a restrictive and unrealistic assumption. The major limitationshared by the existing techniques is that they are agnostic tothe multi-curve data structure and resort to methods developedfor multivariate analysis. Such a flattened view of the data,and the rigid assumptions inherent in multivariate analysis,are inadequate and ineffective.We recognise that global fixed income returns, which spanacross multiple maturities and countries, reside on regularmulti-dimensional data structures referred to as tensors . It isonly through tensor analysis that we have the opportunityto develop sophisticated models capturing the interactionsbetween the entirety of interest rate curves. This motivates thedevelopment of multilinear techniques, which have eventuallyfound its place in many real-world applications where tensorsnaturally reside [10, 11, 12, 13, 14].Accordingly, we have developed a framework that employsthe structure-aware multilinear algebra to rigorously modelthe risk factors shared by an international universe of fixedincome returns. In this way, the estimated risk factors can beanalytically decomposed into two parallel domains of risk: (i) maturity-domain factors which are shared by all countries; and(ii) country-domain factors which are shared by all maturities.By operating within each domain in parallel, the investor candevise rigorous and tractable global portfolio management andhedging strategies, with fewer decision parameters, that aresimultaneously tailored to each risk domain, as a consequenceand natural extension of the proposed multilinear framework.An empirical analysis confirms the existence of commonglobal risk factors shared by eight developed economies.The resulting maturity-domain and country-domain factors areshown to provide compact and physically meaningful insightinto the global macroeconomic environment. a r X i v : . [ q -f i n . P M ] D ec II. P
REREQUISITES OF T ENSOR A LGEBRA
Tensors are manipulated using the mathematical branch of multilinear algebra , for which we provide a comprehensiveintroduction to the subject from an academically rigorous, yetpractitioner-focused perspective. We refer the reader to [10,11] for more details on the topic.In this work, scalars are denoted by lightface font, e.g. x ; vectors by lowercase boldface font, e.g. x ; matrices byuppercase boldface font, e.g. X ; and tensors by boldfacecalligraphic font, e.g. X . A. Nomenclature
The order of a tensor defines the number of dimensions, alsoreferred to as modes . For instance, the tensor X ∈ R I ×···× I N is of order N and has K = (cid:81) Nn =1 I n elements in total.Tensors can be reshaped into mathematically tractable vec-tor and matrix representations, which we can manipulate usinglinear algebra. The vector representation is denoted by x = vec ( X ) ∈ R K (1)and the mode- n unfolding is obtained by a reshaping the tensorinto the matrix X ( n ) = (cid:104) f ( n )1 f ( n )2 · · · f ( n ) KIn (cid:105) ∈ R I n × KIn (2)which contains the set of column vectors, f ( n ) i ∈ R I n ,known as mode- n fibres . Fibres are the multi-dimensionalgeneralization of matrix rows and columns.The operation of mode- n unfolding can be viewed as thereorientation of the mode- n fibres as column vectors of X ( n ) ,as illustrated in Figure 1. Notice that the order- tensor, X , hasalternative, yet equivalent, representations in terms of mode- (left panel), mode- (middle panel) and mode- fibres (rightpanel), that is, columns, rows and tubes, respectively. n = 1 n = 2 n = 3 X X ( n ) Fig. 1: Illustration of the mode- n unfolding, X ( n ) , of theorder-3 tensor, X , in view of the orientation of mode- n fibres. B. Tensor products
Multilinear algebra is based on the class of operatorsknown as tensor products . The
Kronecker product between thematrices A ∈ R I × I and B ∈ R J × J yields the block matrix A ⊗ B = a B · · · a I B ... . . . ...a I B · · · a II B ∈ R IJ × IJ (3) For convenience, we denote the sequence of Kronecker prod-ucts of the matrices U ( n ) ∈ R I n × I n , for n = 1 , ..., N , by Å N ⊗ n =1 U ( n ) ã = U (1) ⊗ · · · ⊗ U ( N ) ∈ R K × K (4)The mode- n product of the tensor X ∈ R I ×···× I N with thematrix U ∈ R J n × I n is denoted by Y = X × n U ∈ R I ×···× I n − × J n × I n +1 ×···× I N (5)and is equivalent to performing the following steps X ( n ) ← X (cid:46) Mode- n unfold Y ( n ) ← UX ( n ) (cid:46) Left matrix multiplication Y ← Y ( n ) (cid:46) Re-tensorizeFor convenience, the sequence of mode- n products of X andthe matrices U ( n ) ∈ R J n × I n , for n = 1 , ..., N , is written as Y = X N × n =1 U ( n ) = X × U (1) × · · · × N U ( N ) ∈ R J ×···× J N (6)and can be expressed in mathematically equivalent vector andmatrix representations, that is y = Å ⊗ n = N U ( n ) ã x (7) Y ( n ) = U ( n ) X ( n ) Ñ ⊗ i = Ni (cid:54) = n U ( i ) T é (8)Figure 2 illustrates the sequence of mode- n products of anorder-3 tensor with matrices U ( n ) , for n = 1 , , . Y = U (1) X U (2) U (3) Fig. 2: Sequence of mode- n products for n = 1 , , . C. Tensor-valued Gaussian random variables
In standard multivariate data analysis, multiple measure-ments are collected at a given trial, experiment or timeinstant, to form the vector-valued sample, x ∈ R K . Anassumption typically adopted in statistical modeling is thatvariables are described by the distribution x ∼ N ( m , Σ ) ,which implies that the covariance matrix, Σ ∈ R K × K , isunstructured. However, if the variables have a natural tensorrepresentation, then it is desirable, if not necessary, to assumethat the covariance matrix, Σ , exhibits a more structured formmotivated by economic considerations.For instance, real-world order- N tensor-valued signals en-countered in finance include:i) interest rates over curve × maturity × country ( N = 3 );ii) futures prices over asset × maturity ( N = 2 );iii) options prices over asset × maturity × strike ( N = 3 ). The statistical properties of tensor-valued random variables areintrinsically linked to that of Gaussian random fields, describedas follows. Consider the zero-mean random variable on an N -dimensional coordinate system, denoted by x : R N (cid:55)→ R , anddescribed by the coordinate-dependent distribution x ( z ) ∼ N (cid:0) , σ ( z ) (cid:1) (9)where z = { z (1) , ..., z ( N ) } ∈ R N is an N -dimensionalcoordinate vector, and z ( n ) ∈ R is the n -th axis coordinate.Furthermore, assume that the random variable is equipped witha covariance operator σ : R N × R N (cid:55)→ R σ ( z , z ) = cov { x ( z ) , x ( z ) } (10)where σ ( z , z ) ≡ σ ( z ) . The random variable is said to exhibita separable covariance structure if and only if the covarianceoperator is separable, that is, if the following condition holds σ ( z , z ) = N (cid:89) n =1 σ ( n ) ( z ( n )1 , z ( n )2 ) , ∀ z , z ∈ R N (11)where σ ( n ) : R × R (cid:55)→ R is the covariance operator specificto the n -th coordinate axis, and independent to the covarianceoperator on the other axes. Remark 1.
For contextual clarity, consider the random vari-able, x ( m, c ) ∈ R , which represents the fixed income returnas a function of the maturity m and country c . This can beviewed as a scalar field on a 2-dimensional coordinate system( maturity × country ). The separability condition asserts that cov { x ( m i , c j ) , x ( m l , c k ) } = σ ( m ) il σ ( c ) jk (12)where σ ( m ) il is the return covariance between the i -th and l -thmaturities, and is independent of the countries. Similarly, σ ( c ) jk is the return covariance between the j -th and k -th countries,and is independent of the maturities.The act of forming the tensor X ∈ R I ×···× I N from K scalar-valued variables is known as tensorization . The scalar-valued samples are ordered as follows [ X ] i ...i N = x ( z (1) i , ..., z ( N ) i N ) , i n = 1 , ..., I n , z ( n ) i n ∈ R (13)Figure 3 illustrates the tensorization of scalar-valued randomvariables to form an order-3 tensor. ( 1 , 1 , 1 ) ......... ... ( I , 1 , 1 )( 1 , 2 , 1 ) ......... ... ( I , 2 , 1 ) ......... ... ( 1 , I − I ) ......... ... ( I , I − I )( 1 , I , I ) ......... ... ( I , I , I ) 12 ... I · · · I ··· I Fig. 3: Tensorization of scalar variables on a 3D coordinatesystem to form an order-3 tensor. Each variable possesses acoordinate -tuple. Owing to the condition in (11), the tensor X is saidto exhibit a Kronecker separable covariance structure, andtherefore has the following statistical properties [15, 16] E (cid:8) (cid:107) X (cid:107) (cid:9) = σ (14) E ¶ X ( n ) X T ( n ) © = σ Θ ( n ) (15) E (cid:8) xx T (cid:9) = σ Å ⊗ n = N Θ ( n ) ã (16)where x = vec ( X ) and (cid:107) · (cid:107) denotes the Frobenius norm.Intuitively, σ is the average variance over all scalar-valuedvariables within the tensor, and Θ ( n ) ∈ R I n × I n is the mode- n covariance density matrix, which has the unit-trace property tr Ä Θ ( n ) ä = 1 , ∀ n . The covariance density, Θ ( n ) , designatesthe proportion of the total variance, σ , allocated to each fibre.The defining feature of the Kronecker separability conditionis that the covariance matrix, Σ , is characterised in terms of fiber-to-fiber (multilinear) covariance parameters, in contrast toelement-to-element (linear) covariance parameters, as impliedby the multivariate normal distribution.In addition, the separability condition provides a stable andparsimonious alternative to an unrestricted estimate of Σ , thelatter being unstable or even unavailable if the dimensionsof the sample tensor are large compared to the number ofsamples. Remark 2.
Consider the covariance matrix, Σ ∈ R K × K ,which constitutes of (cid:0) K + K (cid:1) distinct parameters. In turn,the Kronecker separable counterpart, σ Ä ⊗ n = N Θ ( n ) ä , re-duces to (cid:80) Nn =1 (cid:0) I n + I n (cid:1) distinct parameters. Referringback to Remark 1, consider the case where we observe fixedincome returns for I m = 15 maturities and I c = 8 countries,that is, for K = 120 returns in total. Then, the multivariatecovariance matrix will have (cid:0) + 120 (cid:1) = 7260 distinctparameters, whereas the Kronecker separable counterpart re-duces the model to (cid:0) + 15 + 8 + 8 (cid:1) = 157 . Thepractical utility of such a parameter reduction is evident. D. Multilinear principal component analysis
Consider the eigendecomposition of the mode- n covariancedensity matrix, that is Θ ( n ) = U ( n ) Λ ( n ) U ( n ) T (17)where U ( n ) ∈ R I n × I n is the mode- n eigenvector matrix,and Λ ( n ) ∈ R I n × I n is the mode- n eigenvalue matrix, whichis diagonal and has unit-trace property, tr Ä Λ ( n ) ä = 1 , ∀ n .Following from the properties of the Kronecker product [17],we can express the Kronecker separability conditions in (15)-(16) as follows E ¶ X ( n ) X T ( n ) © = σ U ( n ) Λ ( n ) U ( n ) T (18) E (cid:8) xx T (cid:9) = σ Å ⊗ n = N U ( n ) Λ ( n ) U ( n ) T ã (19)We refer to this result as the multilinear PCA of the tensor-valued random variable X . This result is intrinsically linkedto the well-known multilinear singular value decomposition,also referred to as the Tucker decomposition [18, 19]. III. G
LOBAL F IXED I NCOME F ACTOR A NALYSIS
We next proceed to develop the multilinear model forglobal fixed income returns. The data structures considered inthe sequel are order- tensors, that is, matrix-valued randomvariables, denotes by X ∈ R I × I . The mode- n unfoldings oforder- tensors reduce to X (1) = X ∈ R I × I (20) X (2) = X T ∈ R I × I (21)Although we focus our attention to order-2 tensors, it isimportant to note that the multilinear algebra tools provided inthe previous section, and the model we develop next, naturallygeneralise to tensors of any order N . A. Data preparation
Consider the i.i.d. random variable, x t ( m, c ) ∈ R , whichrepresents the return of a fixed income asset with maturity m and from the country c at a time instant t . For simplicity, wecan assume that the return is distributed according to x t ( m, c ) ∼ N (cid:0) , σ ( m, c ) (cid:1) (22)which is dependent on the maturity and country.When jointly considering the returns of I m maturities and I c countries, we can tensorize the collection of returns ateach time instant t to form the order- tensor-valued randomvariable X t ∈ R I m × I c , given by [ X t ] ij = x t ( m i , c j ) , i = 1 , .., I m , j = 1 , ..., I c , (23)Each tensor has I m I c returns points in total.For clarity, it is important to understand the physical mean-ing of the vector and matrix representations of the consideredtensor. First, define the i -th maturity fibre, f ( m ) i ∈ R I m , and i -th country fibre, f ( c ) i ∈ R I c , respectively as follows f ( m ) i = x ( m , c i ) x ( m , c i ) ...x ( m I m , c i ) , f ( c ) i = x ( m i , c ) x ( m i , c ) ...x ( m i , c I c ) (24)In other words, the i -th maturity fibre, f ( m ) i , contains the returnof all maturities associated to the i -th country (the returnsof an entire domestic curve). In turn, the i -th country fibre, f ( c ) i , contains the returns of all countries associated to the i -thmaturity.With that, the tensor can be expressed in terms of thematurity fibres, or equivalently through the country fibres, asfollows X = î f ( m )1 f ( m )2 · · · f ( m ) I c ó = f ( c ) T f ( c ) T ... f ( c ) T I m (25)Refer to Figure 4 for an illustrative description. X = m m ... m I m c c · · · c I c = c c · · · c I c = m m ... m I m Fig. 4: Illustration of a tensor-valued sample, represented interms of individual returns (left panel), maturity fibres, f ( m ) i ,(middle panel) and country fibres, f ( c ) i , (right panel).Similarly, the vector representation can be visualised, andwritten, in terms of maturity fibres (domestic curves), that is x = c ... c I m = f ( m )1 f ( m )2 ... f ( m ) I c (26) B. Kronecker separability assumptions
We next show that the multilinear algebra allows us to nat-urally decompose the multivariate covariance matrix into twoparallel covariance matrices – maturity-domain and country-domain covariance. This is possible owing to the definingKronecker separability feature of tensor-valued models in (14)-(16), which in this case reduce to E (cid:8) (cid:107) X (cid:107) (cid:9) = σ (27) E (cid:8) XX T (cid:9) = σ Θ ( m ) (28) E (cid:8) X T X (cid:9) = σ Θ ( c ) (29) Σ = E (cid:8) xx T (cid:9) = σ Ä Θ ( c ) ⊗ Θ ( m ) ä (30)Intuitively, σ is the average variance of all fixed incomereturns, Θ ( m ) ∈ R I m × I m is the maturity-domain covariancedensity matrix, and Θ ( c ) ∈ R I c × I c is country-domain covari-ance density matrix.Using the tensor representation in (25) based on fibres, itis clear that Θ ( m ) and Θ ( c ) respectively describe the averagematurity-to-maturity and country-to-country covariance, sincewe can employ the total expectation theorem [20] to show that E (cid:8) XX T (cid:9) = E i ¶ E ¶ f ( m ) i f ( m ) T i ©© (31) E (cid:8) X T X (cid:9) = E i ¶ E ¶ f ( c ) i f ( c ) T i ©© (32)where E i {·} denotes the expectation over the indices i . Equiv-alently, we can inspect the elements of the covariance matrices [ E (cid:8) XX T (cid:9) ] kl = E i { cov { x ( m k , c i ) , x ( m l , c i ) }} (33) [ E (cid:8) X T X (cid:9) ] kl = E i { cov { x ( m i , c k ) , x ( m i , c l ) }} (34)Therefore the statistic [ E (cid:8) XX T (cid:9) ] kl describes the expected return covariance between all assets with maturity m k and m l . Similarly, [ E (cid:8) X T X (cid:9) ] kl describes the expected returncovariance between all assets within countries c k and c l . C. Implied domestic and cross-country dynamics
Equipped with the Kronecker separability conditions, wenext investigate the domestic and cross-country fixed incomereturn interactions implied by the multilinear model.We begin by observing that the multivariate covariancematrix is in fact a block matrix of domestic and cross-countryfixed income return covariance matrices, that is Σ = Σ Σ · · · Σ I c Σ Σ . . . ...... . . . . . . ... Σ I c · · · · · · Σ I c I c (35)where Σ ii ∈ R I m × I m is the domestic covariance matrix ofthe i -th country, and Σ ij ∈ R I m × I m is the cross-countrycovariance matrix between countries i and j .From the Kronecker separability condition in (30), wecan also express the multivariate covariance matrix as thefollowing block matrix Σ = σ θ ( c )11 Θ ( m ) σ θ ( c )12 Θ ( m ) · · · σ θ ( c )1 I c Θ ( m ) σ θ ( c )21 Θ ( m ) σ θ ( c )22 Θ ( m ) . . . ...... . . . . . . ...σ θ ( c ) I c Θ ( m ) · · · · · · σ θ ( c ) I n I c Θ ( m ) (36)where θ ( c ) ij is the ( i, j ) -th element of Θ ( c ) . Therefore themultilinear model asserts that each domestic and cross-countrycovariance matrix takes the form Σ ij = σ θ ( c ) ij Θ ( m ) (37)The multilinear model implicitly assumes that the maturity-domain covariance, given by Θ ( m ) , is identical for all coun-tries. This is to say that all countries exhibit the same level , slope and curvature factors, as is demonstrated empirically inthe Section V-B. Moreover, the term θ ( c ) ij simply scales thevariance parameter, σ , so as to match the observed cross-country variance, i.e. tr ( Σ ij ) = σ θ ( c ) ij . D. Multilinear factor analysis
It is natural to next evaluate and interpret the orthogonalbases spanned by the maturity-domain and country-domaincovariance density matrices, which are obtained through thefollowing eigendecompositions Θ ( m ) = U ( m ) Λ ( m ) U ( m ) T (38) Θ ( c ) = U ( c ) Λ ( c ) U ( c ) T (39)The maturity-domain eigenvector matrix, U ( m ) ∈ R I m × I m ,contains vectors u ( m ) i ∈ R I m which describe orthogonaldirections in maturity-to-maturity covariance. These vectorsrepresent the well-known level, slope and curvature factors.Similarly, U ( c ) ∈ R I c × I c contains vectors u ( c ) i ∈ R I c whichdescribe orthogonal directions in country-to-country covari-ance. These also have economic meaning, as is shown in thesequel. Note that the eigenvector matrices are orthogonal, that is, U ( m ) T U ( m ) = I and U ( c ) T U ( c ) = I .The associated eigenvalue matrices, Λ ( m ) ∈ R I m × I m and Λ ( c ) ∈ R I c × I c , respectively hold the eigenvalues, λ ( m ) i and λ ( c ) i , which describe the fraction of the total variance, σ ,explained by factor u ( m ) i and u ( c ) i . As such, the eigenvaluessum up to unity, tr Ä Λ ( m ) ä = tr Ä Λ ( c ) ä = 1 .Furthermore, we can describe the eigenvectors and eigen-values of the multivariate covariance matrix, Σ , in termsof maturity-domain and country-domain spectral parameters.Upon decomposing the covariance matrix as Σ = UΛU T (40)we can show that the following relationships hold U = Ä U ( c ) ⊗ U ( m ) ä (41) Λ = σ Ä Λ ( c ) ⊗ Λ ( m ) ä (42)By inspecting each eigenvector u i ∈ R I m I c within U ∈ R I m I c × I m I c , as well as its corresponding eigenvalue λ i , wecan obtain the following relationships u i = Ä u ( c ) k ⊗ u ( m ) l ä = u ( c ) k u ( m ) l u ( c ) k u ( m ) l ...u ( c ) kI c u ( m ) l (43) λ i = λ ( c ) k λ ( m ) l (44)where u ( c ) kj is the j -th element in the vector u ( c ) k , and owingto the Kronecker properties, i = ( k − I m + l . Remark 3.
Notice the repeated pattern in (43). The Kroneckerseparable structure asserts that, for a given factor u i , thematurity-domain structure within each country is the same, andequal to u ( m ) l . This follows from the result in (37). Another,less obvious, property is that the country-domain structure isthe same across any maturity, and equal to u ( c ) k . E. Estimation procedure
Unlike existing approaches for joint term structure model-ing, the proposed framework is entirely analytical, and so is theestimation procedure. The estimate of the variance parameter, σ , over T time instants is given by σ = 1 T − T (cid:88) t =1 (cid:107) X t (cid:107) (45)Similarly, the maturity-domain and country-domain covariancedensity matrices are obtained as follows Θ ( m ) = 1 σ ( T − T (cid:88) t =1 X t X T t (46) Θ ( c ) = 1 σ ( T − T (cid:88) t =1 X T t X t (47)These are the maximum likelihood estimators of the tensor-valued Gaussian distribution, which have been shown to bestatistically consistent [16]. IV. G
LOBAL P ORTFOLIO M ANAGEMENT AND H EDGING
The direct application of domestic principal components forevaluating market risk and constructing hedged portfolios hasbeen widely studied and implemented in the financial industry[21, 22, 23, 24]. However, understanding the commonalitiesbetween different country’s term structures is also importantfor assessing the potential for international diversification andmanaging the risk of global fixed income portfolios.The portfolio risk measure we consider is the portfoliovariance , which is a function of the multivariate covariancematrix, Σ ∈ R I m I c × I m I c . Given a vector of portfolio weights, w ∈ R I m I c , the portfolio variance is given by σ p = w T Σw (48)We have demonstrated in the previous section that, whenconsidering an international basket of fixed income assets,the covariance, Σ , exhibits the Kronecker separable structurein (30). In light of this, we can show that if we choose ourportfolio vector to match the Kronecker separable structure,that is, if we set w = Ä w ( c ) ⊗ w ( m ) ä (49)where w ( m ) ∈ R I m and w ( c ) ∈ R I c are respectively thematurity-domain and country-domain weights, then we arriveat parsimonious and compact solutions for global portfoliomanagement. This owes to the reduction in parameters re-quired for portfolio optimization from I m I c to ( I m + I c ) . Thisevident advantage is only achieved via the decomposition ofoverall risk into parallel domain – maturity and country. A. Minimum variance portfolio
The capital-constrained portfolio which attains the mini-mum variance is obtained through the optimization problem min w σ p , s.t. w T = 1 (50)the solution of which is given by the well known minimumvariance portfolio w = Σ − T Σ − (51)Notice that for the Kronecker separable case the optimalportfolio reduces to w = Ä Θ ( c ) ⊗ Θ ( m ) ä − T Ä Θ ( c ) ⊗ Θ ( m ) ä − = Ç Θ ( c ) − T Θ ( c ) − ⊗ Θ ( m ) − T Θ ( m ) − å (52)This results asserts that the portfolio optimization can be sep-arated into parallel problems within the maturity and countrydomains. This is equivalent to solving for w ( m ) and w ( c ) independently through the following minimizations min w ( m ) w ( m ) T Θ ( m ) w ( m ) , s.t. w ( m ) T = 1 and min w ( c ) w ( c ) T Θ ( c ) w ( c ) , s.t. w ( c ) T = 1 B. Hedging
Hedging of fixed income securities remains one of themost challenging problems faced by financial institutions. Thesensitivity of a portfolio, w , with respect to a risk factor, u , issimply given by the inner product u T w . The aim is to forma portfolio which is orthogonal to the risk factor, that is, theaim is to attain u T w = 0 .Within the considered international setup, risk factors areKronecker separable, that is, u = (cid:0) u ( c ) ⊗ u ( m ) (cid:1) . Uponsetting the portfolio vector to match the Kronecker separablestructure, the risk exposure simplifies to u T w = Ä u ( c ) T w ( c ) ä Ä u ( m ) T w ( m ) ä (53)This result asserts that orthogonality, u T w = 0 , can be at-tained independently in either the maturity or country domain,and thus it is not necessary to do so in both simultaneously.In other words, orthogonality is achieved by either attaining u ( c ) T w ( c ) = 0 or u ( m ) T w ( m ) = 0 . We next consider real-world applications of this result.
1) Hedging long-term bonds:
Consider hedging a long-onlyportfolio of international long-term fixed income assets with i -th maturity (e.g. 30 years), using an international portfolioof shorter-term assets. The hedged portfolio must satisfy thefollowing constraints within the maturity domain only: δ T i w ( m ) = 1 (54) T w ( m ) = 0 (55) U ( m ) T w ( m ) = (56)where δ i ∈ R I m is a vector of zeros with the i -th elementequal to 1. Intuitively, the first condition reflects the long-only position in the asset with the i -th maturity, the secondconstrains the strategy to be self-financing, while the lastenforces orthogonality with the maturity-domain factors.
2) Hedging domestic bonds:
Conversely, consider hedginga domestic long-only portfolio within the i -th country, usingan international portfolio. The hedged portfolio must satisfythe following constraints within the country domain only: δ T i w ( c ) = 1 (57) T w ( c ) = 0 (58) U ( c ) T w ( c ) = (59)where δ i ∈ R I c . In this case, the first condition reflects thelong-only position in the i -th country, the second constrains thestrategy to be self-financing, while the last condition enforcesorthogonality with the country-domain factors.The portfolio hedging problems reduce to solving the linearsystems A ( m ) w ( m ) = b ( m ) and A ( c ) w ( c ) = b ( c ) , re-spectively. The optimal maturity-domain and country-domainweights are given by w ( m ) = A ( m )+ b ( m ) and w ( c ) = A ( c )+ b ( c ) , where ( · ) + denotes the Moore-Penrose inverseoperator. V. E
MPIRICAL A NALYSIS
We next provide an empirical analysis of the global termstructure of the international interest rate swaps (IRS) mar-ket using the proposed multilinear factor model. The datacomprised of weekly IRS rate curves , ranging in the period2015-01-01 to 2019-07-01, for eight developed economies,including Switzerland, Euro Area, United Kingdom, Japan,Australia, New Zealand, Canada and United States. Eachdomestic IRS curve consisted of swaps with maturities { , , , , , , , , , , , , , , } years. Therefore,at each time instant, we observed I m = 15 IRS returns foreach of the I c = 8 economies, resulting in I m I c = 120 dailyobservations in total. Figure 5 displays the historical IRS ratesemployed in the analysis. A glance at the collective behaviourof the historical data helps us understand the importance ofglobal factors in driving the co-movement of fixed incomesecurities across advanced economies. A. Domestic analysis
Firstly, as a complementary and preliminary assessmentof the commonality of returns within different country IRScurves, we performed a principal component analysis for eachof the eight economies independently, to obtain their dominantdomestic principal components, that is, their domestic level,slope and curvature factors. The loadings of the three leadingfactors within each domestic IRS curve is shown in Figure 5,and the percentage of variance explained by each componentis reported in Table I.In agreement with the existing literature, all economiesexhibit similar loadings across the three leading principal com-ponents. Moreover, the explanatory power of the componentsis consistent across all economies, whereby the first principalcomponent (level) explains ≈ , and the second principalcomponent (slope) explains ≈ , and the third principal com-ponent (curvature) explains ≈ of the variation in interestrate changes. We interpret this as indicating the existence ofthree leading dominant factors, that is the global level , globalslope and global curvature factors. With these preliminary andsuggestive results, we now proceed to evaluate the commonglobal risk factors with the proposed multilinear model. Economy Level Slope Curvature SF .
88 10 .
02 1 . EU .
15 4 .
78 0 . GB .
29 3 .
83 0 . JP .
04 14 .
10 2 . AU .
84 4 .
94 0 . NZ .
30 5 .
76 0 . CA .
14 5 .
74 0 . US .
30 4 .
06 0 . TABLE I: Explanatory power [%] of each principal componentfor the eight economies considered. The swap rate is the fixed interest rate that the receiver of the IRS demandsin exchange for the uncertainty of having to pay the short-term floating LIBORrate over time. R a t e [ % ] − L oa d i n g PC1 PC2 PC3 (a) Switzerland (SF) R a t e [ % ] − L oa d i n g PC1 PC2 PC3 (b) Euro Area (EU) R a t e [ % ] − L oa d i n g PC1 PC2 PC3 (c) Great Britain (GB) R a t e [ % ] − L oa d i n g PC1 PC2 PC3 (d) Japan (JP) R a t e [ % ] − L oa d i n g PC1 PC2 PC3 (e) Australia (AU) R a t e [ % ] − L oa d i n g PC1 PC2 PC3 (f) New Zealand (NZ) R a t e [ % ] − L oa d i n g PC1 PC2 PC3 (g) Canada (CA) R a t e [ % ] − L oa d i n g PC1 PC2 PC3 (h) United States (US)
Fig. 5: Weekly swap rates for each economy with maturi-ties { , , , , , , , , , , , , , , } years (respec-tively coloured from blue to red) during the period 2015-01-01to 2019-07-01 (left panel) and their corresponding level, slopeand curvature components obtained from the PCA of the swapweekly returns (right panel). Source: Bloomberg.
B. Global analysis
In this section we evaluate the results obtained from mul-tilinear analysis of the international IRS dataset. The imple-mentation procedure is summarised as follows :(i) The weekly IRS returns at the t -th week were tensorizedto form the matrix-valued sample, X t ∈ R I m × I c , asdescribed in (23);(ii) The parameters of the model ( σ , Θ ( m ) , Θ ( c ) ) wereestimated using the analytic estimators in (45)-(47);(iii) The global maturity-domain and country-domain factors, U ( m ) and U ( c ) , and their associated eigenvalues, Λ ( m ) and Λ ( c ) , were obtained from the eigendecompositionsof Θ ( m ) , and Θ ( c ) , as shown in (38)-(39).The loadings of the three leading maturity-domain factors, { u ( m ) i } i =1 , are plotted in Figure 6(a), and their correspondingexplanatory powers, { λ ( m ) i } i =1 , are presented in Table II. Theinterpretation of the maturity-domain loadings is analogous tothat of traditional domestic PCA. The maturity-domain factorloadings resemble the components obtained from domesticprincipal components (see Figure 5), and therefore confirm theexistence of a common set of bases shared by all economies.Furthermore, the explanatory powers of these factors are inline with that observed from the domestic analyses, whichfurther strengthens the validity of our findings. The obtainedmaturity-domain factors clearly serve as a stencil for describ-ing the term structure within each domestic IRS curve, and assuch we refer to these as the global level , global slope and global curvature .Additionally, the country-domain factors loadings, { u ( c ) i } I c i =1 , are visualized in Figures 6(b)–6(c), and theircorresponding explanatory powers, { λ ( c ) i } I c i =1 , are presentedin Table III. The most dominant factor, u ( c )1 , has positiveloadings across all economies, and can be thought of asthe global risk premium , analogous to the level factor inthe maturity-domain. Notice that this factor also explainsa significant portion of the international IRS variance. Theremaining country-domain factors represent interpretablemacroeconomic factors concerning subsets of the consideredeconomies. These results demonstrate the direct applicabilityof the proposed approach for gaining physical insight into theglobal macroeconomic environment in a straightforward andcompact manner, owing to the small number of parametersrequired to fully describe the global fixed income universe.With reference to the previous section, the maturity-domain(listed in Table II) and country-domain (listed in Table III)factor loadings can be directly employed for global macroe-conomic hedging and risk management. We conclude thissection by reiterating the practical advantage of the multilinearframework, namely: (i) the reduction in parameters required tooptimize the global portfolio, which for I m = 15 and I c = 8 reduces from I m I c = 120 to ( I m + I c ) = 23 portfolio weightparameters; and (ii) the parsimonious description of the globalrisk in terms of parallel maturity-domain and country-domainrisk factors that can facilitate the investor’s decision makingprocess. The data analysis was implemented using our own Python Higher-OrderTensor ToolBOX (HOTTBOX) [25]. − L oa d i n g u ( m ) u ( m ) u ( m ) (a) Maturity-domain factor loadings. S F E U G B J P AU N Z C A U S Country − L oa d i n g u ( c ) u ( c ) u ( c ) u ( c ) (b) Country-domain factor loadings (1–4). S F E U G B J P AU N Z C A U S Country − L oa d i n g u ( c ) u ( c ) u ( c ) u ( c ) (c) Country-domain factor loadings (5–8). Fig. 6: Loadings of the three leading maturity-domain globalfactors (top panel) and of the country-domain global factors(middle and bottom panels).
Factor Symbol Economicinterpretation Varianceexplained [%]1 u ( m )1 Global level . u ( m )2 Global slope . u ( m )3 Global curvature . TABLE II: Economic interpretation and explanatory power ofthe three leading global factors in the maturity-domain.
Factor Symbol Economicinterpretation Varianceexplained [%]1 u ( c )1 Global risk premium . u ( c )2 (AU, NZ) vs. rest . u ( c )3 (SF, EU, GB, JP) vs.(AU, NZ, CA, US) . u ( c )4 AU vs. rest . u ( c )5 GB vs (SF, EU) . u ( c )6 CA vs US . u ( c )7 SF vs EU . u ( c )8 JPY . TABLE III: Economic interpretation and explanatory power ofthe eight leading global factors in the country-domain.
VI. C
ONCLUSIONS
A unifying tensor-valued framework for modelling theglobal risk factors shared by multiple domestic term structureshas been introduced. By virtue of the multilinear approach(as opposed to the current “flat-view” multivariate ones),the proposed approach has been shown to decompose theoverall multivariate covariance structure of international assetreturns into maturity-domain covariance and country-domaincovariance. In this way, the proposed analysis: (i) achievesa significant reduction in the number of parameters requiredto fully describe the international investment universe; and(ii) offers a physically interpretable setting for estimating andidentifying global risk factors. As a natural extension of theproposed framework, we have derived analytic solutions toglobal hedging and portfolio management, which allows theinvestor to gain enhanced control over the portfolio risk withintwo independent domains – maturity and country. An empiricalanalysis has been performed on the interest rate swaps curvesfor eight developed economies, and the results have confirmedthe existence of common global risk factors. The results aresupported by our own Python toolbox for tensor analysis [25].Although we have focused the analysis on fixed income assets,the methodology can be generalised to any asset class.A
CKNOWLEDGMENTS
The author would like to express his sincere gratitude toDanilo P. Mandic, Vladimir Lucic and Anoosh Lachin for theirvery constructive comments.R
EFERENCES[1] R. Litterman and J. Scheinkmann, “Common Factors Affecting BondReturns.”
Journal of Fixed Income , vol. 1, pp. 54–61, 1991.[2] I. T. Jolliffe,
Principal Component Analysis . New York: Springer–Verlag, 1986.[3] A. P. Rodrigues, “Term Structure and Volatility Shocks,”
Working Paper,Federal Reserve Bank of New York , 1997.[4] J. Driessen, B. Melenberg, and T. Nijman, “Common Factors in Inter-national Bond Returns,”
Journal of International Money and Finance ,vol. 22, pp. 629–656, 2003.[5] A. Novosyolov and D. Satchkov, “Global Term Structure Modeling usingPrincipal Components Analysis,”
Journal of Asset Management , vol. 9,pp. 49–60, 2008.[6] B. Flury,
Common Principal Components and Related MultivariateModels . New York: Wiley, 1988.[7] J. Juneja, “Common Factors, Principal Components Analysis, and theTerm Structure of Interest Rates,”
International Review of FinancialAnalysis , vol. 24, pp. 48–56, 2012.[8] L. R. Tucker, “An Inter-Battery Method of Factor Analysis,”
Psychome-trika , vol. 23, pp. 111–136, 1958.[9] C. P´erignon, D. R. Smith, and C. Villa, “Why Common Factors in Inter-national Bond Returns Are Not so Common,”
Journal of InternationalMoney and Finance , vol. 26, pp. 284–304, 2007.[10] T. G. Kolda and B. W. Bader, “Tensor Decompositions and Applica-tions,”
SIAM Review , vol. 51, no. 3, pp. 455–500, 2009.[11] A. Cichocki, D. P. Mandic, A. H. Phan, C. F. Caiafa, G. Zhou, Q. Zhao,and L. De Lathauwer, “Tensor Decompositions for Signal ProcessingApplications,”
IEEE Signal Processing Magazine , vol. 145, pp. 145–163, 2015.[12] A. Cichocki, A. H. Phan, Q. Zhao, N. Lee, I. Oseledets, and D. P.Mandic, “Tensor Networks for Dimensionality Reduction and Large-Scale Optimizations. Part 1: Low–Rank Tensor Decompositions,”
Foun-dations and Trends in Machine Learning , vol. 9, no. 4–5, pp. 249–429,2017.[13] A. Cichocki, A. H. Phan, Q. Zhao, N. Lee, I. Oseledets, M. Sugiyama,and D. P. Mandic, “Tensor Networks for Dimensionality Reduction and Large-Scale Optimizations. Part 2: Applications and Future Perspec-tives,”
Foundations and Trends in Machine Learning , vol. 9, no. 6, pp.431–673, 2017.[14] N. D. Siridopoulos, L. De Lathauwer, X. Fu, K. Huang, E. E. Papalex-akis, and C. Faloutsos, “Tensor Decomposition for Signal Processing andMachine Learning,”
IEEE Transactions on Signal Processing , vol. 65,no. 13, pp. 3551–3582, 2017.[15] P. D. Hoff, “Separable Covariance Arrays via the Tucker Product, withApplications to Multivariate Relational Data,”
Bayesian Analysis , vol. 6,no. 2, pp. 179–196, 2011.[16] B. Scalzo Dees and D. P. Mandic, “A Statistically Identifiable Model forTensor-Valued Gaussian Random Variables,” arXiv:1911.02915 , 2019.[17] J. R. Magnus and H. Neudecker, “Matrix Differential Calculus withApplications to Simple, Hadamard, and Kronecker Products,”
Journalof Mathematical Psychology , vol. 29, pp. 474–492, 1985.[18] L. R. Tucker, “Some Mathematical Notes on Three-Mode Factor Anal-ysis,”
Psychometrika , vol. 31, no. 3, pp. 279–311, 1966.[19] L. De Lathauwer, B. D. Moor, and J. Vandewalle, “A MultilinearSingular Value Decomposition,”
SIAM Journal on Matrix Analysis andApplications , vol. 21, no. 4, pp. 1253–1278, 2000.[20] N. A. Weiss, P. T. Holmes, and M. Hardy,
A Course in Probability .Pearson Addison Wesley, 2005.[21] “Principles of Principal Components: A Fresh Look at Risk, Hedging,and Relative Value,”
Research Report, Solomon Smith Barney , 2000.[22] “PCA Unleashed,”
Research Report, Credit Suisse , 2015.[23] “Introducing a Relative Value Tool for Swaps,”
Research Report, Stan-dard Chartered , 2013.[24] “Market Musings – Relative Value Across the U.S. Swap Surface: APCA Approach,”