Identifying and compensating for model mis-specification in factor risk models

Abstract

Techniques for using factor risk models to more accurately estimate the risk or active risk of an investment portfolio are disclosed. Inherent modeling error in factor risk models is identified and compensated for. One or more factors are added to compensate for factors that are unspecified or unattributed in the original factor risk model and which lead to modeling error. The approach can be used with a variety of different factor risk models, and for a variety of securities. Knowledge of the risk associated with modeling error can be utilized when estimating risk or active risk using factor risk models or when constructing optimal portfolios by mean-variance optimization or other portfolio construction strategies using factor risk models.

Claims

1. A computer-based method embodied on a non-transitory computer readable storage medium comprising an improvement in the technical field of monitoring risk of an investment portfolio by unconventionally enhancing one or more correlations between possible investments predicted by an original factor risk model: electronically inputting a set of N possible investment opportunities; electronically receiving by a programmed computer an original factor risk model defined for the N possible investment opportunities, said factor risk model comprising a matrix of factor exposures, a matrix of factor covariances, and a matrix of specific covariances; electronically inputting a portfolio of actual investments whose risk is to be estimated; computing a conventional risk estimate for the portfolio of actual investments using the original factor risk model; determining by a programmed computer if the conventional risk estimate under-predicts risk and is therefore inadequate; in the case where the conventional risk estimate is inadequate, unconventionally enhancing the original correlations between the possible investments predicted by the original factor risk model by first dynamically computing an N-dimensional new factor vector representing a projection of the actual investment portfolio into a null space of the transpose of said matrix of factor exposures so that a vector inner product of said new factor and any column of said matrix of factor exposure vanishes; unconventionally creating a new modified factor risk model in which a modified matrix of factor exposures comprises a union of the original matrix of factor exposures and the new dynamic factor; computing an unconventional risk estimate for the portfolio of investments using the modified factor risk model; determining by the programmed computer the unconventional risk estimate more accurately predicts risk; electronically outputting comparison results obtained utilizing the conventional risk estimate and results obtained utilizing the unconventional risk estimate.

2. The computer-based method of claim 1 further comprising: performing backtests to evaluate comparative performance of the original factor risk model and the new modified factor risk model.

3. The computer-based method of claim 2 further comprising: producing a side by side comparison of the original factor risk model and the new modified factor risk model.

4. The computer-based method of claim 3 wherein the side by side comparison comprises: a table displaying backtest data for the original factor risk model and for the new modified factor risk model for a backtest with a maximum estimated risk.

5. The computer-based method of claim 4 further comprising: a table displaying side by side lists of assets for two portfolios of investments and percentage holdings of each asset determined for the same specified risk, one for the original factor risk model and one for the new modified factor risk model.

6. The computer-based method of claim 5 wherein said table highlights differences in top holdings of the two portfolios.

7. The computer-based method of claim 1 further comprising: a table displaying side by side lists of assets for two portfolios of investments and percentage holdings of each asset determined for the same specified risk, one for the original factor risk model and one for the new modified factor risk model.

8. The computer-based method of claim 7 wherein said table highlights differences in top holdings of the two portfolios.

Description

BRIEF DESCRIPTION OF THE DRAWINGS

(1) FIG. 1 shows a computer-based system which may be suitably utilized to implement the present invention;

(2) FIG. 2A illustrates a flowchart of a first process in accordance with the present invention;

(3) FIG. 2B illustrates a flowchart of a second process in accordance with the present invention;

(4) FIG. 3 illustrates a flowchart of a third process in accordance with the present invention; and

(5) FIGS. 4A and 4B are tables each of which illustrates a comparison of an exemplary risk factor model not utilizing the present invention with an exemplary modified risk factor model in accordance with the present invention.

DETAILED DESCRIPTION

(6) The present invention may be suitably implemented as a computer-based system, in computer software which resides on computer readable media, such as solid state storage devices, such as RAM, ROM, or the like, magnetic storage devices such as a hard disk or floppy disk media, optical storage devices, such as CD-ROM or the like, or as methods implemented by such systems and software.

(7) FIG. 1 shows a block diagram of a computer system 100 which may be suitably used to implement the present invention. System 100 is implemented as a computer 12, such as a personal computer, workstation, or server. One likely scenario is that the system of the invention will be implemented as a personal computer or workstation which connects to a server or other computer running software to implement the processes of the present invention either through a network, Internet or other connection. As shown in FIG. 1 and described in further detail below, the system 100 includes software that is run by the central processing unit of the computer 12. The computer 12 may suitably include a number of standard input and output devices, including a keyboard 14, a mouse 16, CD-ROM drive 18, disk drive 20, monitor 22, and printer 24. In addition, the computer 12 may suitably include an Internet or network connection 26 for downloading software, data and updates or for providing outputs to remote system users. It will be appreciated, in light of the present description of the invention, that the present invention may be practiced in any of a number of different computing environments without departing from the spirit of the invention. For example, the system 100 may be implemented in a network configuration with individual workstations connected to a server. Also, other input and output devices may be used, as desired. For example, a remote user could access the server with a desktop computer, a laptop utilizing the Internet or with a wireless handheld device such as a Blackberry, Treo, or the like.

(8) One embodiment of the invention has been designed for use on a stand-alone personal computer or workstation with an Intel Xeon, Pentium, or later microprocessor, using as an operating system Debian GNU/Linux 3.1 (or later versions). This embodiment of the invention employs approximately 128 MB of random-access memory.

(9) According to one aspect of the invention, it is contemplated that the computer 12 will be operated by a user, such as a portfolio manager, an investment advisor or investor in an office setting. However, if desired, it would also be possible to practice the invention with the user using an off-site computer and either loading the below-described software onto the off-site computer or connecting to a server computer running the software. In the situation in which the computer 12 is operated by an investment advisor, that advisor may receive information from a client or clients, for example, by having the client fill out a form, by conducting an interview, or the like addressing the client's requirements regarding acceptable risk, expected return, preferred investments, disfavored investments, and the like. For example, an investor may wish to maximize the return of his or her portfolio subject to having a maximum acceptable risk of 3%, prefer investment in large capitalization stocks, and not want to invest in Japanese equities, dotcom stocks, or some other investment where he or she had sustained a large loss in the past.

(10) As illustrated in FIG. 1, and as described in greater detail below, additional inputs 28 may suitably include databases of historical data for backtesting and the like, data sources for assets which may be included in portfolios, such as the asset symbols, tickers, or identification numbers, the current prices of stocks, bonds, commodities, currencies, options, other investment vehicles, and the like, client data, such as current client portfolios, risk and return data, and the like. This data may also include historical information on macroeconomic variables, such as inflation and the rates for United States Treasury bonds of various maturities, for example. It will be recognized that a wide variety of additional inputs may be provided including without limitation other complementary or supplementary portfolio modeling software, such as portfolio optimization modeling software, for example.

(11) As further illustrated in FIG. 1, and as described in greater detail below, the system outputs 30 may suitably include a measure of portfolio risk, a recommended new portfolio with an expected higher return, a recommended new portfolio having lower expected risk, or the like. The output information may appear on the monitor 22 or may also be printed out at the printer 24. The output information may also be electronically sent to a broker or some other intermediary for execution. Other devices and techniques may be used to provide outputs, as desired.

(12) In the present embodiment of the invention, software is utilized to generate a number of computer display screens for receiving inputs from, and providing outputs to, a user of the system.

(13) It is anticipated that the models of the present disclosure will be implemented in software. The software may be stored in any appropriate computer readable medium, such as RAM. The software may be executed on any appropriate computer system, such as the system 12 as shown in FIG. 1.

(14) Before addressing the detailed mathematics of a presently preferred embodiment of the invention, several advantageous aspects of the present invention are described more generally. To this end, FIG. 2A illustrates several aspects of a process 200 in accordance with the present invention. Process 200 may be suitably implemented as software or in a computer-based system, such as the system 100 of FIG. 1. In a typical prior art factor risk model, the elements are known deterministically or probabilistically. As discussed above, such models when utilized to select a portfolio may underestimate the risk of the portfolio. Simply as an example, for an investor wanting to invest in a portfolio having an acceptable level of risk of no greater than 3%, the true risk may be actually closer to 5%. For a detailed discussion of a series of actual backtests on actual portfolios, see the discussion below in connection with FIGS. 4A and 4B, and Tables 1 and 2.

(15) To address such problems, the process 200 begins at step 202, by selecting a portfolio whose risk is to be estimated. In step 204, a new factor and factor exposure are determined based on the mathematical properties of the exposure matrix of the factor risk model. In step 205, the magnitude of the new factor and factor exposure are determined for the portfolio whose risk is to be estimated. In step 206, the new factor and factor exposure determined in steps 204 and 205 are added to the fully specified factor risk model to produce a modified factor risk model. Finally, in step 208, the modified factor risk model is utilized to calculate an adjusted risk estimate for the portfolio.

(16) A second process 250 in accordance with the present invention is illustrated in FIG. 2B. Like process 200, process 250 may be suitably implemented as software or in a computer-based system, such as the system 100 of FIG. 1. In step 254, a correction to the matrix of factor covariances or the matrix of specific risk is determined based on the mathematical properties of the exposure matrix of the factor risk model. This correction is generally expressed in terms of matrices. In step 255, the magnitude for the correction is determined. In step 256, the correction multiplied by the correction magnitude is added to the fully specified factor risk model to produce a modified factor risk model. Finally, in step 258, the modified factor risk model is utilized to calculate an adjusted risk estimate for the portfolio.

(17) A third process 300 in accordance with the present invention is illustrated in FIG. 3. Like processes 200 and 250, process 300 may be suitably implemented as software or in a computer-based system, such as the system 100 of FIG. 1. The process 300 addresses a computer implemented method for determining measures of risk for investment portfolios. In step 302, a characteristic of a first investment portfolio, w, that causes a factor risk model with fixed risk factors to underestimate the risk for the first investment portfolio is identified. In step 304, a variable risk factor, or factors, which varies dependent upon the factor risk model and the first investment portfolio is determined. In step 305, the magnitude of the variable risk factor, or factors, is determined. In step 306, the variable risk factor determined in steps 304 and 305 is added to the factor risk model to compensate for said underestimated risk. In step 308, a second investment portfolio sharing the characteristic of the first investment portfolio is analyzed to determine a second modified risk model to produce a more accurate measure of risk for the second investment portfolio. Alternatively, in step 310, a sequence of portfolios is analyzed with a modified factor risk model determined for each portfolio to produce a new investment portfolio having a desired level of risk.

(18) With this background in mind, we turn to a detailed mathematical discussion of a presently preferred embodiment of the invention and its context.

(19) In a factor risk model, the asset covariance matrix is modeled as
Q=BB.sup.T+.sup.2
where

(20) Q is an NN covariance matrix

(21) B is an NM matrix of factor exposures (also called factor loadings)

(22) is an MM matrix of factor covariances

(23) .sup.2 is an NN matrix of security specific risk variances

(24) It is assumed that M<<N, and it is also assumed that an N-dimensional column vector of portfolio weights w is known where w represents the fraction of available wealth invested in each asset or security or, for active risk, the difference in weights between a managed portfolio and a benchmark portfolio.

(25) As with all estimated models, the model above for Q has modeling and estimation error. As a result, when Q is used to predict the risk of a portfolio, the resulting risk prediction may be inaccurate. For example, if risk is measured as the portfolio variance, V=w.sup.TQw, the estimate of variance V may be too small or too large for a particular portfolio w. Within the context of portfolio construction and optimal portfolio construction, portfolios are sought that minimize risk. Consequently, algorithms for portfolio construction and optimal portfolio construction may bias the portfolios they construct to include those for which the predicted risk is underestimated.

(26) In many cases, a portfolio or a mathematical family of portfolios can be identified in which the risk prediction is suspected of being underestimated. For example, consider portfolios w that lie in the null space of B.sup.T. That is to say, consider portfolios such that B.sup.Tw=0. The variance predicted for such a portfolio is V=w.sup.TQw=w.sup.TBB.sup.Tw+w.sup.T.sup.2w=w.sup.T.sup.2w. In other words, for such a portfolio, there is only the specific risk associated with .sup.2. No risk is associated with or attributable to the factors modeled by B and . This risk estimate may be low because most portfolios have some risk associated with the factors that are not identified in this particular factor risk model.

(27) Another possible family of portfolios whose risk prediction may be smaller than desired could be portfolios that are significantly different than the portfolio currently held by an investor. Even if the risk prediction of a portfolio that was very different than the current holdings was accurate, it may be desirable to increase the risk prediction associated with that portfolio in order to make such a portfolio less desirable to a portfolio construction algorithm to avoid the high transaction costs of significantly altering the portfolio.

(28) In an analogous fashion, families of portfolios can be identified in which the risk prediction is too large. However, since most portfolios are constructed to minimize risk, such portfolios are normally of less practical interest.

(29) For purposes of the discussion below, it is assumed that a family of portfolios have been identified for which it is believed that the risk estimate derived from the original factor risk model is underestimated. The goal of the present invention is to compensate for this prediction by modifying the factor risk model so that its predictions for that family of portfolios are larger. To do this, a modified factor risk model, {tilde over (Q)} is constructed which includes an additional k factors:

(30) $\tilde{Q}={\left[B{f}_1{f}_2\mathrm{.Math.}{f}_k\right]\left[\begin{array}{ccccc}& 0& 0& \mathrm{.Math.}& 0\\ 0& {}_1^2& 0& \mathrm{.Math.}& 0\\ 0& 0& {}_2^2& \mathrm{.Math.}& 0\\ \mathrm{.Math.}& \mathrm{.Math.}& \mathrm{.Math.}& \mathrm{.Math.}& \mathrm{.Math.}\\ 0& 0& 0& \mathrm{.Math.}& {}_k^2\end{array}\right]\left[B{f}_1{f}_2\mathrm{.Math.}{f}_k\right]}^T+{}^2$
where

(31) f.sub.i, i=1, . . . , k, are k N-dimensional column vectors of as yet unspecified factor exposures (also called factor loadings), and

(32) .sub.i.sup.2, i=1, . . . , k, are k scalars that scale the variance magnitude of each unspecified factor. In the less practical case of compensating for overestimated risk, the scalars .sub.i.sup.2 may be negative.

(33) The variance predicted by {tilde over (Q)} is

(34) $V=w^T\tilde{Q}w=w^{T}B{B}^{T}w+\underset{i=1}{\overset{k}{.Math.}}{{}_i^2\left(w^T{f}_i\right)}^2+w^T{}^{2}w$

(35) The goal is to strategically determine the number of unspecified factors, k, the unspecified factor exposures f.sub.i and unspecified factor variances .sub.i.sup.2 so as to appropriately alter the risk prediction for those portfolios in the family of portfolios with incorrectly predicted risk. Once each of these has been specified, then {tilde over (Q)} is fully determined and can be used to estimate risk.

(36) In order to properly define k, f.sub.i, and .sub.i.sup.2, a number of mathematical constraints are imposed. Some of these constraints may be arbitrary. For example, one may choose k=1 for ease of computation.

(37) One constraint that has been found to work well is to require the unspecified factor exposures f.sub.i to belong to the same family as the portfolios whose variance is incorrectly estimated. Notice that in the expression for the additional variance created by adding the new unspecified factors,

(38) $\underset{i=1}{\overset{k}{.Math.}}{{}_i^2\left(w^T{f}_i\right)}^2,$
there is symmetry between the portfolio w and the new unspecified factor exposures, f.sub.i. If one interchanges w and f.sub.i, one obtains the same variance. Requiring f.sub.i and w to belong to the same family further reinforces this symmetry.

(39) The constraints chosen to define k, f.sub.i, and .sub.i.sup.2 may vary widely depending on the nature of the family of portfolios whose variance is underestimated.

(40) There are a number of methods by which the number of unspecified factors and unspecified factor variances .sub.i.sup.2 can potentially be specified. The choice of the most appropriate value will depend on the purpose for which the new factor risk model is used.

(41) In the context of portfolio analysis, the number of unspecified factors and total unspecified factor variance or the individual unspecified variances themselves may be prescribed arbitrarily. For example, the total unspecified variances,

(42) $\underset{j=1}{\overset{k}{.Math.}}{}_j^2,$
could be arbitrarily set as a fixed ratio of the total specific variance of a benchmark portfolio, or a fixed fraction of the total variance, or to a predetermined constant.

(43) Alternatively, in the context of optimal portfolio construction strategies, the best choices for these parameters may be determined by running backtests on historical data to determine acceptable, strategic or optimal choices.

(44) In order to further illustrate the invention disclosed herein, the specific details are described for one particular family of portfolios whose variance is likely to be underestimated.

(45) Consider the family of portfolios that lie in the null space of B.sup.T (i.e., B.sup.T w=0). For simplicity, assume also that only one factor (k=1) is to be employed, and .sub.1.sup.2 is set equal to some fixed constant. In this example, the unspecified factor exposure f.sub.1 is required to lie in the null space of B.sup.T (i.e., B.sup.T f.sub.1=0). f.sub.1 is normalized such that f.sub.1=1, where represents a suitable norm in the N-dimensional vector space of f.sub.1. While these constraints help determine possible solutions for f.sub.1, they do not completely specify f.sub.1. The null space of B.sup.T has at least NM dimensions, so there are many possible vectors that are potential candidates for f.sub.1. The definition of f.sub.1 is finalized by requiring it to be the vector that maximizes w.sup.T{tilde over (Q)}w for any given w. This requirement is equivalent to choosing f.sub.1 to maximize (w.sup.Tf.sub.1).sup.2.

(46) For this particular example, then, Q is defined by the constraints: w prescribed and fixed k=1 .sub.1.sup.2=a fixed constant B.sup.T f.sub.1=0 f.sub.1=1 f.sub.1 maximizes (w.sup.Tf.sub.1).sup.2

(47) By selecting the f.sub.1 that maximizes the underestimated risk compensation in all cases, a portfolio optimization or construction algorithm will have difficulty finding a portfolio from the family of underestimated risk portfolios in which the risk is not compensated.

(48) This particular formulation of the problem enjoys the advantage of having a closed-form solution. In fact, the solution for f.sub.1 is the normalized projection of w into the null space of B.sup.T. To better see this, denote the projection matrix into the null space of B.sup.T by P.sub.N(B.sub.T.sub.) and define y=P.sub.N(B.sub.T.sub.)w. Then, the solution to the problem of maximizing w.sup.T{tilde over (Q)}w or equivalently (w.sup.Tf.sub.1).sup.2 as specified above is f.sub.1=y/y, and the additional variance is .sub.1.sup.2w.sup.Tf.sub.1f.sub.1.sup.Tw=.sub.1.sup.2w.sup.TP.sub.N(B.sub.T.sub.)w. The projection matrix can be written as P.sub.N(B.sub.T.sub.)=IB(B.sup.T B).sup.+B.sup.T where X.sup.+ indicates the pseudo-inverse of X and I is the identity matrix. Hence,
.sub.1.sup.2w.sup.Tf.sub.1f.sub.1.sup.Tw=.sub.1.sup.2(w.sup.Tww.sup.TB(B.sup.TB).sup.+B.sup.Tw)
Using the fact that P=P.sup.2 for projection matrices, this expression can be written as
.sub.1.sup.2w.sup.Tf.sub.1f.sub.1.sup.Tw=.sub.1.sup.2y.sup.Ty y=wB(B.sup.TB).sup.+B.sup.Tw
This gives the closed form expression for f.sub.1 in terms of w and B.

(49) This same closed-form solution can be written as a modification of the original factor risk model matrices as follows: (1) The original factor risk model factor covariance matrix E is replaced by the modified factor covariance matrix =.sub.1.sup.2(B.sup.TB).sup.+ (2) The original factor risk model specific variance matrix .sup.2 is replaced by the modified specific variance matrix (.sup.2)=.sup.2+.sub.1.sup.2I
For this particular closed-form solution, the modification of the factor risk model therefore only depends on B and the user's choice for the correction magnitude, .sub.1. The new factor, f.sub.1, need not be explicitly computed.

(50) If no closed-form solution exists for {tilde over (Q)}, a numerical solution can be found by, for example, using a general numerical optimization routine.

(51) Although the invention has been illustrated for one particular example, there are many alternative ways in which the revised factor risk model may be specified. For example, one could specify: k=2, .sub.1.sup.2+.sub.2.sup.20.1, f.sub.1.sup.Tf.sub.1=1, f.sub.2.sup.Tf.sub.2=1, f.sub.1.sup.Tf.sub.2=0, B.sup.Tf.sub.1=0, B.sup.Tf.sub.2=0, and choose f.sub.1 and f.sub.2 to maximize .sub.1.sup.2(w.sup.Tf.sub.1).sup.2+.sub.2.sup.2(w.sup.Tf.sub.2).sup.2. This particular definition may or may not have a closed-form solution.

(52) In some cases, it may be preferable to use the alternative risk estimate provided by this invention to improve portfolio construction but to continue to use the original risk model to formally estimate the risk of the portfolio in performance attribution. Since risk estimates utilized by performance attribution are often used to evaluate fund managers, it is important that the definition of risk to be used by such attribution be unambiguous. Having two risk models could potentially add ambiguity to a manager's evaluation.

(53) In such cases, it may be preferable to restate the portfolio construction and risk estimate problems solved by this invention in an alternative mathematical format. Specifically, the mathematical statement of the portfolio construction problem with adjusted risk is the following:

(54) Outer Maximization/Portfolio Construction to Determine w maximize w.sup.T (the expected return) subject to w.sup.T1=1 (the budget constraint) w.sup.TQw.sub.0.sup.2 (the original risk budget) (f.sub.1.sup.Tw).sup.2.sub.1.sup.2 (the adjusted risk budget)

(55) Inner Maximization/Risk Estimation to Determine f.sub.1 maximize (f.sub.1.sup.Tw).sup.2 subject to f.sub.1=1 B.sup.Tf.sub.1=0
where

(56) w=an N dimensional column vector of portfolio weights (typically active weights)

(57) f.sub.1=an N dimensional column vector giving the new factor

(58) =an N dimensional column vector of expected returns

(59) 1=an N dimensional column vector of all ones

(60) Q=an NN matrix of asset covariances

(61) .sub.0.sup.2=the original risk budget, still to be used for performance attribution

(62) .sub.1.sup.2=the adjusted risk budget, independent of performance attribution

(63) B=an NM matrix of asset exposures or factor loadings

(64) By solving the problem in this way, the portfolio construction process selects portfolios with limited risk associated with the factor f.sub.1 while simultaneously allowing the portfolio to have the full risk budget associated with the original risk model. In this statement of the invention, there is no adjusted risk model that can be confused with the original risk model. The term (f.sub.1.sup.Tw).sup.2 can be called a metric of model uncertainty.

(65) The underestimation of risk for portfolios lying in the null space of B.sup.T is illustrated below with a specific simplified numerical example. In this example, it is assumed that there are five stocks: A, B, C, D, and E in an investment portfolio, and the true asset-asset covariance matrix for these five stocks is given by Q.sub.true below:

(66) $Q_{\mathrm{true}}=\left[\begin{array}{ccccc}0.20& 0.065& 0.015& -0.055& -017\\ 0.065& 0.045& 0.0075& -0.0175& -0.055\\ 0.015& 0.0075& 0.045& 0.0075& 0.015\\ -0.055& -0.0175& 0.0075& 0.0289& 0.065\\ -0.17& -0.055& 0.015& 0.065& 0.1936\end{array}\right]$
where the rows and columns of Q.sub.true correspond to the assets A, B, C, D, and E in that order. In other words, the variance of A is the element in row 1, column 1, 0.20; the covariance of A and B is the element in row 1, column 2, 0.065; and so on.

(67) Assume we wish to compute the risk for a portfolio whose wealth is distributed as 0% in asset A, 23.82% in asset B, 22.47% in asset C, 13.10% in asset D, and 40.61% in asset E. In vector notation, this asset allocation corresponds to the vector

(68) $w=\left[\begin{array}{c}0\\ 0.2382\\ 0.2247\\ 0.1310\\ 0.4061\end{array}\right]$
The true variance of this portfolio is given by V.sub.true=w.sup.TQ.sub.truew=0.03641, which corresponds to a volatility of 19.08%. The volatility of a portfolio is the square root of its variance, and the risk of a portfolio is often expressed in terms of volatility rather than variance.

(69) Now consider a one-factor, factor risk model that models Q.sub.true. In this example, market capitalization of the stocks has been chosen as the factor, and it is further assumed that the market capitalization of the stocks increases from A, the smallest, to E, the largest. In order to closely match the true asset-asset covariance matrix, Q.sub.true, the following factor risk model has been chosen for purposes of illustration:

(70) Factor exposures:

(71) $B=\left[\begin{array}{c}-3\\ -1\\ 0\\ 1\\ 3\end{array}\right]$

(72) Factor-factor covariance: =[0.021]

(73) Specific risk:

(74) ${}^2=\left[\begin{array}{ccccc}0.010404& 0& 0& 0& 0\\ 0& 0.023409& 0& 0& 0\\ 0& 0& 0.0441& 0& 0\\ 0& 0& 0& 0.007396& 0\\ 0& 0& 0& 0& 0.004096\end{array}\right]$
For this one-factor model, the modeled asset-asset covariance matrix is

(75) $Q_{\mathrm{one}-\mathrm{factor}}=\left[\begin{array}{ccccc}0.1994& 0.063& 0& -0.063& -0.189\\ 0.063& 0.04441& 0& -0.021& -0.063\\ 0& 0& 0.0441& 0& 0\\ -0.063& -0.021& 0& 0.0284& 0.063\\ -0.189& -0.063& 0& 0.063& 0.1931\end{array}\right]$
Note that the elements lying along the diagonal of Q.sub.one-factor and on many of the off-diagonal elements are similar to the elements of the true asset-asset covariance matrix Q.sub.true. However, there are a number of off-diagonal elements that are not very similar in these two matrices. This dissimilarity occurs because the one-factor factor-risk model is only an approximation of the true asset-asset covariance matrix.

(76) For the one-factor risk model, the variance of the portfolio w is V.sub.one-factor=w.sup.TQ.sub.one-factorw=0.03028, which corresponds to a volatility of 17.40%. This volatility value underestimates the true value of 19.08%.

(77) When we compute B.sup.Tw we obtain the vector 1.111, so w does not fully lie in the null space of B.sup.T. Nevertheless, the approach of the present invention can be used to correct the factor risk model for this particular portfolio, w, using the null-space factor adjustment method as described above. Specifically, the following constraints are utilized: k=1 .sub.1.sup.2=1 B.sup.Tf.sub.1=0 f.sub.1={square root over (f.sub.1.sup.T.sub.1)}=0.15 f.sub.1 maximizes (w.sup.Tf.sub.1).sup.2 for the w given above.
The solution for the new factor exposure to this problem for this particular w is

(78) 0$f=\left[\begin{array}{c}0.052404\\ 0.092362\\ 0.070659\\ 0.02371\\ 0.075288\end{array}\right]$
Note that as discussed above, the factor exposure to add depends on the portfolio whose variance is to be estimated. For a different w, the factor exposure that is added may be different than the f.sub.1 determined for this example.

(79) For any given w, the corrected factor risk model, asset-asset covariance matrix, Q.sub.corrected, is not unique. For the above example, it can be calculated in terms of the projection matrix as Q.sub.corrected=Q.sub.one-factor+.sub.1.sup.2f.sub.1.sup.2[IB(B.sup.T B).sup.+B.sup.T]. For this formula, we obtain

(80) $Q_{\mathrm{corrected}}=\left[\begin{array}{ccccc}0.21178& 0.05963& 0.00000& -0.05963& -0.17888\\ 0.05963& 0.06578& 0.00000& -0.01988& -0.05963\\ 0.00000& 0.00000& 0.06660& 0.00000& 0.00000\\ -0.05963& -0.01988& 0.00000& 0.04977& 0.05963\\ -0.17888& -0.05963& 0.00000& 0.05963& 0.20547\end{array}\right]$
If, on the other hand, we calculate it using the formula Q.sub.corrected=Q.sub.one-factor+.sub.1.sup.2f.sub.1f.sub.1.sup.T, we obtain

(81) $Q_{\mathrm{corrected}}=\left[\begin{array}{ccccc}0.20215& 0.06784& 0.00370& -0.06176& -0.18505\\ 0.06784& 0.05294& 0.00653& -0.01881& -0.05605\\ 0.00370& 0.00653& 0.04909& 0.00168& 0.00532\\ -0.06176& -0.01881& 0.00168& 0.02896& 0.06479\\ -0.18505& -0.05605& 0.00532& 0.06479& 0.19876\end{array}\right]$
The values of the corrected variance and volatility are unique. For either Q.sub.corrected given above, the variance of the portfolio w is V.sub.corrected=w.sup.TQ.sub.correctedw=0.0354, which corresponds to a volatility of 18.82% which is much closer to the true value of 19.08%.

(82) The differences between an unmodified risk factor model and a modified risk factor model in accordance with the present invention with one unspecified factor have been evaluated using two series of backtests. For each backtest, portfolios were constructed using a long-short, dollar-neutral, maximize return portfolio management strategy over a period of ten years with monthly rebalancings. In each portfolio generated by the monthly rebalancings, there were between 900 and 1500 equities chosen from an asset universe of approximately 1800 equities. The actual number of equities varied from month to month depending on the events of the preceding month, such as mergers, IPOs, bankruptcies, and the like. In addition to the risk and the dollar-neutrality constraints, asset bounds of plus or minus 5% were imposed and the long and short holdings in each industry were restricted to be no more than 25%. Total long holdings were limited to 100% and turnover (buys plus sells) for each monthly rebalance was limited to 150%. In the first set of backtests, the goal was to limit risk to 3 percent. In the second set of backtests, the goal was to limit the risk to 6 percent.

(83) A single missing factor orthogonal to the existing factors was assumed. Backtesting revealed that a good value for the variance of the unspecified factor was .sub.1.sup.2=0.16 with f.sub.1=1.

(84) For each level of risk, three separate backtests were performed to evaluate the performance of the model.

(85) In the first set of backtests in which the goal was to limit risk to 3 percent, the first backtest was performed with the original factor risk model and no additional factors. At the end of the 10 year backtest with monthly rebalancings, the realized portfolio return and risk were computed. In addition, the Sharpe ratio, the ratio of the return to realized risk, was computed. In general, larger Sharpe ratios indicate superior performance.

(86) The difference in the backtest results without and with the unspecified factor can be demonstrated either by comparing the realized annual risk or by comparing the Sharpe ratio of the results. The results for the first back test are given on the first line of Table 1A which is labeled backtest 1. For this backtest, the realized annual return was 10.77% and the realized annual risk was 4.84%, giving a Sharpe ratio of 2.23.

(87) This backtest illustrates that the maximum estimated risk using the original 3% risk model is underestimated for these optimized portfolios since the realized risk is 4.84%. One possible approach to improving the prediction is to artificially reduce the risk allowed for the maximum estimated risk. For example, since risk was underestimated by approximately 38%, one could simply reduce the maximum allowable estimated risk by approximately 38% so that the realized risk would hopefully become approximately 3%, assuming all other things were constant. The second row illustrates the results of conducting this experiment. When the maximum allowable risk is set to a value 40% below 3%, or 1.8% in the table, the realized risk now becomes 2.95%, which is close to the original risk budget of 3%. For this backtest, the realized annual return was 7.25%, so that the resulting Sharpe ratio was 2.46. In terms of Sharpe ratio, this backtest was superior to the first because of the larger Sharpe ratio.

(88) Finally, rather than simply reducing the effective estimated risk budget, in the third backtest, the estimated risk budget was kept at 3%, but now the additional factor was added to the risk model. The results of this backtest are shown on the third line of Table 1A, marked backtest 3. For this backtest, the annual return was 8.32% and the annual realized risk was 3.05%, a far more accurate risk result than shown in backtest 1. The Sharpe ratio for this backtest was 2.73, a significant improvement over both the first and second backtests. Hence, for this set of three backtests, adding one new factor to the risk model was both more accurate and superior to either using the unmodified risk model or simply reducing the original budget for estimated risk.

(89) TABLE-US-00001 TABLE 1A The first set of backtests. Maximum Number of Annual Estimated Unspecified Annual Realized Sharpe Backtest Risk Factors Return Risk Ratio 1 3% 0 10.77% 4.84% 2.23 2 1.8% 0 7.25% 2.95% 2.46 3 3% 1 8.32% 3.05% 2.73

(90) A second set of backtests similar to those in Table 1A is shown in Table 1B. In this set of backtests, the nominal maximum estimated risk was 6%. Backtest 4 shows the results for the original factor risk model with a risk budget of 6%. Backtest 5 shows the same risk model with a reduced risk budget of 4% chosen to make the realized risk close to 6%. Finally, backtest 6 shows the results using an additional risk factor.

(91) In this set of backtests, the reduced risk budget, backtest 5, performs worse than the original backtest in that the observed Sharpe ratio was reduced from 2.11 to 2.04. However, as in the results shown in Table 1A, the Sharpe ratio for the risk model with one added factor was significantly higher than either of the other two tests. In addition, the prediction of realized risk was more accurate.

(92) TABLE-US-00002 TABLE 1B The second set of backtests. Maximum Number of Annual Estimated Unspecified Annual Realized Sharpe Backtest Risk Factors Return Risk Ratio 4 6% 0 16.77% 7.93% 2.11 5 4% 0 12.41% 6.08% 2.04 6 6% 1 14.84% 6.15% 2.41

(93) In order to get a better sense for the kinds of differences that occurred in the backtests described above, FIG. 4A shows a distribution 400 of the top three and bottom two holdings in two portfolios 410 and 420 that are estimated to have 3% risk. On the left, the risk estimation is made with the unadjusted factor risk model. On the right, the risk estimate is made with a factor risk model with one, dynamic factor added. As can be seen in FIG. 4A, the holdings are different for the two portfolios. With no adjustment, 1.84% of the portfolio 410 is held in shares of Public Service Enterprise Group Inc., a stock listed on the New York Stock Exchange, under the symbol PEG and 3.77% of the portfolio 410 is held in shares of Apartment Investment & Management Co., a stock listed on the New York Stock Exchange under the symbol AIV. These are the largest positive and largest negative holdings. A positive holding represents a long position in the stock and a negative holding representing a short position in the stock. For the adjusted risk model, portfolio 420 has a top positive holding of 0.51% in ticker Valero Energy Corp. (VLO) and the top negative holding is 0.70% in AIV. In the portfolio 420 for the adjusted risk model, there may be non-zero holdings in tickers PEG, BXP, CEI, and CNA which are not listed. However, these holding percentages, if any, lie between the top positive and top negative holdings listed.

(94) FIG. 4B shows the distribution 425 of the top six and bottom five holdings 430 for the portfolio 410 after the first month's rebalancing using the unmodified or unadjusted risk model. These holdings 430 include portfolio weights for tickers PEG, BXP, CEI, ED, FCE/A, TXU, HCP, AYE, NYB, CNA and AIV. FIG. 4B also shows the top six and bottom five holdings 440 for the portfolio 410 after rebalancing with the modified risk or adjusted risk model. It is noted that the holdings 440 of all of these particular tickers have decreased.

(95) An additional backtest example follows below. For each backtest in this further example, the portfolios were constructed using a long-only, fully invested, maximize return portfolio management strategy over a period of ten years with monthly rebalancings. In each portfolio generated by the monthly rebalancing, there were between 90 and 850 equities chosen from an asset universe of approximately 1000 equities. The actual number of equities available to select changed from month to month depending on the events of the preceding month such as mergers, IPOs, bankruptcies, and the like. The only constraints were the active risk constraint, and the long-only, fully invested constraints. The goal was to limit active risk versus a benchmark to 3 percent. The benchmark was chosen as the market portfolio of the top 1000 stocks by market capitalization.

(96) A single missing factor orthogonal to the existing factors was assumed, in other words, factor exposures in the null space of B.sup.T. Backtesting revealed that a good value for the variance of the unspecified factor was .sub.1.sup.2=0.04 and f.sub.1=1.

(97) Initially, a factor risk model was used without accounting for modeling error. Subsequently, modeling error was explicitly accounted for by incorporating a single unknown factor.

(98) The difference in the backtest results without and with the unspecified factor can be demonstrated either by comparing the realized annual risk or by comparing the Sharpe ratio of the results.

(99) The realized annual risk was 3.33% when modeling error was ignored whereas it was 2.92% when modeling error was explicitly accounted for. These results indicate that the risk predictions modeled using one unspecified factor are more accurate than those modeled without an unspecified factor.

(100) The Sharpe ratio of the modified risk model results was substantially higher than with the original unmodified risk model. The Sharpe ratio increased from 2.11 to 2.81. Within the context of investment portfolio analysis, larger Sharpe ratios indicate superior portfolio performance.

(101) Hence, by both measures, the use of the modified factor risk model with one unspecified factor has improved the portfolio performance. The results of this further example are shown in Table 2 below.

(102) TABLE-US-00003 TABLE 2 Maximum Number of Annual Estimated Unspecified Annual Realized Sharpe Risk Factors Return Risk Ratio 3% 0 7.92% 3.33% 2.11 3% 1 8.21% 2.92% 2.81

(103) The modified risk model of the present invention is different from both traditional statistical risk modeling approaches and standard principal component analysis. In traditional statistical risk modeling approaches and principal component analysis, a single factor risk model is determined. There is no risk model correction with a variable correction magnitude that can be used to optimally adjust the risk model and its estimates of portfolio volatility.

(104) By contrast, in the typical prior art approach, all components of the risk models are known in advance, including the probabilistic manner in which they vary. In other words, B, , and .sup.2, are all known in advance. They may vary probabilistically, but the complete specification is known in advance. For example, the variance of the asset may be normally distributed around a mean value with a known standard deviation.

(105) In the present invention, additional factor covariances and factor exposures that will be added to the factor risk model are not known in advance and cannot be determined until a specific portfolio has been selected and analyzed to determine the factor covariances and factor exposures. In this sense, the additional factor or factors added to the risk model must be determined dynamically or on the fly.

(106) While the present invention has been disclosed in the context of various aspects of presently preferred embodiments, it will be recognized that the invention may be suitably applied to other environments consistent with the claims which follow.