Empirical Modeling and Analytics @ Empirics.com RISK BUDGETING Yi Wang Empirical Modeling and Analytics @Empirics.com June 2000 ABSTRACT We show how well established principles of modern portfolio theory can be applied to risk budgeting at the portfolio level, at the asset class level, and at the total fund level. Specifically we describe in detail how to use a factor model of returns to quantify risk at the three levels and illustrate how quantification of risk can be used to: Monitor risk Project the impact of proposed transactions Optimize portfolio modifications Construct optimal portfolios Choose a optimal mix of asset classes We also outline the algorithms for achieving these objectives. Empirical Modeling and Analytics @ Empirics.com RISK BUDGETING WHAT IS RISK BUDGETING Risk budgeting is a relatively new innovation in the investment world. Simply put, it is a process of carefully allocating a risk budget among assets or asset classes, closely monitoring changes in the portfolio or fund’s risk level, and timely adjusting the allocation to keep the portfolio or fund’s actual risk consistent with the desired level. It builds on the acknowledgement that market conditions can and will change and on the fact that portfolio or fund managers today can be more precise in the way they acquire the risk data and analytical tools to calculate and control their risk exposure. While traditional asset allocation is based on infrequent estimates of volatility and correlation as long term averages, risk budgeting focuses on quantifying and reallocating risk in a timely fashion. In addition to keeping portfolio or fund risk on track, research shows that risk budgeting can significantly enhance portfolio or fund performance, measured by some risk-reward ratio, such as the active ratio (active return / active risk, as defined below). DEFINING RISK To allocate, monitor, and keep portfolio or fund risk on target, we need to define and quantify risk. The classical definition of investment risk is uncertainty of total returns as measured by their standard deviation. Investments with greater risk are expected to earn greater returns than less risky alternatives. Asset allocation models help investors 2 Empirical Modeling and Analytics @ Empirics.com choose the asset mix (e.g., a mix of stocks, bonds, and cash) with the highest expected return given their risk constraints. Once investors have selected a desired asset mix, they often enlist specialized asset managers to implement their investment goals. The performance of an asset manager is usually compared with a benchmark. Accordingly the risk of the portfolio is defined by its deviation from the benchmark rather than by the standard deviation of its total returns. The risk thus defined is the active risk of the portfolio, since it stems from active management of the portfolio, in contrast to passive replication of the benchmark. This definition of risk makes it clear that the portfolio manager bears the risk of deviating from the benchmark, while the benchmark risk belongs to the plan sponsor. The active risk of a portfolio is measured by its tracking error, the standard deviation of the return difference between the portfolio and the benchmark. MODELING PORTFOLIO RETURNS Typically, to develop and apply a risk budgeting model, a relationship is first established between individual security returns and a set of risk factors that drive them. This relationship forms the bridge by which market experience in the form of past returns can be applied to characterize the expected distribution of future returns. Assume that a portfolio manager’s investment universe consists of a finite set of N securities. The performance of the entire universe over the coming period can then be represented by an N 1 random vector r of individual security total returns. A return model attempts to explain the return ri on any security i in terms of broader market movements. A set of M risk factors (M << N) is chosen to represent the primary sources of 3 Empirical Modeling and Analytics @ Empirics.com risk (and return) to which the portfolio may be exposed. The extent to which security i is exposed to a particular risk factor j is modeled by a fixed factor loading fij. The 1 M row vector fi thus characterizes the exposure of security i to systematic risk. The return of any security i can be expressed in terms of the M 1 random factor vector x by M (1) ri fij x j i fix+i, j 1 where fi = {fij} is the vector of factor loadings that characterizes security i, and i is the non-systematic random error. That is, i, also referred to as residual error, is the portion of the return ri that is not explained by the systematic risk model. This reflects the possibility of events specific to security i. If we let F be the N M matrix containing one row for the factor loading vector of each of the N securities in the investment universe, and denote by the N 1 vector of non-systematic random errors, we can restate Equation (1) in matrix form, (2) r = Fx + . We can represent a given portfolio p by a 1 N allocation vector qp, which states the proportion of the market value of the portfolio allocated to each of the N securities in the investment universe. The portfolio return rp is then given by 4 Empirical Modeling and Analytics @ Empirics.com (3) rp = qpr = qpFx + qp = fpx + qp where fp = qpF is the portfolio factor loading vector that summarizes the systematic risk exposure of a portfolio as a weighted sum of the exposures of its constituent securities. QUANTIFYING PORTFOLIO RISK Of primary importance in assessing portfolio risk are the secondmoment statistics, that is, the return variances. The variances 2p and 2b of the portfolio and benchmark returns rp and rb may be expressed as (4) 2p = VAR(rp) = fpfpT + qpqpT (5) 2b = VAR(rb) = fbfbT + qbqbT where is an M M matrix, which contains the variances and covariances of the common risk factors, with jk = COV(xj, xk) and jj = VAR(xj), and is an N N matrix, which contains the variances and covariances of the security residual returns, with ij = COV(i, j) and ii = VAR(i). The total volatility of the portfolio, measured by 2p, thus has two components: the first component, measured by fpfpT, is due to its exposure to the common risk factors, while the second component, measured by qpqpT, is due to its exposure to security-specific risk. 5 Empirical Modeling and Analytics @ Empirics.com Accordingly, the total portfolio risk can be seen to be composed of systematic risk and non-systematic risk. There are no cross terms in Equations (4) and (5), due to the commonly made assumption that the error vector and the systematic factor vector x are uncorrelated and that the error terms have a mean of zero. An additional simplifying assumption is that residual terms are uncorrelated to each other. In that case, ij = COV(i, j) = 0 and becomes an N N diagonal matrix. The tracking error squared of the portfolio may be expressed as (6) 2TE = VAR(rp - rb) =( fp – fb)(fp – fb )T + (qp – qb)(qp – qb)T From this expression, it is clear that tracking error arises from mismatches of composition between the portfolio and the benchmark (qp qb, and so fp fb as well). The closer the portfolio tracks the benchmark, the smaller the value of TE. If the portfolio and the benchmark are identically composed (qp = qb, and so fp = fb as well), then rp will be identical to rb under all possible states of the world, and we will have TE = 0. Indeed this is the only way that a zero tracking error may be achieved. From Equation (6), it is also clear that, like the total risk, the risk of the portfolio relative to the benchmark has two components: one is systematic and the other is non-systematic. While the non-systematic tracking error arises because mismatches of composition lead to differences in the exposures to security-specific risk (qp qb), the systematic tracking error arises because mismatches of composition 6 Empirical Modeling and Analytics @ Empirics.com lead to differences in the sensitivities to common, systematic risk factors ( fp fb ). Traditional portfolio optimization techniques are based on quantification of total risk, as in Equation (4). Their focus is on maximizing return for a given level of total risk. Risk budgeting, however, builds on quantification of active risk, as in Equation (6), and the objective could be either minimization of active risk or maximization of active return for a given level of active risk. Quantification of tracking error allows us to translate the structural differences of the portfolio and the benchmark into a forecast of tracking error. The systematic component is projected based on the differences in risk factor exposures, while the non-systematic component is projected based on the differences in security-specific risk exposures. Conversely, by tracing tracking error back to the structural differences between the portfolio and the benchmark, we can give quantitative answers to questions such as: How different are they in exposures to market risk, to size-related risk, or to financial distress risk? How different are they in exposures to certain sectors? How different are they in exposures to security-specific risk? By how much will a given transaction increase (or decrease) systematic (or non-systematic) risk? What transactions will reduce the risk of the portfolio relative to the benchmark to the desired level without significantly increasing portfolio turnover? Quantitative answers to such questions allow us to keep the risk of the portfolio on target in an efficient way. 7 Empirical Modeling and Analytics @ Empirics.com USING HISTORICAL RISK DATA To develop and apply a risk budgeting model, we need to estimate the factor loadings of the securities in our investment universe and the variances and covariances of the risk factors and the residual returns on the individual securities. In other words, we need to estimate F, , and . Typically these parameters are estimated by analyzing historical data. Therefore the assumption is that the variances and covariances of the historical risk factor and security return realizations provide a reasonable characterization of their distribution for the coming period. Since risk factors themselves are unobservable, returns on factormimicking portfolios are usually used as proxies for risk factors. Thus, for example, we may estimate the variances and covariances of the risk factors directly from the monthly returns on the factor-mimicking portfolios in the recent past. To estimate the factor loadings of individual securities, we may run regressions of their monthly returns on the corresponding returns on the factor-mimicking portfolios. Such regressions produce as output not only the coefficients, which are our estimates of the factor loadings, but also the residual terms, from which we can calculate the variances and covariances of the residual returns on individual securities. The estimated variances and covariances of the risk factors and factor loadings of the securities provide a basis on which we estimate and analyze the systematic component of tracking error, while the estimated variances and covariances of security residual returns provide a basis on which we estimate and analyze its non-systematic component. If we estimate F, , and from monthly return data, we probably want to update our estimates monthly. With monthly updated volatility 8 Empirical Modeling and Analytics @ Empirics.com and correlation information, we are able to monitor the portfolio risk and keep it on track on a monthly basis. MONITORING PORTFOLIO ACTIVE RISK Traditionally, portfolio managers are measured by returns relative to benchmarks or peers. Little attention is paid to contribution to portfolio risk. Such measures of performance are inadequate. Many firms today use risk-adjusted measures to evaluate portfolio performance. Control of risk is the focus. A high return achieved by a series of successful but risky plays may not please a risk-conscious pension plan sponsor. A more modest return, achieved while maintaining much lower or targeted risk, might be seen as a healthier approach over the long run. This point of view can be implemented in two ways. Traditionally, it is done by adjusting performance by the amount of risk taken. Under risk budgeting, we specify in advance the acceptable level of risk for the portfolio (i.e., targeted or budgeted risk). In any case, the portfolio manager should be cognizant of the risk inherent in the active management of the portfolio and weight it against the anticipated gain in return. A risk-budgeting model based on quantification of tracking error can be used as a monitoring tool. Whatever position a portfolio manager has taken relative to the benchmark, the model will quantify how much systematic/non-systematic risk has been assumed relative to the benchmark. This helps measure the risk of the exposures taken to express a market view and may point out the potential unintended risk in the portfolio. 9 Empirical Modeling and Analytics @ Empirics.com PROJECTING EFFECT OF PROPOSED TRANSACTIONS Using a risk budgeting model based on quantification of tracking error, we can easily quantify the effect of proposed trades on the portfolio risk relative to the benchmark. Consider a one-for-one substitution, that is, selling one security in the portfolio and using the proceeds to buy another. To see how we can quantify its effect on the portfolio tracking error, let us rewrite the formula for tracking error (Equation (6)) directly in terms of the security allocation vectors qp and qb for the portfolio and the benchmark, respectively: (7) 2TE = ( qp – qb)FFT(qp – qb )T + (qp – qb)(qp – qb)T. The gradient G of 2TE with respect to qp is given by (8) G = 2( qp – qb)FFT + 2(qp – qb). The ith element of the gradient vector G, Gi, represents the sensitivity of the portfolio tracking error to change in portfolio position in security i. Now let us investigate the effect on tracking error of a small change in portfolio weights. If we denote the new portfolio allocation vector by qp’ = qp + , we can see that the tracking error of the new portfolio is: (9) ’2TE = ( qp’ – qb)FFT(qp’ – qb )T + (qp’ – qb)(qp’ – qb)T 10 Empirical Modeling and Analytics @ Empirics.com = 2TE + 2( qp – qb)FFTT + 2(qp – qb)T + FFTT + T. We can express the resulting change of tracking error in terms of the change in portfolio weights and the gradient G as (10) 2TE = ’2TE - 2TE = GT + (FFT + )T. The transaction to sell security m and purchase an equivalent amount of security l can be represented by a vector whose elements are given by (11) i l, m 0 i = x x il im where x is the size of the trade, which is the percentage market value of the purchase or the sale relative to the total portfolio market value. For this special type of transaction, with only two non-zero entries in the change vector , the expression for the change in tracking error squared can be simplified dramatically. For instance, the first term of Equation (10), which gives the first-order effect of the change in portfolio composition, can be expressed as (11) GT = N G i 1 i i x(Gl Gm ) , 11 Empirical Modeling and Analytics @ Empirics.com where Gi is the ith element of the gradient vector G, which quantifies the sensitivity of the portfolio tracking error to change in portfolio position in security i. The second term of Equation (10) incorporates the second-order effects of the transaction, as well as the cross-effects of changing positions in more than one security at a time. This term, which considers both systematic and non-systematic risk, can be simplified as well, and the change in tracking error squared can be expressed as (12) 2TE = x(Gl – Gm) + x2[(fl – fm)( fl – fm)T + ( 2 2 ) ] l m where fi and 2 are the factor loading vector and residual variance of i security i, respectively. As discussed above, we can estimate security factor loadings and residual variances and risk factor variances and covariances by analyzing historical return data. Once we have estimated those risk parameters we can use Equation (8) to calculate gradient Gi for each security. With the risk parameter and gradient information in store, we can quickly quantify the impact of any proposed transaction on the portfolio tracking error using Equation (12). OPTIMIZING PORTFOLIO MODIFICATIONS Suppose we want to find a one-for-one transaction that will reduce the portfolio risk relative to the benchmark the most. We need to determine both the optimal securities and size to trade. A procedure can be developed in a risk budgeting model that will allow us to do this quickly. 12 Empirical Modeling and Analytics @ Empirics.com For a one-for-one transaction, Equation (12) gives the change in tracking error as a quadratic function of the transaction size x, (12) 2TE = bx + ax2 where (13) b = (Gl – Gm) (14) a = [(fl – fm)( fl – fm)T + ( 2 2 ) ]. l m To find the optimal market value to trade, we set d2TE / dx = 0, and find that 2TE has a minimum at (15) x b . 2a (This is clearly a minimum because is a variance-covariance matrix, which is positive semi-definite, guaranteeing that a 0.) The resulting change in tracking error (as long as a 0) is (16) 2TEmin = b2 . 4a (Note that since 2TEmin is negative, it represents the maximum reduction in tracking error by the transaction of the selected pair of securities.) 13 Empirical Modeling and Analytics @ Empirics.com Thus, for any selected pair of securities, it is straightforward to calculate the optimal transaction size as well as the resulting tracking error. In principle, the best such transaction could be found by a brute force search across all pairs of securities in our investment universe, selecting the one that gives the largest reduction in tracking error. However, for N securities, there are (N–1)N/2 possible pairs of securities, making this a time-consuming operation. Additionally, we may have other reasons to favor one security over another. A procedure can be designed that will make the search easily and quickly. From the above equations, it is easy to see that the more negative Gl is, the lower 2TE is; and the more positive Gm is, the lower 2TE is. Thus, the most reduction in tracking error would be achieved by selling the security with the most positive gradient and using the proceeds to buy the security with the most negative gradient. Accordingly, the first step of the procedure would be to compute and sort the gradients of all the securities in our investment universe. Thus, if we wish to first pick the security to buy, the procedure presents the securities in ascending order of their gradients; if we prefer to first select the security to sell, it presents the securities in descending order of their gradients. Once we have chosen a security, the procedure computes the amount to buy (or sell) and the resulting tracking error for each of the possible substitutions between this security and all other N-1 securities. The results are then presented with the lowest tracking error first. Our discussion in this section has focused on finding a one-for-one transaction that will reduce the portfolio tracking error the most. Various procedures can be developed depending on our objectives. For example, 14 Empirical Modeling and Analytics @ Empirics.com a procedure can be designed that will help us quickly find a transaction that will increase the active ratio of the portfolio the most. CONSTRUCTING OPTIMAL PORTFOLIOS The procedures used to find an optimal transaction can be extended into models for constructing optimal portfolios. As in any portfolio optimization procedure, the first step is to choose the set of securities that may be purchased. The composition of this investable universe is critical. It should be large enough to provide flexibility in matching benchmark risk exposures, yet contain only those securities that are acceptable candidates for purchase. Once the investable universe has been determined, a portfolio optimization model begins an iterative process, searching for one-for-one transactions that will achieve our objective. In the simplest case, the objective is to minimize the risk of the portfolio relative to the benchmark. In that case, the securities are ranked in terms of reduction in tracking error per unit of each security purchased. The model indicates which security, if purchased, will lead to the steepest decline in tracking error. Once a security has been selected for purchase, the model offers a list of possible one-for-one transactions of this security with various securities in the portfolio (with the optimal transaction size for each transaction), sorted in order of possible reduction in tracking error. Thus, at each step of the iterative process, the model guides us to select the transaction that will reduce the portfolio tracking error the most. The procedure is repeated until the portfolio has reached the desired level of risk relative to the benchmark. 15 Empirical Modeling and Analytics @ Empirics.com Minimization of tracking error is the most basic application of risk budgeting. It is ideal for passive investors who want their portfolios to track the benchmark as closely as possible. Risk budgeting also can be used by those investors who hope to outperform the benchmark mainly on the basis of security selection, without expressing views on markets or sectors. For example, given a carefully selected universe of securities, a procedure can be designed to help build security picks into a portfolio with no significant systematic exposures relative to the benchmark. For more active portfolios, the objective is no longer minimization of tracking error. When minimizing tracking error, we may find that the most efficient way is to reduce the largest differences in composition between the portfolio and the benchmark. But what if, for example, the portfolio is meant to be overweighted in a particular sector to express a view. These views certainly should not be “optimized” away. However, unintended exposures need to be minimized, while keeping the intentional ones. In cases like this, a procedure can be designed that will allow the investor to keep the desired sector exposure and optimize to reduce the components of tracking error due to all other risk exposures. More complicated procedures also can be developed to maximize various portfolio risk-reward ratios, such as the active ratio. QUANTIFYING ACTIVE RISK AT THE ASSET CLASS LEVEL So far, our discussion of risk budgeting has focused on individual portfolio manager level. After choosing a mix of asset classes, a plan sponsor typically enlists a number of portfolio managers within each asset class. For such an investor, risk budgeting ultimately must be at the 16 Empirical Modeling and Analytics @ Empirics.com total fund level. So we need to quantify the total fund tracking error and identify its sources. The total fund tracking error depends on the tracking error for each asset class, which in turn depends on the tracking error taken by each of the active portfolio managers. Thus, to quantify the total fund tracking error we must first use information on the tracking error for each manager to arrive at an asset class tacking error. Table 1 illustrates how to compute the tracking error within an asset class. The example uses a hypothetical set of U.S. Large Cap managers, each of which is measured relative to the S&P 500 index. Information on the tracking error and weight of each manager is shown in the table. Table 1: U.S. Large Cap Tracking Error ================================= Manager Tracking Error (bp) Weight -------------------------------------------------------1 400 60% 2 600 40% -------------------------------------------------------Total 380* 100% ================================= * See derivation below It is important to keep in mind that the asset class tracking error depends on the correlation of active returns between managers within the asset class. As the correlation between managers increases, the investment in the asset class is taking on more exposures to the same underlying source of active risk. Consequently, the tracking error for the 17 Empirical Modeling and Analytics @ Empirics.com asset class also increases, since the sources of active risk in the asset class are becoming less diversified. Thus, one practical implication of this analysis is that investors should look for investment styles that are complementary (i.e., a correlation of active returns close to zero). By so doing, the sources of tracking error can be better diversified and the asset class tracking error reduced. For the example in Table 1, we assume that two managers have been enlisted within the U.S. Large Cap and that the correlation of their active returns is 0.25. In that case, the formula for portfolio variance applies: (Tracking Error U.S. Large Cap)2 = (0.6*400)2 + (0.4*600)2 + 2*0.25*(0.6*400)*(0.4*600) = 144000 and Tracking Error U.S. Large Cap = 380 Thus, the tracking error for the asset class is 380 basis points, which is less than the tracking error of either of the two managers. QUANTIFYING ACTIVE RISK AT THE TOTAL FUND LEVEL Once we have calculated the tracking error for each asset class in the asset mix, we can determine the tracking error for the total fund given its asset allocation. Table 2 provides an example. 18 Empirical Modeling and Analytics @ Empirics.com Table 2: Total Fund Tracking Error ==================================================== Asset Class Benchmark Tracking Error (bp) Weight ---------------------------------------------------------------------------------------U.S. Large Cap S&P 500 380 45% U.S. Small Cap RUSSELL 2000 500 25% U.S. Fixed Income Lehman Aggregate 125 20% International Equity EAFE 400 10% ---------------------------------------------------------------------------------------Total Fund 217 100% ==================================================== Just as an assumption about the correlation of active returns across managers within an asset class is required to determine the tracking error of an asset class, so, too, is an assumption required about the correlation of active returns across asset classes. This assumption can be interpreted as measuring the degree to which active managers across asset classes are exposed to the same active risk factors. The total fund tracking error in Table 2 is calculated under the assumption that active returns are uncorrelated across asset classes. The correlation of active returns across asset classes has an important impact on the total fund tracking error. Table 3 illustrates how sensitive it is to the correlation assumption. It assumes that the fund invests 60 percent of its capital in U.S. Large Cap and 40 percent in U.S. Small Cap. As the correlation between active returns of the two asset classes increases from = 0 (uncorrelated) to = 1.0 (perfectly correlated), the tracking error of the total fund increases from 303 basis points to 428 basis points. This type of sensitivity analysis is important, as it provides a sense as to the potential ranges for the predicted total fund tracking error. 19 Empirical Modeling and Analytics @ Empirics.com Table 3: Total Fund Tracking Error and Asset Class Correlation ==================================================== Asset Class Weight TE (bp) TE (bp) TE (bp) ( = 0) ( = 0.5) ( = 1.0) ---------------------------------------------------------------------------------------U.S. Large Cap 60% 380 380 380 U.S. Small Cap 40% 500 500 500 ---------------------------------------------------------------------------------------Total Fund 100% 303 371 428 ==================================================== The total fund tracking error is a figure of interest in and of itself, as it provides information about the risk of deviating from a purely passive strategy at the total fund level. However, from the perspective of adjusting risk allocation, additional information is required. In particular, we need to identify the sources of the total fund tracking error. Return to the example in Table 2. In that case, the bulk of the total fund active risk is being taken in U.S. Equity. Only a small portion is attributable to U.S. Fixed Income. Suppose we assume that U.S. Fixed Income has a positive active return. Under this assumption, the fund could possibly better diversify its total active risk away from U.S. Equity and into U.S. Fixed Income. Because U.S. Fixed Income has a tracking error (125 bps) less than one third of the tracking error of either U.S. Large Cap (380 bps) or U.S. Small Cap (500 bps), shifting funds from U.S. Equity to U.S. Fixed Income can largely reduce the total fund tracking error. 20 Empirical Modeling and Analytics @ Empirics.com RISK BUDGETING AT THE TOTAL FUND LEVEL Reiterating a point made earlier, risk budgeting ultimately must be at the total fund level. At the total fund level, risk budgeting starts with a strategic asset allocation. The process of asset allocation usually begins by determining the desired characteristics of the total risk of the fund. After the targeted total fund risk has been set, a set of relevant asset classes is selected. At this point, an optimizer based on modern portfolio theory is generally used to develop an efficient frontier that shows the return-maximizing asset allocations at various risk levels. From this efficient frontier, an asset allocation is selected that best meets the overall desired risk characteristics of the total fund and is taken as the strategic asset allocation. In principle, the investor has the option of passively investing in the underlying indices of the asset classes selected. For example, the investor can choose to invest in the S&P 500 and Russell 2000 index funds for the capital allocated in the asset classes of U.S. Large Cap and U.S. Small Cap. The strategic asset allocation in the underlying indices for the asset classes chosen involves no active risk and constitutes the strategic benchmark for the fund. Most institutional investors take active risk relative to their strategic benchmarks. Indeed, the trend is towards blending passive and active investments. Active portfolio managers are enlisted because of their presumed ability to add value at the total fund level. But, in which asset classes should active risk be taken? What percentage of the assets should be actively versus passively managed? Resolutions to such issues 21 Empirical Modeling and Analytics @ Empirics.com depend on what the targeted active risk is and what the objective function is at the total fund level. Suppose that, for a given target of total fund active risk, the objective is to maximize the active ratio, i.e., the ratio of active return to active risk, at the total fund level. In such a case, a computer model can be used to determine the optimal allocation of active risk across the asset classes and across the portfolio managers within each asset class. The procedure typically starts with some relatively arbitrary allocation of active risk and then recursively adjust it until the active ratio at the total fund level is maximized. The resulting allocation of active risk is optimal and subsequent deviations from it can be monitored by computing and analyzing the active risk at the total fund level. 22