Skip to content

Appraisal ratio

Appraisal ratio measures regression alpha per unit of residual risk. See why it differs from active return over tracking error and how model, fees and frequency matter.

The appraisal ratio is a portfolio’s regression alpha divided by the standard deviation of its regression residual. It measures model-unexplained expected return per unit of model-unexplained risk. A manager with a large alpha can still have an unremarkable appraisal ratio if the residual returns are very volatile.

Its practical use is evaluating an active component alongside exposure to a benchmark. The ratio asks how much additional expected return accompanies the risk that remains after removing the specified factor exposures. It does not say how much the portfolio beats an unadjusted index, or establish that the unexplained return is skill.

Specify the regression first

For a single benchmark, let RP,tR_{P,t} be the portfolio’s simple return, RB,tR_{B,t} the benchmark’s simple return and Rf,tR_{f,t} the matching risk-free return, all for the same observation period. Write

RP,t−Rf,t=α+β(RB,t−Rf,t)+εt,R_{P,t}-R_{f,t}=\alpha+\beta(R_{B,t}-R_{f,t})+\varepsilon_t,
with E[εt]=0\mathbb{E}[\varepsilon_t]=0 and Cov(εt,RB,t−Rf,t)=0\text{Cov}(\varepsilon_t,R_{B,t}-R_{f,t})=0. The coefficient β\beta is the portfolio’s benchmark beta, and σε\sigma_\varepsilon is its residual standard deviation. Provided 0<σε<∞0<\sigma_\varepsilon<\infty,
AR=ασε.AR=\frac{\alpha}{\sigma_\varepsilon}.
Here alpha is an intercept, not a single period’s realised residual. The time-varying unexplained component is α+εt\alpha+\varepsilon_t. Sherif’s performance-measure methodology, §3.1 gives the appraisal ratio as alpha relative to non-systematic risk.

With several factors, replace β(RB−Rf)\beta(R_B-R_f) by the specified sum of factor exposures and use the residual from that same regression. Changing the factors can change both the numerator and denominator. An omitted style exposure can appear in the residual; “unexplained” means unexplained by this model, not free of economic risk.

The attraction of scaling alpha by residual risk is clear in the Treynor–Black model: an active idea’s contribution depends on its alpha and its specific risk, rather than alpha alone. Portfolio construction also needs residual correlations across ideas, exposures and constraints. A ratio for one manager is not a complete allocation rule.

Why it can differ from the information ratio

Use information ratio here to mean expected active return divided by tracking error:
IR=E[RP−RB]σ(RP−RB).IR=\frac{\mathbb{E}[R_P-R_B]}{\sigma(R_P-R_B)}.

Some authors also use “information ratio” for alpha divided by residual risk. Check the underlying return series whenever comparing reports; the label alone does not fix the convention.

Subtracting the benchmark from the single-factor equation gives

RP−RB=α+(β−1)(RB−Rf)+ε.R_P-R_B=\alpha+(\beta-1)(R_B-R_f)+\varepsilon.

Let μB=E[RB−Rf]\mu_B=\mathbb{E}[R_B-R_f] and σB=σ(RB−Rf)\sigma_B=\sigma(R_B-R_f). Orthogonality then yields

E[RP−RB]=α+(β−1)μB,σ2(RP−RB)=(β−1)2σB2+σε2.\begin{aligned} \mathbb{E}[R_P-R_B]&=\alpha+(\beta-1)\mu_B,\\ \sigma^2(R_P-R_B)&=(\beta-1)^2\sigma_B^2+\sigma_\varepsilon^2. \end{aligned}
For a fictional annual model, take alpha of 2%, residual volatility of 8%, beta of 1.3, benchmark expected excess return of 6% and benchmark volatility of 15%. These are annual arithmetic-return inputs; no return path is being compounded. The appraisal ratio is
AR=0.02/0.08=0.25.AR=0.02/0.08=0.25.

The active-return calculation retains the extra benchmark exposure:

E[RP−RB]=0.02+0.3(0.06)=0.038,Tracking error=0.32(0.15)2+(0.08)2≈0.09179,IR=0.038/0.09179≈0.414.\begin{aligned} \mathbb{E}[R_P-R_B]&=0.02+0.3(0.06)=0.038,\\ \text{Tracking error}&=\sqrt{0.3^2(0.15)^2+(0.08)^2}\approx0.09179,\\ IR&=0.038/0.09179\approx0.414. \end{aligned}

The 3.8% expected active return includes 1.8 percentage points from the beta mismatch. Its 9.18% tracking error includes that mismatch’s variability. Neither is the regression quantity used in the appraisal ratio.

In a fictional annual model, the appraisal ratio stays at 0.25 as beta changes. At beta 1.3, expected active return is 3.8%, tracking error 9.18% and information ratio 0.414. At beta 1 both ratios are 0.25.
In a fictional annual model, the appraisal ratio stays at 0.25 as beta changes. At beta 1.3, expected active return is 3.8%, tracking error 9.18% and information ratio 0.414. At beta 1 both ratios are 0.25.
Original model illustration. Only beta changes; alpha, residual volatility, benchmark moments and residual orthogonality stay fixed. Both ratios use the same annual basis.

When β=1\beta=1 in this single-factor setup, active return is α+ε\alpha+\varepsilon, so the population ratios agree. With other beta values they generally differ, although particular inputs can produce accidental numerical equality. In a multifactor model, the corresponding structural condition is matching the benchmark’s factor exposures. Finite-sample reports can still differ because of denominator conventions or estimation methods.

Estimating a ratio is not testing an intercept

Fit the regression with an intercept to aligned portfolio, benchmark and risk-free observations. An ordinary least squares estimate is

AR^=α^sε,sε=∑t=1Tε^t2T−K−1,\widehat{AR}=\frac{\hat\alpha}{s_\varepsilon},\qquad s_\varepsilon=\sqrt{\frac{\sum_{t=1}^{T}\hat\varepsilon_t^2}{T-K-1}},

where KK counts fitted factor slopes and T>K+1T>K+1. Report the sample window, frequency and residual degrees-of-freedom convention. A residual standard deviation computed with another divisor changes the reported estimate.

The denominator is not the standard error of α^\hat\alpha. Dividing alpha by its own standard error produces a tt-statistic, which addresses estimation uncertainty. The appraisal ratio scales by the volatility of residual returns. A short sample may produce a high ratio with very imprecise alpha. Selection from many backtests adds another reason a large fitted ratio may fail on later data.

Use consistently gross or net portfolio returns and identify the costs included. A flat, deterministic per-period fee subtracts from the intercept without changing residual variance in this additive model. Performance fees, changing financing costs and trading costs can also change the residual series; recomputing from net returns is more informative than mechanically subtracting an annual fee from a reported ratio.

Frequency and tail risk remain part of the result

A monthly appraisal ratio is not numerically comparable with an annual one. Under a stable additive model with serially uncorrelated residuals, mm periods give alpha mαm\alpha and residual volatility m σε\sqrt m\,\sigma_\varepsilon, hence

ARm periods=m ARone period.AR_{m\text{ periods}}=\sqrt m\,AR_{\text{one period}}.
For example, monthly alpha of 0.15% and residual volatility of 2.5% give a monthly ratio of 0.06 and an annualised ratio of 12(0.06)≈0.208\sqrt{12}(0.06)\approx0.208 under these assumptions. This is an additive annualisation convention, not compounding alpha as a guaranteed investment return. Serial dependence adds covariance terms to aggregate residual variance; overlapping observations, smoothing or changing exposures can invalidate the square-root scaling. Sharpe’s 1994 discussion of time dependence explains the same aggregation issue for reward-to-variability ratios.

Normality is unnecessary to define the ratio when the relevant moments exist. But mean and standard deviation do not describe skewness, crash exposure or liquidity. A high estimated ratio can coexist with severe losses in states absent from the sample; if residual variance is infinite, the population ratio above is not defined. Sharpe’s mean–variance discussion also explains why two moments may be insufficient for evaluating a strategy.

For a broader comparison, SelfMadeQuant’s Risk-Adjusted Performance Measures covers appraisal, information, Sharpe and Treynor measures. A public preview is available; the full lesson requires membership.

Sources

  • Mohamed Sherif (2016), Ethical Dow Jones indexes and investment performance: international evidence, §3.1, “The Treynor Black Appraisal Ratio,” author-deposited manuscript p.7.
  • Lorenzo Naranjo, The Treynor–Black Model, “Model Setup,” “Solving the Model” and “The Information Ratio.” The notes use information ratio for alpha over residual volatility; this article explicitly distinguishes that convention from raw active return over tracking error.
  • William F. Sharpe (1994), The Sharpe Ratio, “The Ratio,” “Time Dependence,” “Related Measures” and “Scale Independence.”
Last updated
  • performance evaluation
  • regression alpha
  • residual risk
  • benchmarking
Keep connecting the ideas
← Back to the glossaryBack to top ↑