PORT 4 – Using Multifactor Models

Multifactor models are now the working language of quantitative portfolio management. Asset managers and asset owners use them to build portfolios with deliberate exposures to specific sources of risk, to decompose an active manager’s return and risk into their underlying causes, to compare the risk of equity, fixed-income and other asset class holdings on a common footing, and to size active bets relative to a benchmark. This reading explains where multifactor models come from, the theory that justifies them, the main varieties in current use, and the four practical roles they play: return attribution, risk attribution, portfolio construction and strategic portfolio decisions.

From Markowitz to a single market factor

In 1952 Markowitz proposed a framework for building portfolios by weighing every holding against the whole portfolio rather than in isolation. That framework, modern portfolio theory, modeled asset returns with a multivariate normal distribution, so that mean returns, return variances and return correlations completely describe the distribution of outcomes. Its central insight is that any correlation among asset returns below one creates room for risk reduction through diversification.

In 1964 Sharpe built on that foundation with the capital asset pricing model. The CAPM gave the field vocabulary that is still in daily use: alpha, beta and systematic risk. Its key claim is that risk which can be diversified away inside a portfolio, because it is offset by the risks of other holdings, should not earn extra expected return. Only non-diversifiable risk, systematic risk, should be priced. In the CAPM that systematic risk is captured entirely by a single number, beta, which measures how sensitive an asset’s return is to the return of the market portfolio. Differences in expected return across assets are explained by differences in market beta alone: greater market risk should carry greater expected return.

Decades of subsequent equity market evidence made clear that a single market factor is an incomplete description of risk. Models that incorporate several sources of systematic risk explain the cross-section of returns considerably better than the CAPM does on its own. Practice has moved to models built around macroeconomic surprises, or around observable company and security attributes, in place of a single market beta. The remainder of this reading works through the theory behind that shift, the major families of multifactor model, and the uses to which practitioners put them.

Ross developed arbitrage pricing theory in the 1970s as an alternative to the CAPM. Rather than committing to a single factor, APT starts from any assumed multifactor process for returns and derives the equilibrium expected return that must follow if no arbitrage opportunity is available. Suppose K factors are assumed to generate returns. The general return-generating process for asset i is then written as a linear function of the factors:

The intercept a_i is the expected return to asset i when every factor takes a value of zero. In any single period the error term keeps the model from fitting exactly, but the error is assumed to average to zero over time. A widely used variant restates Equation 1 in excess-return terms throughout, netting the risk-free rate out of both the asset’s return on the left side and out of at least one factor on the right side; the Carhart model, taken up next, is built exactly this way.

The three assumptions

APT resembles the CAPM in describing a financial market equilibrium, but it rests on markedly weaker assumptions. It requires only three:

1. A factor model describes asset returns.
2. Enough assets exist that a diversified investor can build a portfolio in which each holding’s asset-specific risk washes out almost entirely.
3. No riskless arbitrage opportunity remains available once portfolios are built this way.

An arbitrage is a risk-free operation requiring no net investment of money that nonetheless earns a positive expected profit. An arbitrage opportunity is the chance to conduct one. The first assumption does not commit to any particular number of factors. The second is empirically well supported: once a portfolio holds many stocks, each holding’s asset-specific risk contributes almost nothing to the variance of the whole portfolio. The third is simply the condition of equilibrium, that prices leave no free lunch on the table.

Given these three assumptions, the expected return on any well-diversified portfolio p is linear in its factor sensitivities:

The factor risk premium fp_j is the expected reward for holding a sensitivity of exactly 1 to factor j and exactly 0 to every other factor. A portfolio with that profile is called a pure factor portfolio for factor j. Its exact interpretation depends on the model: in the Carhart model discussed next, the market factor’s risk premium is the market’s expected excess return over the risk-free rate, while each of the other three risk premiums is the mean return of a long portfolio in one style minus the mean return of a related, opposite portfolio.

The mechanics are simple to see with a single factor. Suppose a portfolio has a sensitivity of exactly 1 to Factor 1 and 0 to every other factor. If its expected return is 12% and the risk-free rate is 4%, then 0.12 = 0.04 + fp_1, so fp_1 = 0.08, an 8% risk premium for that factor.

The APT equation only holds if no arbitrage opportunity is available. When a new portfolio’s expected return and factor sensitivity are inconsistent with the equation fitted from other well-diversified portfolios, an arbitrage becomes possible: a portfolio can be built from the existing portfolios that matches the new portfolio’s factor sensitivity but not its expected return, and the mispriced portfolio is either undervalued or overvalued relative to that replica.

The Carhart four-factor model

The Carhart model, developed by Carhart in 1997, extends the three-factor model of Fama and French to add a momentum factor, and is among the most referenced multifactor models in current equity practice. Its starting observation is empirical: three groups of stocks have historically earned higher returns than their market beta alone would predict, namely small-capitalization stocks, low price-to-book “value” stocks, and stocks whose prices have recently been rising, “momentum” stocks. The model turns each anomaly into a systematic risk factor: small minus big (SMB), high minus low (HML), and winners minus losers (WML).

Because the expected value of the error term and of alpha (assuming the four factors capture all systematic risk) is zero, the equilibrium expected-return version follows directly:

Viewed from the CAPM, size, value and momentum are anomalies, capital market regularities the theory does not explain. Viewed from the Carhart model, they are priced systematic risk factors, and exposure to them is expected to be compensated with differences in mean return. Whichever label is used, size, value and momentum remain common building blocks in equity portfolio construction and continue to serve actively in risk decomposition and return attribution, which are covered later in this reading.

Multifactor models used in practice fall into three broad groups by the type of factor they use.

• Macroeconomic factor models. The factors are surprises in macroeconomic variables that significantly explain returns, such as interest rates, inflation, business cycle risk and credit spreads. For equities, these surprises can be understood as moving either a company’s expected future cash flows or the rate used to discount those cash flows to the present.
• Fundamental factor models. The factors are company or security attributes that account for why prices differ from one stock to the next at a point in time, among them book-value-to-price, market capitalization, the price-to-earnings ratio and financial leverage.
• Statistical factor models. Statistical methods are applied to historical security returns to extract portfolios of securities, defined by portfolio weights, that explain the observed returns. Factor analysis models find the portfolios that best reproduce historical return covariances; principal components models find the portfolios that best reproduce historical return variances.

Statistical models make minimal assumptions, which is their main advantage, but the resulting factors are generally hard to interpret economically. A statistical factor with weights resembling a market index might reasonably be called a market factor, but in general no such label is available. Industry practice leans toward macroeconomic and fundamental models because they are easier to interpret and rely less on data mining, though statistical models retain real use.

Fixed-income multifactor models

The same three families translate to fixed income. A macroeconomic bond model might tie bond i’s return to surprises in inflation and GDP growth:

Fundamental bond models instead organize risk around categories unique to fixed income, commonly duration (from cash to long-dated bonds), credit (from government securities to high yield), currency (home currency versus foreign developed and emerging market currencies), and geography (specific developed and emerging markets).

A simplified sector framework divides a broad bond benchmark into short, intermediate and long government sectors plus investment-grade credit, mortgage-backed and securitized debt, and high yield. Each sector carries its own spread and duration profile, and the resulting model is at once macroeconomic, because spread over similar-duration governments tracks the growth factor, and fundamental, because duration is itself a security attribute. Historic sector weights are found by a constrained regression of portfolio returns against the sector factors, with the constraint that the weights sum to 100%.

A related fixed-income application checks whether a portfolio’s sector exposures actually match the mandate a manager claims to follow. A strategy described as intermediate-duration and investment-grade should show sector weights concentrated in government and investment-grade sectors, and should carry little or no high-yield exposure; a heavy weight in long-duration or below-investment-grade sectors would signal a mismatch between the stated mandate and the actual portfolio, regardless of what the manager’s marketing materials claim.

Risk and style factors across asset classes

A further category of multifactor approach uses risk, or style, factors that apply thematically across several asset classes at once, most commonly momentum, value, carry and volatility. These are often built the same way regardless of asset class: value as a real, inflation-adjusted yield; momentum as the prior 12-month excess return; carry as a term spread. Statistical models are the easiest of the three families to extend across asset classes, because they require no asset-class-specific tuning, whereas macroeconomic and fundamental models both need adjustment and repurposing to fit the specifics of bond investing.

In a macroeconomic factor model, returns are correlated only with surprises in variables tied to the aggregate economy. Surprise is defined generally as actual value minus predicted value:

If a forecaster expects 0.4% inflation for the month and actual inflation comes in at 0.5%, the surprise is 0.5% − 0.4% = 0.1%. Only the surprise enters the model as an independent variable, because the predicted value should already be reflected in prices, and so in expected returns, before the period begins. The intercept a_i captures the effect of the predicted values of the macro variables on the asset’s expected return; the surprise term captures new information; the error term captures whatever is left over, asset-specific risk.

Consider two factors, inflation and GDP growth. For stock i:

The coefficient b_{i1} is the factor sensitivity, or factor loading: the return contribution of a 1 percentage point surprise in inflation, holding the GDP surprise constant, and similarly for b_{i2} with respect to GDP growth. Empirically, the risk premium associated with the GDP growth factor is typically positive, while the risk premium associated with the inflation factor is typically negative, so an asset with positive inflation sensitivity, one that tends to do well when inflation surprises to the upside, commands a lower required return than an otherwise identical asset with negative inflation sensitivity, because it is valued for its inflation-hedging ability.

The same two-factor logic extends naturally to asset allocation. Different asset classes and securities tend to perform differently across combinations of growth and inflation expectations, giving a simple two-factor way to organize a portfolio’s building blocks.

Once the parameters for the individual assets in a portfolio are known, whether from regression or from a specialist consulting firm’s estimates, the portfolio’s own parameters follow as a value-weighted average of the individual assets’ parameters, with each asset’s weight equal to its share of the portfolio’s total market value.

Fundamental factor models share the general form of a macroeconomic model, but the terms mean something different. The factors are stated as returns rather than surprises, so they do not generally have an expected value of zero, and the intercept is no longer a clean estimate of expected return. Factor sensitivities are instead attributes of the security itself, rescaled into what is called a standardized beta: take the asset’s own reading on the attribute, subtract the cross-sectional average reading, and divide that gap by how much the attribute typically varies across the whole stock universe.

Standardization lets every factor be interpreted on the same scale regardless of its units. Suppose an investment has a dividend yield of 3.5%, the average dividend yield across the stock universe is 2.5%, and the cross-sectional standard deviation of dividend yields is 2%. The stock’s sensitivity to the dividend yield factor is (3.5% − 2.5%) ÷ 2% = 0.50, half a standard deviation above average. A stock exactly at the average has a sensitivity of 0; a stock one standard deviation below average has a sensitivity of −1. Binary attributes, such as industry membership, are the exception: they are represented with a dummy variable equal to 1 if the stock belongs to the industry and 0 otherwise.

A second distinction concerns the order of estimation. In macroeconomic models the factor surprise series is built first, and factor sensitivities are then estimated by regression. In fundamental models the factor sensitivities, the attributes, are specified first, and factor returns are then estimated by regression. Fundamental factor models generally fall into three broad groups: company fundamental factors tied to internal performance, such as earnings growth, earnings variability and financial leverage; company share-related factors that incorporate investor expectations directly, such as valuation multiples, dividend yield, market capitalization, share price momentum and trading activity; and macroeconomic-style factors including industry or sector membership, CAPM beta and yield curve sensitivity. Global models often classify factors as country, industry or style, where country and industry are dummy variables and style factors relate to earnings, risk and valuation.

Active managers are judged against a benchmark, and multifactor models let an analyst decompose exactly why an active portfolio’s return differed from that benchmark. The starting definition is simple:

A factor model splits active return into two pieces. The return from factor tilts is the product of the manager’s over- or underweight to each factor, relative to the benchmark, and that factor’s realized return. Security selection is whatever active return the factor tilts do not explain, reflecting the manager’s skill at overweighting outperforming securities and underweighting underperforming ones, security by security, independent of factor bets.

Fundamental factor models are the workhorse for this kind of decomposition, because their factors are thematically understandable and translate readily into a narrative for a client, in contrast to statistical factors, and because they express style choices and security characteristics in more direct detail than macroeconomic factors typically allow.

Active risk, commonly called tracking error or tracking risk, is the standard deviation of active returns:

Active return and tracking error must be stated on a consistent time basis. Assuming returns are serially uncorrelated, a daily tracking error is annualized by multiplying by the square root of 250 trading days, and a monthly tracking error by the square root of 12.

The information ratio (IR) measures mean active return per unit of active risk, playing a role for active management analogous to the Sharpe ratio for absolute returns:

If a portfolio earned a mean return of 9% while its benchmark earned 7.5%, and tracking error over the same period was 6%, the information ratio is (9% − 7.5%) ÷ 6% = 0.25. Setting a minimum acceptable information ratio, alongside a maximum acceptable tracking error, is a common way institutional investors keep a manager’s active risk and style aligned with the mandate they were hired to run.

Because variances of uncorrelated variables are additive while standard deviations are not, risk decomposition works with active risk squared rather than tracking error itself:

Active factor risk is the share of active risk squared coming from the portfolio’s different-from-benchmark exposures to the factors in the risk model. Active specific risk, also called security selection risk, is the active non-factor, or residual, risk the manager assumes; a manager expects to earn a positive average return from security selection as compensation for bearing it. For a single asset class, active specific risk is built directly from each holding’s active weight and residual risk:

Active factor risk is found indirectly, as active risk squared minus active specific risk, or directly by first building each factor’s active exposure as the active-weighted sum of every holding’s sensitivity to that factor, and then combining exposures across factors using their covariances:

Analysts turn to this decomposition to answer a specific set of questions about a manager: which active exposures contributed most to tracking error, whether the manager can articulate a rationale for the exposures assumed, whether those exposures fit the manager’s stated philosophy, and which active bets earned an adequate return for the risk taken.

Multifactor models are as useful in building a portfolio as in analyzing one after the fact. A factor portfolio is constructed with a sensitivity of exactly 1 to one target factor and 0 to every other factor, giving a manager a pure vehicle for hedging that risk or for speculating on it directly.

Multifactor models serve three broad roles in constructing portfolios. In passive management, where a fund tracks an index using a sample of its constituents, the models let an analyst replicate the index’s factor exposures rather than holding every name. In active management, they let a manager predict alpha or relative return and translate that view into a desired factor profile as part of a broader strategy. In rules-based, alternative-index management, they let a strategy tilt systematically toward factors such as size, value, quality or momentum, aiming to capture, mechanically and at low cost, exposures that were traditionally credited to manager skill.

Combining pure factor portfolios: benchmark versus risk parity weighting

A practical demonstration builds pure factor portfolios for eight common style factors, each formed by buying the top 20% of stocks and shorting the bottom 20% ranked on that factor, both legs equally weighted and rebalanced monthly.

Two simple ways to combine the eight pure factor portfolios into a single multifactor portfolio are compared. The benchmark (BM) approach weights the eight equally. The risk parity (RP) approach instead weights each factor so that it contributes an equal share of the combined portfolio’s overall risk, which requires estimating the variance-covariance matrix among the eight factor portfolios and re-optimizing periodically as that matrix rolls forward. Both approaches are constrained to be long-only, with weights on the eight factor portfolios non-negative and summing to 100%.

To avoid look-ahead bias, testing this kind of strategy uses rolling-window backtesting twice over: once to build each pure factor portfolio from the rolling factor data, and a second time to estimate the covariance matrix used to set the BM or RP weights, using only information available up to each rebalancing date, then measuring realized performance in the following out-of-sample period, and repeating the process every month across the full backtest horizon.

The broader lesson from this comparison is not that risk parity is always superior, but that a straightforward equal weighting of pure factor bets is a reasonable starting point in practice, often performing at least as well as more elaborate optimization, while risk parity’s more deliberate handling of factor correlation and volatility can materially improve risk-adjusted, rather than raw, performance.

Beyond attribution and construction, a sound multifactor model lets an investor ask a strategic question the CAPM cannot: relative to other investors, which risks am I well placed to bear, and which am I not? University endowments, with very long horizons and few near-term liabilities, often have a comparative advantage in bearing the business cycle risk of traded equities or the illiquidity of private equity, and can tilt strategic allocation toward capturing the associated risk premiums for risks that barely touch them day to day. The same endowments may simultaneously be at a comparative disadvantage in bearing inflation risk, if the costs of the activities they fund have historically outpaced average inflation.

This is a materially richer framework than the CAPM offers. Under the CAPM, every investor optimally holds only two things, the market portfolio and the risk-free asset, varying the split between them according to risk tolerance. A multifactor lens instead lets an investor tilt deliberately away from the market portfolio toward whichever risks that investor is better placed than average to bear.

The logic applies equally to individuals. An investor who depends on salary or self-employment income is exposed to business cycle risk twice over, once through employment and again through any procyclical holdings, and given that exposure might reasonably demand a larger premium before taking on more of it. An investor with independent wealth and no job-loss concern has a comparative advantage in bearing business cycle risk and might rationally tilt toward greater-than-average exposure to it, all else equal, precisely because a recession does not threaten that investor’s basic financial security. Being aware of which priced risks an investor actually faces, and by how much, is the starting point for using this framework well.

Multifactor models also improve diversification for its own sake. A portfolio’s return characteristics are typically explained better by a combination of size, value and momentum factors alongside the market factor than by the market factor alone, which is exactly the empirical motivation that produced the Carhart model earlier in this reading. Compared with single-factor thinking, a multifactor approach gives an investor a richer set of dimensions along which to search for improvements in portfolio selection, whether the goal is attribution, construction or a strategic allocation decision.