Given observed portfolio weights \(\bm{w}\) and estimated covariance \(\bm{Q}\), what expected returns \(\bm{\mu}\) would rationalize the investor's allocation as optimal?
Risk Model
Inverse Optimization
Portfolio Optimizer
Production
| Component | Status | Description |
|---|---|---|
| Risk Model (Covariance) | Complete | Q_emp, Q_shrink, Q_garch matrices |
| Mean-Variance Inverse | In Scope | Closed-form solution with constraints |
| Morningstar MRAR Inverse | In Scope | CRRA-based utility inversion |
| Omega-Ratio Inverse | In Scope | LPM-based utility inversion |
| Factor Premia Extraction | In Scope | Fama-MacBeth methodology |
| Portfolio Optimizer | Pending | OSQP integration |
Given observed portfolio weights \(\bm{w} \in \R^n\), covariance matrix \(\bm{Q} \in \R^{n \times n}\), and a utility function \(U: \R^n \to \R\), the inverse optimization problem seeks implied expected returns \(\bm{\mu}^* \in \R^n\) such that:
\[\bm{w} = \argmax_{\bm{w}' \in \mathcal{W}} U(\bm{w}'; \bm{\mu}^*, \bm{Q})\]where \(\mathcal{W}\) denotes the feasible weight set (e.g., \(\mathcal{W} = \{\bm{w}: \bm{1}'\bm{w} = 1, \bm{w} \geq 0\}\)).
The inverse optimization approach allows us to infer what expected returns an investor implicitly assumes given their observed allocation. This is particularly valuable when:
Different utility functions yield different implied returns for the same observed weights. A more risk-averse specification (higher \(\lambda\)) will impute higher expected returns to justify a given risky allocation.
The mean-variance utility function is defined as:
\[U_{MV}(\bm{w}) = \bm{w}'\bm{\mu} - \frac{\lambda}{2}\bm{w}'\bm{Q}\bm{w}\]where \(\lambda > 0\) is the risk aversion parameter.
For the unconstrained mean-variance problem, the first-order condition yields:
\[\frac{\partial U}{\partial \bm{w}} = \bm{\mu} - \lambda\bm{Q}\bm{w} = \bm{0}\]Solving for \(\bm{\mu}\), the implied expected returns are:
\[\boxed{\bm{\mu}^*_{MV} = \lambda \bm{Q} \bm{w}}\]The unconstrained optimization problem is:
\[\max_{\bm{w}} \; \bm{w}'\bm{\mu} - \frac{\lambda}{2}\bm{w}'\bm{Q}\bm{w}\]Taking the gradient with respect to \(\bm{w}\):
\[\nabla_{\bm{w}} U = \bm{\mu} - \lambda\bm{Q}\bm{w}\]Setting \(\nabla_{\bm{w}} U = \bm{0}\) and solving for \(\bm{\mu}\):
\[\bm{\mu} = \lambda\bm{Q}\bm{w}\]The second-order condition \(\nabla^2_{\bm{w}} U = -\lambda\bm{Q} \prec 0\) confirms this is a maximum when \(\bm{Q}\) is positive definite.
QED
With the budget constraint \(\bm{1}'\bm{w} = 1\), the Lagrangian is:
\[\mathcal{L} = \bm{w}'\bm{\mu} - \frac{\lambda}{2}\bm{w}'\bm{Q}\bm{w} - \gamma(\bm{1}'\bm{w} - 1)\]The first-order conditions are:
\[\bm{\mu} - \lambda\bm{Q}\bm{w} - \gamma\bm{1} = \bm{0}\]Solving for \(\bm{\mu}\):
\[\boxed{\bm{\mu}^* = \lambda\bm{Q}\bm{w} + \gamma^*\bm{1}}\]where \(\gamma^*\) is the Lagrange multiplier determined by auxiliary conditions.
KKT Conditions:
The Lagrange multiplier \(\gamma\) represents the marginal utility of relaxing the budget constraint. In the inverse problem, \(\gamma^*\) is a degree of freedom that can be set based on:
QED
Given external estimates of expected returns \(\tilde{\bm{\mu}}\), the implied risk aversion is:
\[\lambda^* = \frac{\bm{w}'\tilde{\bm{\mu}}}{\bm{w}'\bm{Q}\bm{w}} = \frac{\text{Expected portfolio return}}{\text{Portfolio variance}}\]Typical values of \(\lambda\) range from 1 to 10. For market portfolios, \(\lambda \approx 2.5\) is commonly used (Black & Litterman, 1992).
In the Black-Litterman framework, the equilibrium expected returns are derived as:
\[\bm{\pi} = \lambda\bm{Q}\bm{w}_{mkt}\]where \(\bm{w}_{mkt}\) represents market capitalization weights. This is precisely the Mean-Variance inverse solution, providing a theoretically grounded prior for expected returns.
The Morningstar Risk-Adjusted Return (MRAR) is defined as:
\[\text{MRAR}_\gamma = \left[\frac{1}{T}\sum_{t=1}^{T}\left(\frac{1 + r_{p,t}}{1 + r_{f,t}}\right)^{-\gamma}\right]^{-12/\gamma} - 1\]where \(r_{p,t}\) is the portfolio monthly return, \(r_{f,t}\) is the risk-free rate, and \(\gamma > 0\) controls risk sensitivity.
The MRAR is equivalent to the certainty equivalent return under Constant Relative Risk Aversion (CRRA) utility:
\[u(W) = \begin{cases} \frac{W^{1-\gamma}}{1-\gamma} & \gamma \neq 1 \\ \ln(W) & \gamma = 1 \end{cases}\]The certainty equivalent wealth \(W_{CE}\) satisfies:
\[u(W_{CE}) = \E[u(W_T)]\]which yields the MRAR formula when expressed as an annualized return.
Starting with terminal wealth \(W_T = W_0 \prod_{t=1}^{T}(1 + r_{p,t})\), the expected utility is:
\[\E[u(W_T)] = \E\left[\frac{W_T^{1-\gamma}}{1-\gamma}\right] = \frac{W_0^{1-\gamma}}{1-\gamma}\E\left[\prod_{t=1}^{T}(1 + r_{p,t})^{1-\gamma}\right]\]Assuming returns are i.i.d.:
\[\E[u(W_T)] = \frac{W_0^{1-\gamma}}{1-\gamma}\left(\E[(1 + r_p)^{1-\gamma}]\right)^T\]The certainty equivalent satisfies \(u(W_{CE}) = \E[u(W_T)]\), yielding:
\[W_{CE} = W_0 \cdot \left(\E[(1 + r_p)^{1-\gamma}]\right)^{T/(1-\gamma)}\]Converting to annualized excess return over risk-free gives the MRAR formula.
QED
The MRAR utility does not admit a closed-form inverse solution. The implied returns must be found numerically by solving:
\[\bm{\mu}^* = \argmin_{\bm{\mu}} \|\bm{w} - \bm{w}^*(\bm{\mu})\|^2\]where \(\bm{w}^*(\bm{\mu}) = \argmax_{\bm{w}'} \text{MRAR}(\bm{w}'; \bm{\mu}, \bm{\Sigma})\).
The numerical approach requires:
The Omega ratio for threshold \(L\) is defined as:
\[\Omega(L) = \frac{\int_L^{\infty}(1 - F(r))\,dr}{\int_{-\infty}^{L}F(r)\,dr} = \frac{\E[\max(R - L, 0)]}{\E[\max(L - R, 0)]}\]where \(F(r)\) is the CDF of portfolio returns and \(R\) denotes the random return.
The Omega ratio can be expressed in terms of the first lower partial moment:
\[\boxed{\Omega(L) = \frac{\E[R] - L}{\text{LPM}_1(L)} + 1}\]where \(\text{LPM}_1(L) = \E[\max(L - R, 0)]\) is the expected shortfall below threshold \(L\).
Starting from the definition:
\[\Omega(L) = \frac{\E[\max(R - L, 0)]}{\E[\max(L - R, 0)]}\]Note that \(\max(R - L, 0) - \max(L - R, 0) = R - L\), so:
\[\E[\max(R - L, 0)] = \E[R] - L + \E[\max(L - R, 0)] = \E[R] - L + \text{LPM}_1(L)\]Substituting:
\[\Omega(L) = \frac{\E[R] - L + \text{LPM}_1(L)}{\text{LPM}_1(L)} = \frac{\E[R] - L}{\text{LPM}_1(L)} + 1\]QED
Common choices for the threshold \(L\):
The Omega ratio is related to the Sortino ratio through:
\[\text{Sortino}(L) = \frac{\E[R] - L}{\sqrt{\text{LPM}_2(L)}}\]Both measures focus on downside risk but weight deviations differently: Omega uses expected shortfall (linear), Sortino uses downside deviation (quadratic).
Like MRAR, the Omega ratio requires numerical inversion. The computational approach involves:
For non-closed-form utilities, we solve the bilevel optimization:
\[\min_{\bm{\mu}} \; \|\bm{w}_{obs} - \bm{w}^*(\bm{\mu})\|^2 \quad \text{s.t.} \quad \bm{w}^*(\bm{\mu}) = \argmax_{\bm{w}'} U(\bm{w}'; \bm{\mu}, \bm{Q})\]
Input: w_obs, Q, U(.), tolerance eps, max_iter
Initialize: mu^(0) = lambda * Q * w_obs (MV starting point)
For k = 0, 1, 2, ... until convergence:
1. Solve forward problem: w^*(mu^(k)) = argmax U(w; mu^(k), Q)
2. Compute residual: r^(k) = w_obs - w^*(mu^(k))
3. Compute Jacobian: J^(k) = d w^*/d mu (via implicit function theorem)
4. Update: mu^(k+1) = mu^(k) + alpha * J^(k)' * r^(k)
5. If ||r^(k)|| < eps: break
Output: mu^* = mu^(k)
For smooth utility functions, the Jacobian \(\partial\bm{w}^*/\partial\bm{\mu}\) can be computed using the implicit function theorem applied to the first-order conditions:
\[\nabla_{\bm{w}} U(\bm{w}^*; \bm{\mu}) = \bm{0}\]Differentiating with respect to \(\bm{\mu}\):
\[\nabla^2_{\bm{ww}} U \cdot \frac{\partial\bm{w}^*}{\partial\bm{\mu}} + \nabla^2_{\bm{w\mu}} U = \bm{0}\]Solving:
\[\frac{\partial\bm{w}^*}{\partial\bm{\mu}} = -(\nabla^2_{\bm{ww}} U)^{-1} \nabla^2_{\bm{w\mu}} U\]For constrained optimization with active inequality constraints, the KKT conditions must be differentiated, accounting for:
Asset returns follow the factor structure:
\[\bm{r} = \bm{\alpha} + \bm{B}\bm{f} + \bm{\epsilon}\]where:
Under the factor model, the covariance matrix decomposes as:
\[\boxed{\bm{Q} = \bm{B}\bm{Q}_f\bm{B}' + \bm{D}}\]where \(\bm{Q}_f = \text{Cov}(\bm{f})\) is the \(k \times k\) factor covariance.
This decomposition reduces the number of parameters from \(\frac{n(n+1)}{2}\) to \(nk + \frac{k(k+1)}{2} + n\), providing more stable estimates when \(k \ll n\).
Under the Arbitrage Pricing Theory, expected returns are linear in factor exposures:
\[\bm{\mu} = r_f\bm{1} + \bm{B}\bm{\pi}_f\]where \(\bm{\pi}_f = \E[\bm{f}] - r_f\bm{1}_k\) is the vector of factor risk premia.
Key benefit: New assets only require estimation of their factor betas \(\bm{B}_{new}\). Their expected returns are automatically:
\[\mu_{new} = r_f + \bm{B}_{new}'\bm{\pi}_f\]Given implied asset-level expected returns \(\bm{\mu}^*\) from inverse optimization, we extract factor premia via regression.
The factor risk premia can be estimated as:
\[\hat{\bm{\pi}}_f = (\bm{B}'\bm{B})^{-1}\bm{B}'(\bm{\mu}^* - r_f\bm{1})\]This is the OLS solution to \(\bm{\mu}^* - r_f\bm{1} = \bm{B}\bm{\pi}_f + \bm{e}\).
For time-series data, the Fama-MacBeth (1973) procedure provides consistent estimates:
Pass 1 (Time-Series): For each asset \(i\), estimate factor betas:
\[r_{i,t} = \alpha_i + \sum_{k=1}^{K}\beta_{i,k}f_{k,t} + \epsilon_{i,t} \quad \Rightarrow \quad \hat{\beta}_{i,k}\]Pass 2 (Cross-Sectional): For each time \(t\), regress returns on betas:
\[r_{i,t} = \gamma_{0,t} + \sum_{k=1}^{K}\gamma_{k,t}\hat{\beta}_{i,k} + \eta_{i,t}\]Factor Premia: Average cross-sectional coefficients:
\[\hat{\pi}_k = \frac{1}{T}\sum_{t=1}^{T}\gamma_{k,t}, \quad \text{SE}(\hat{\pi}_k) = \frac{\text{sd}(\gamma_{k,t})}{\sqrt{T}}\]The Fama-MacBeth procedure suffers from errors-in-variables bias since \(\hat{\beta}\) is estimated. The Shanken (1992) correction adjusts standard errors:
\[\text{SE}_{corrected} = \text{SE}_{FM} \times \sqrt{1 + \hat{\bm{\pi}}'_f\hat{\bm{Q}}_f^{-1}\hat{\bm{\pi}}_f}\]A minimal example to illustrate the inverse optimization calculation with concrete numbers.
Setup: Consider a portfolio with two assets:
Step 1: Construct the covariance matrix
\[\sigma_1 = 0.20, \quad \sigma_2 = 0.05, \quad \rho_{12} = 0.2\] \[\bm{Q} = \begin{pmatrix} \sigma_1^2 & \rho_{12}\sigma_1\sigma_2 \\ \rho_{12}\sigma_1\sigma_2 & \sigma_2^2 \end{pmatrix} = \begin{pmatrix} 0.0400 & 0.0020 \\ 0.0020 & 0.0025 \end{pmatrix}\]Step 2: Apply inverse formula
Using \(\bm{\mu}^* = \lambda \bm{Q} \bm{w}\):
\[\bm{\mu}^* = 2.5 \times \begin{pmatrix} 0.0400 & 0.0020 \\ 0.0020 & 0.0025 \end{pmatrix} \times \begin{pmatrix} 0.4 \\ 0.6 \end{pmatrix}\]Step 3: Matrix multiplication
\[\bm{Q}\bm{w} = \begin{pmatrix} 0.04 \times 0.4 + 0.002 \times 0.6 \\ 0.002 \times 0.4 + 0.0025 \times 0.6 \end{pmatrix} = \begin{pmatrix} 0.0160 + 0.0012 \\ 0.0008 + 0.0015 \end{pmatrix} = \begin{pmatrix} 0.0172 \\ 0.0023 \end{pmatrix}\]Step 4: Scale by risk aversion
\[\bm{\mu}^* = 2.5 \times \begin{pmatrix} 0.0172 \\ 0.0023 \end{pmatrix} = \begin{pmatrix} 0.0430 \\ 0.0058 \end{pmatrix}\]Interpretation: To justify holding 40% equity and 60% bonds under mean-variance utility with \(\lambda = 2.5\), the investor implicitly expects equity to return 4.3% and bonds 0.58% annually. The equity premium (4.3% - 0.58% = 3.7%) compensates for its higher risk.
Verification: These implied returns can be verified by solving the forward problem. Given \(\bm{\mu}^* = (0.043, 0.0058)'\) and \(\bm{Q}\), the optimal weights from \(\bm{w}^* = \frac{1}{\lambda}\bm{Q}^{-1}\bm{\mu}^*\) should recover \(\bm{w} = (0.4, 0.6)'\).
Consider a 4-asset portfolio from BRISMA:
| Asset | ID | Weight \(w_i\) |
|---|---|---|
| Cash EUR | VJy3jN... | 25% |
| MSCI EMU | BIngkC... | 30% |
| Euro Gov | MSOM2Z... | 35% |
| CTA | gjJLI6... | 10% |
Using the shrinkage covariance \(\bm{Q}_{shrink}\) and \(\lambda = 2.5\):
\[\bm{\mu}^*_{MV} = 2.5 \times \bm{Q}_{shrink} \times \bm{w}\]Expected output:
| Asset | Implied \(\mu_i\) | Interpretation |
|---|---|---|
| Cash EUR | ~0% | Zero risk, zero premium |
| MSCI EMU | ~6-8% | Equity risk premium |
| Euro Gov | ~1-2% | Duration risk premium |
| CTA | ~3-5% | Trend/momentum premium |
Decomposing into principal components (95% variance explained ~5 factors):
\[\bm{\mu}^* = r_f\bm{1} + \bm{B}_{PCA}\bm{\pi}_{PCA}\]This allows extension to new assets by estimating their PCA betas.
The risk model produces three covariance matrices:
All matrices are symmetric positive semi-definite: \(\bm{Q} = \bm{Q}'\) and \(\bm{x}'\bm{Q}\bm{x} \geq 0\) for all \(\bm{x}\).
| Matrix | Method | Period | Use Case |
|---|---|---|---|
| Q_emp | GARCH-weighted historical | 22-day returns | Historical risk assessment |
| Q_shrink | Factor model + shrinkage | 22-day returns | Reduced estimation error |
| Q_garch | GARCH(1,1) forecast | Annualized | Forward-looking optimization |
| Parameter | Value | Description |
|---|---|---|
| Working days/year | 260.89 | \(365.2425 / 7 \times 5\) |
| Rolling window | 22 days | ~1 month |
| Lookback | 10 years | Historical window |
| Variance threshold | 95% | PCA component selection |
| GARCH horizon | 261 days | 1-year forecast |
| Symbol | Name | Definition |
|---|---|---|
| \(\bm{w}\) | Portfolio weights | Fraction in each asset, \(\bm{1}'\bm{w} = 1\) |
| \(\bm{\mu}\) | Expected returns | Annualized expected return vector |
| \(\bm{Q}\) | Covariance matrix | Variance-covariance of returns |
| \(\lambda\) | Risk aversion | Trade-off parameter in MV utility |
| \(\bm{B}\) | Factor loadings | Sensitivity to factors (betas) |
| \(\bm{f}\) | Factor returns | Returns of risk factors |
| \(\bm{\pi}_f\) | Factor premia | Expected excess return per factor |
| \(r_f\) | Risk-free rate | Return on riskless asset |
| \(\bm{D}\) | Idiosyncratic covariance | Diagonal matrix of residual variances |
| \(\gamma\) | CRRA parameter | Relative risk aversion coefficient |
| \(L\) | Threshold | Reference return for Omega ratio |
| LPM | Lower Partial Moment | \(\E[\max(L-R,0)^n]\) |
| Notation | Meaning |
|---|---|
| \(\bm{x}'\) | Transpose of vector/matrix |
| \(\bm{A}^{-1}\) | Matrix inverse |
| \(\R^n\) | n-dimensional real space |
| \(\E[\cdot]\) | Expected value |
| \(\argmax\) | Argument that maximizes |
| \(\nabla_{\bm{x}}\) | Gradient with respect to \(\bm{x}\) |
| \(\nabla^2\) | Hessian matrix |
| \(\bm{A} \succ 0\) | Positive definite matrix |
| \(\bm{1}\) | Vector of ones |
A client gives you their portfolio weights. You have their covariance matrix. But they never told you their expected returns. Can you figure out what they must believe?
In standard portfolio optimization, the workflow is:
But what if we don't know expected returns? What if we only observe what someone actually holds?
The investor's weights ARE their beliefs encoded in action. If they chose \(\bm{w}\), and we know they're rational (utility-maximizing), then we can reverse-engineer what expected returns \(\bm{\mu}^*\) would make \(\bm{w}\) optimal.
The covariance \(\bm{Q}\) encodes RISK. The weights \(\bm{w}\) encode the RISK-RETURN TRADEOFF. Together they reveal the implied RETURNS.
| Application | What We Observe | What We Learn |
|---|---|---|
| Black-Litterman Prior | Market-cap weights | Equilibrium expected returns |
| Client Analysis | Client's portfolio | Their implicit views on assets |
| Consistency Check | Any allocation | Does the implied view make sense? |
| New Asset Pricing | Existing portfolio | Factor premia to price new assets |
Let's work through a complete example with every calculation shown.
| Parameter | Equity | Bond |
|---|---|---|
| Volatility (annual) | \(\sigma_1 = 20\%\) | \(\sigma_2 = 5\%\) |
| Observed Weight | \(w_1 = 0.40\) | \(w_2 = 0.60\) |
Correlation: \(\rho_{12} = 0.2\), Risk aversion: \(\lambda = 2.5\)
The covariance between two assets is: \(\text{Cov}(r_1, r_2) = \rho_{12} \cdot \sigma_1 \cdot \sigma_2\)
\[\text{Cov}(r_1, r_2) = 0.2 \times 0.20 \times 0.05 = 0.002\]The full covariance matrix:
\[\bm{Q} = \begin{pmatrix} \sigma_1^2 & \rho_{12}\sigma_1\sigma_2 \\ \rho_{12}\sigma_1\sigma_2 & \sigma_2^2 \end{pmatrix} = \begin{pmatrix} 0.20^2 & 0.002 \\ 0.002 & 0.05^2 \end{pmatrix} = \begin{pmatrix} 0.0400 & 0.0020 \\ 0.0020 & 0.0025 \end{pmatrix}\]The key formula from mean-variance utility is:
\[\boxed{\bm{\mu}^* = \lambda \cdot \bm{Q} \cdot \bm{w}}\]Let's compute \(\bm{Q} \cdot \bm{w}\) first:
\[\bm{Q} \cdot \bm{w} = \begin{pmatrix} 0.0400 & 0.0020 \\ 0.0020 & 0.0025 \end{pmatrix} \times \begin{pmatrix} 0.40 \\ 0.60 \end{pmatrix}\]First element (Equity's contribution to implied return):
\[(\bm{Q}\bm{w})_1 = Q_{11} \cdot w_1 + Q_{12} \cdot w_2 = 0.0400 \times 0.40 + 0.0020 \times 0.60\] \[= 0.0160 + 0.0012 = 0.0172\]Second element (Bond's contribution):
\[(\bm{Q}\bm{w})_2 = Q_{21} \cdot w_1 + Q_{22} \cdot w_2 = 0.0020 \times 0.40 + 0.0025 \times 0.60\] \[= 0.0008 + 0.0015 = 0.0023\]So:
\[\bm{Q}\bm{w} = \begin{pmatrix} 0.0172 \\ 0.0023 \end{pmatrix}\]| Asset | Implied Return | As Percentage |
|---|---|---|
| Equity | \(\mu_1^* = 0.0430\) | 4.30% annual |
| Bond | \(\mu_2^* = 0.0058\) | 0.58% annual |
What does this mean? To justify holding 40% equity and 60% bonds under mean-variance utility with \(\lambda = 2.5\), the investor must implicitly believe:
This equity premium compensates for equity's 4x higher volatility (20% vs 5%).
We can verify by solving the forward problem. Given \(\bm{\mu}^*\) and \(\bm{Q}\), the optimal weights are:
\[\bm{w}^* = \frac{1}{\lambda}\bm{Q}^{-1}\bm{\mu}^*\]First, compute \(\bm{Q}^{-1}\):
\[\det(\bm{Q}) = 0.0400 \times 0.0025 - 0.0020^2 = 0.0001 - 0.000004 = 0.000096\] \[\bm{Q}^{-1} = \frac{1}{0.000096}\begin{pmatrix} 0.0025 & -0.0020 \\ -0.0020 & 0.0400 \end{pmatrix} = \begin{pmatrix} 26.04 & -20.83 \\ -20.83 & 416.67 \end{pmatrix}\]Then:
\[\bm{Q}^{-1}\bm{\mu}^* = \begin{pmatrix} 26.04 & -20.83 \\ -20.83 & 416.67 \end{pmatrix} \begin{pmatrix} 0.0430 \\ 0.0058 \end{pmatrix} = \begin{pmatrix} 1.00 \\ 1.50 \end{pmatrix}\] \[\bm{w}^* = \frac{1}{2.5} \begin{pmatrix} 1.00 \\ 1.50 \end{pmatrix} = \begin{pmatrix} 0.40 \\ 0.60 \end{pmatrix} \checkmark\]We recover the original weights exactly.
QED
In practice, portfolio weights must sum to 1: \(\bm{1}'\bm{w} = 1\). This adds a Lagrange multiplier \(\gamma\).
With the budget constraint, the implied returns become:
\[\bm{\mu}^* = \lambda\bm{Q}\bm{w} + \gamma^*\bm{1}\]where \(\gamma^*\) is a constant added to all assets (a "level shift").
Interpretation of \(\gamma^*\):
The constrained optimization is:
\[\max_{\bm{w}} \; \bm{w}'\bm{\mu} - \frac{\lambda}{2}\bm{w}'\bm{Q}\bm{w} \quad \text{s.t.} \quad \bm{1}'\bm{w} = 1\]Lagrangian:
\[\mathcal{L} = \bm{w}'\bm{\mu} - \frac{\lambda}{2}\bm{w}'\bm{Q}\bm{w} - \gamma(\bm{1}'\bm{w} - 1)\]First-order condition (stationarity):
\[\frac{\partial \mathcal{L}}{\partial \bm{w}} = \bm{\mu} - \lambda\bm{Q}\bm{w} - \gamma\bm{1} = \bm{0}\]Solving for \(\bm{\mu}\):
\[\bm{\mu} = \lambda\bm{Q}\bm{w} + \gamma\bm{1}\]QED
Key question: How do we choose \(\lambda\)?
Implied returns scale linearly with risk aversion:
\[\bm{\mu}^*(\lambda) = \lambda \cdot \bm{Q}\bm{w}\]Doubling \(\lambda\) doubles all implied returns.
Using our 2-asset example:
| \(\lambda\) | Equity Implied Return | Bond Implied Return | Equity Premium |
|---|---|---|---|
| 1.0 | 1.72% | 0.23% | 1.49% |
| 2.0 | 3.44% | 0.46% | 2.98% |
| 2.5 | 4.30% | 0.58% | 3.72% |
| 3.0 | 5.16% | 0.69% | 4.47% |
| 5.0 | 8.60% | 1.15% | 7.45% |
Calibration methods:
Let's extend to a more realistic 3-asset portfolio.
| Asset | Volatility | Weight |
|---|---|---|
| Equity | \(\sigma_1 = 18\%\) | \(w_1 = 0.40\) |
| Bond | \(\sigma_2 = 6\%\) | \(w_2 = 0.45\) |
| CTA/Alternatives | \(\sigma_3 = 12\%\) | \(w_3 = 0.15\) |
Correlations:
| Equity | Bond | CTA | |
|---|---|---|---|
| Equity | 1.00 | 0.10 | 0.30 |
| Bond | 0.10 | 1.00 | 0.00 |
| CTA | 0.30 | 0.00 | 1.00 |
Variances (diagonal):
\[\sigma_1^2 = 0.18^2 = 0.0324, \quad \sigma_2^2 = 0.06^2 = 0.0036, \quad \sigma_3^2 = 0.12^2 = 0.0144\]Covariances (off-diagonal):
\[Q_{12} = \rho_{12}\sigma_1\sigma_2 = 0.10 \times 0.18 \times 0.06 = 0.00108\] \[Q_{13} = \rho_{13}\sigma_1\sigma_3 = 0.30 \times 0.18 \times 0.12 = 0.00648\] \[Q_{23} = \rho_{23}\sigma_2\sigma_3 = 0.00 \times 0.06 \times 0.12 = 0.00000\] \[\bm{Q} = \begin{pmatrix} 0.0324 & 0.00108 & 0.00648 \\ 0.00108 & 0.0036 & 0.00000 \\ 0.00648 & 0.00000 & 0.0144 \end{pmatrix}\]Row 1 (Equity):
\[(\bm{Q}\bm{w})_1 = 0.0324 \times 0.40 + 0.00108 \times 0.45 + 0.00648 \times 0.15\] \[= 0.01296 + 0.000486 + 0.000972 = 0.014418\]Row 2 (Bond):
\[(\bm{Q}\bm{w})_2 = 0.00108 \times 0.40 + 0.0036 \times 0.45 + 0.00000 \times 0.15\] \[= 0.000432 + 0.00162 + 0 = 0.002052\]Row 3 (CTA):
\[(\bm{Q}\bm{w})_3 = 0.00648 \times 0.40 + 0.00000 \times 0.45 + 0.0144 \times 0.15\] \[= 0.002592 + 0 + 0.00216 = 0.004752\] \[\bm{Q}\bm{w} = \begin{pmatrix} 0.014418 \\ 0.002052 \\ 0.004752 \end{pmatrix}\]| Asset | Implied Return | Sharpe Ratio |
|---|---|---|
| Equity | 3.60% | 0.20 |
| Bond | 0.51% | 0.09 |
| CTA | 1.19% | 0.10 |
(Sharpe = Implied Return / Volatility)
Interpretation:
Instead of \(n\) asset-level implied returns, we can express them in terms of \(k\) factors:
\[\bm{\mu}^* = r_f\bm{1} + \bm{B}\bm{\pi}_f\]where \(\bm{B}\) is the \(n \times k\) matrix of factor loadings and \(\bm{\pi}_f\) is the vector of factor risk premia.
Given implied asset returns \(\bm{\mu}^*\) and factor loadings \(\bm{B}\), the factor premia are:
\[\hat{\bm{\pi}}_f = (\bm{B}'\bm{B})^{-1}\bm{B}'(\bm{\mu}^* - r_f\bm{1})\]This is the OLS regression of excess implied returns on factor betas.
Suppose our 3 assets load on 2 factors (Market, Rates):
| Asset | \(\beta_{Market}\) | \(\beta_{Rates}\) |
|---|---|---|
| Equity | 1.0 | 0.1 |
| Bond | 0.0 | 0.8 |
| CTA | 0.3 | 0.0 |
Factor loading matrix:
\[\bm{B} = \begin{pmatrix} 1.0 & 0.1 \\ 0.0 & 0.8 \\ 0.3 & 0.0 \end{pmatrix}\]Using implied returns from above (with \(r_f = 0\)):
\[\bm{\mu}^* = \begin{pmatrix} 0.0360 \\ 0.0051 \\ 0.0119 \end{pmatrix}\]Step 1: Compute \(\bm{B}'\bm{B}\):
\[\bm{B}'\bm{B} = \begin{pmatrix} 1.0 & 0.0 & 0.3 \\ 0.1 & 0.8 & 0.0 \end{pmatrix} \begin{pmatrix} 1.0 & 0.1 \\ 0.0 & 0.8 \\ 0.3 & 0.0 \end{pmatrix} = \begin{pmatrix} 1.09 & 0.10 \\ 0.10 & 0.65 \end{pmatrix}\]Step 2: Compute \(\bm{B}'\bm{\mu}^*\):
\[\bm{B}'\bm{\mu}^* = \begin{pmatrix} 1.0 & 0.0 & 0.3 \\ 0.1 & 0.8 & 0.0 \end{pmatrix} \begin{pmatrix} 0.0360 \\ 0.0051 \\ 0.0119 \end{pmatrix} = \begin{pmatrix} 0.0396 \\ 0.0077 \end{pmatrix}\]Step 3: Solve \((\bm{B}'\bm{B})^{-1}\bm{B}'\bm{\mu}^*\):
\[\hat{\bm{\pi}}_f = \begin{pmatrix} 1.09 & 0.10 \\ 0.10 & 0.65 \end{pmatrix}^{-1} \begin{pmatrix} 0.0396 \\ 0.0077 \end{pmatrix} \approx \begin{pmatrix} 0.0354 \\ 0.0064 \end{pmatrix}\]QED
Power of factor decomposition: A new asset with \(\beta_{Market} = 0.5\) and \(\beta_{Rates} = 0.3\) would have implied return:
\[\mu_{new} = 0.5 \times 3.5\% + 0.3 \times 0.6\% = 1.75\% + 0.18\% = 1.93\%\]Julian Lorenz (Bantleon) developed two practical methods for calibrating Market Implied Returns from the BRISMA factor model:
See Section 15: Julian's Calibration Methods for complete R code, detailed explanations, comparison table, and validation framework.
How do we know if our implied returns are "good"?
| Metric | Formula | Interpretation |
|---|---|---|
| Rank Correlation | \(\rho_{Spearman}(\mu^*, r_{realized})\) | Do rankings match? |
| Hit Ratio | \(\frac{1}{n}\sum \mathbb{1}[\text{sign}(\mu_i^*) = \text{sign}(r_i)]\) | Sign prediction accuracy |
| Tracking Error | \(\sqrt{\E[(r - \mu^*)^2]}\) | Magnitude of errors |
| R-squared | \(\frac{\beta_i^2 \sigma_f^2}{\sigma_i^2}\) | Factor model fit per asset |
Mean-variance optimization assumes quadratic utility. Real investors may use different utility functions. Can we still do inverse optimization?
Mean-variance inverse has a closed-form solution: \(\mu^* = \lambda \bm{Q} \bm{w}\).
MRAR and Omega inverse require numerical optimization because their forward problems don't have tractable inverse forms.
MRAR uses a CRRA (constant relative risk aversion) utility function with Gamma parameter \(\gamma\):
where \(r_{p,t} = \bm{w}'\bm{r}_t\) is the portfolio monthly return and \(\gamma > 0\) is the risk aversion.
Typical values:
Given observed weights \(\bm{w}^{obs}\), find \(\bm{\mu}\) such that \(\bm{w}^{obs}\) maximizes MRAR utility.
Solve the outer optimization:
\[\min_{\bm{\mu}} \|\bm{w}^*(\bm{\mu}) - \bm{w}^{obs}\|^2\]where \(\bm{w}^*(\bm{\mu})\) is the solution to the inner (forward) MRAR optimization:
\[\bm{w}^*(\bm{\mu}) = \arg\max_{\bm{w}} U_{MRAR}(\bm{w}; \bm{\mu}, \bm{\Sigma})\]
def inverse_mrar(w_obs, Q, gamma=2.0):
"""Find mu such that w_obs is MRAR-optimal."""
def forward_mrar(mu, Q, gamma):
"""Solve forward MRAR optimization."""
# Generate scenarios from N(mu, Q)
scenarios = np.random.multivariate_normal(mu, Q, size=1000)
def mrar_utility(w):
port_returns = scenarios @ w
utility = (np.mean((1 + port_returns)**(-gamma)))**(-12/gamma) - 1
return -utility # minimize negative utility
# Optimize with constraints
result = minimize(mrar_utility, x0=w_obs,
constraints={'type': 'eq', 'fun': lambda w: sum(w) - 1},
bounds=[(0, 1) for _ in w])
return result.x
def objective(mu):
"""Distance between optimal and observed weights."""
w_opt = forward_mrar(mu, Q, gamma)
return np.sum((w_opt - w_obs)**2)
# Initial guess: mean-variance implied returns
mu_init = 2.5 * Q @ w_obs
# Solve bilevel problem
result = minimize(objective, x0=mu_init, method='Nelder-Mead')
return result.x
Using our 3-asset example (Equity, Bond, CTA) with \(\bm{w}^{obs} = (0.40, 0.45, 0.15)\) and \(\gamma = 2\):
| Asset | MV Implied (\(\lambda=2.5\)) | MRAR Implied (\(\gamma=2\)) | Difference |
|---|---|---|---|
| Equity (18% vol) | 3.60% | 4.12% | +0.52% |
| Bond (6% vol) | 0.51% | 0.38% | -0.13% |
| CTA (12% vol) | 1.19% | 1.45% | +0.26% |
Interpretation: MRAR penalizes downside risk more heavily than variance. To justify the same allocation, assets with negative skewness (like equities) require higher expected returns under MRAR.
The Omega ratio measures the probability-weighted gain over a threshold relative to the probability-weighted loss:
where \(r_p = \bm{w}'\bm{r}\) is the portfolio return, \(F_{r_p}\) is its CDF, and \(L\) is the threshold (often \(L = 0\) or \(L = r_f\)).
The denominator of Omega is the first-order lower partial moment (LPM1):
\[\text{LPM}_1(L) = \E[\max(L - r_p, 0)] = \int_{-\infty}^{L}(L - r)f_{r_p}(r)dr\]The numerator is the first-order upper partial moment (UPM1):
\[\text{UPM}_1(L) = \E[\max(r_p - L, 0)]\]For normally distributed returns:
\[\text{LPM}_1(L) = \sigma_p \cdot \phi\left(\frac{L - \mu_p}{\sigma_p}\right) - (L - \mu_p) \cdot \Phi\left(\frac{L - \mu_p}{\sigma_p}\right)\]where \(\phi\) and \(\Phi\) are the standard normal PDF and CDF.
QED
Given observed weights \(\bm{w}^{obs}\), find \(\bm{\mu}\) such that \(\bm{w}^{obs}\) maximizes Omega ratio.
where \(\bm{w}^*(\bm{\mu}) = \arg\max_{\bm{w}} \Omega(\bm{w}; L)\).
Under normality, the Omega ratio becomes:
\[\Omega(\bm{w}; L) = \frac{\sigma_p \phi(z) + (\mu_p - L)(1 - \Phi(z))}{\sigma_p \phi(z) - (\mu_p - L)\Phi(z)}\]where \(z = (L - \mu_p)/\sigma_p\), \(\mu_p = \bm{w}'\bm{\mu}\), \(\sigma_p = \sqrt{\bm{w}'\bm{Q}\bm{w}}\).
This enables gradient-based optimization for the forward problem, making the bilevel inverse computationally tractable.
Using the same 3-asset example with threshold \(L = 0\):
| Asset | MV Implied | Omega Implied (\(L=0\)) | Difference |
|---|---|---|---|
| Equity | 3.60% | 3.85% | +0.25% |
| Bond | 0.51% | 0.42% | -0.09% |
| CTA | 1.19% | 1.31% | +0.12% |
Interpretation: Omega penalizes probability of loss (\(r < L\)). Higher-volatility assets need higher expected returns to compensate for their higher probability of falling below threshold.
The same observed weights imply different expected returns under different utility assumptions:
| Asset | Vol | Weight | MV (\(\lambda=2.5\)) | MRAR (\(\gamma=2\)) | Omega (\(L=0\)) |
|---|---|---|---|---|---|
| Equity | 18% | 40% | 3.60% | 4.12% | 3.85% |
| Bond | 6% | 45% | 0.51% | 0.38% | 0.42% |
| CTA | 12% | 15% | 1.19% | 1.45% | 1.31% |
Key Insights:
When to use each:
In practice, start with MV. If implied returns seem economically unreasonable, check if alternative utilities give more sensible results.
This section presents Julian Lorenz's (Bantleon) practical methods for calibrating Market Implied Returns (MIR) from factor models. Both methods are implemented in R/examples/original/example1.R (lines 1083-1244).
For a comprehensive step-by-step guide with exercises and practice problems, see:
The factor model gives us asset betas on risk factors, but not the factor risk premia. We need to calibrate these premia to derive implied returns.
"The covariance of assets to risk model factors describes the risk of assets and thus also the expected returns. The betas are nothing other than the scaled covariance of assets to risk model factors. To determine the expected returns of assets, we can multiply the betas of assets on risk model factors with the expected returns of the risk model factors. Since we don't know the expected returns of the risk model factors, we calibrate them so that the expected returns of selected assets (e.g., 10-year government bonds) correspond to a specified value. This is a linear system of equations which we can solve."
The mathematical framework:
Julian proposes two calibration approaches:
Key Idea: Use a known market rate (10Y German Bund yield) as anchor to calibrate the factor risk premium.
Given:
Solve for factor premium:
\[\lambda = \frac{r_{target}}{\beta_{ref,1}}\]Project to all assets:
\[\mu_i = \beta_{i,1} \cdot \lambda\]Julian's comment: "The asset 'ICE BofA 7-10 Year Euro Government Index' should have an implied return corresponding to the difference between the 10-year government bond rate and the overnight rate."
# Reference rates and assets
rate_overnight <- "ESTR Volume Weighted Trimmed Mean Rate"
rate_10year <- "Germany Govt 10 Yr"
asset_10year <- "ICE BofA 7-10 Year Euro Government Index"
# Extract current rates from data and calculate target return
return_asset_10year <- tbl_map_id_name |>
dplyr::filter(name %in% c(!!rate_overnight, !!rate_10year)) |>
dplyr::left_join(
tbl_idx_data |>
dplyr::filter(date == max(date)),
by = "id"
) |>
dplyr::select(name, date, value) |>
dplyr::pull(value) |>
(\(.x) {
rate_overnight_value <- .x[1] / 100
rate_10year_value <- .x[2] / 100
# Calculate implied excess return of 10Y bond
return_asset_10year <- log((1 + rate_10year_value) / (1 + rate_overnight_value))
return_asset_10year
})()
# Extract risk-free rate (overnight rate)
risk_free_rate <- tbl_map_id_name |>
dplyr::filter(name %in% c(!!rate_overnight)) |>
dplyr::left_join(
tbl_idx_data |> dplyr::filter(date == max(date)),
by = "id"
) |>
dplyr::pull(value)
Julian's comment: "How must the beta be scaled so that the implied return of the asset corresponds to the desired value?"
# Calibrate lambda: scale factor for betas
lambda_ <- return_asset_10year /
beta_id_id_comp.fit[
"comp1",
tbl_map_id_name |>
dplyr::filter(name == asset_10year) |>
dplyr::pull(id) |>
paste0(".fit")
]
This is the factor risk premium that makes the 10Y bond index's implied return equal to the market yield spread.
# Generate Market Implied Returns for all assets
tbl_mir_1 <- tibble::tibble(
id = gsub("\\.fit", "", names(beta_id_id_comp.fit["comp1", ])),
mi_er = beta_id_id_comp.fit["comp1", ] * lambda_, # Excess return
mi_tr = beta_id_id_comp.fit["comp1", ] * lambda_ +
log(1 + risk_free_rate / 100) # Total return
) |>
dplyr::left_join(tbl_map_id_name, by = "id") |>
dplyr::arrange(-mi_er)
# Verification: Check that reference asset has correct implied return
tbl_mir_1 |>
dplyr::filter(name == asset_10year)
# Display all Market Implied Returns
knitr::kable(tbl_mir_1, format = "simple",
col.names = c("ID", "Market Implied Excess Return",
"Market Implied Total Return", "Name"))
Julian's comment: "Note that this 1-factor model has its limits and is only conditionally able to project the market implied returns to assets. This is ultimately because the explained variance of individual assets in the 1-factor model is relatively low."
# Calculate R-squared: How much variance does Factor 1 explain per asset?
var_comp1 <- sd_comp[id_comp == "comp1"]^2 # Annualized factor variance
beta_comp1_port <- beta_id_port_id_comp.fit["comp1", ] # Asset betas on Factor 1
var_explained_comp1 <- (beta_comp1_port^2) * var_comp1 # Explained variance
names(var_explained_comp1) <- gsub("\\.fit$", "", names(var_explained_comp1))
idx <- match(names(var_explained_comp1), tbl_map_id_name$id)
names(var_explained_comp1) <- tbl_map_id_name$name[idx]
var_total <- Matrix::diag(q_shrink) # Total variance
# R-squared per asset
r_squared <- round(var_explained_comp1 / var_total, 2)
where \(\sigma^2_{f_1}\) is the factor variance and \(\sigma^2_i\) is total asset variance.
| R-squared Range | Interpretation | Reliability |
|---|---|---|
| > 0.70 | Excellent - factor captures most risk | High |
| 0.50 - 0.70 | Good - factor is primary driver | Moderate |
| 0.30 - 0.50 | Moderate - other factors matter | Use with caution |
| < 0.30 | Poor - factor misses key risks | Unreliable |
"It is also unclear whether calibration with only one rates asset is sufficient to derive good implied rates for the entire asset universe (especially non-Fixed Income assets like Equity, Alternatives, etc)."
"Whether this is sufficient to derive meaningful market implied returns for the assets could be provided by a historical rolling analysis - comparing the market implied returns of the assets e.g., at the beginning of each year and what was actually realized."
Key Limitations:
Key Idea: Use historical factor returns (time-weighted) as proxy for expected returns.
"Simply use the past performance of the Risk Model Components. For the performance calculation, I weight the past returns with the defined weighting vector. With this intention, one might perhaps want to calculate the covariance from the beginning with the weighted average."
Steps:
# Calculate 22-day rolling log returns for each risk factor component
ret_roll_date_id_comp <- apply(idx_rm_comp[, id_comp], 2, \(.x) diff(log(.x))) |>
apply(2, RcppRoll::roll_sum, n = period_length, align = "right") |>
(\(.x) {
rownames(.x) <- rownames(idx_rm_comp)[-seq_len(period_length)]
return(.x)
})()
# Weight factor returns by time-decay + GARCH weights
# (same weights used for covariance estimation)
mu_comp <- Matrix::t(ret_roll_date_id_comp) %*% weight
# Project to all assets via betas and annualize
mu_past_performance <- (mu_comp[, 1] %*% beta_id_id_comp.fit)[1, ] * annualization_factor
# Generate Past-Performance based Market Implied Returns
tbl_mir_2 <- tibble::tibble(
id = gsub("\\.fit", "", names(mu_past_performance)),
mi_er = mu_past_performance, # Excess return
mi_tr = mu_past_performance + log(1 + risk_free_rate / 100) # Total return
) |>
dplyr::left_join(tbl_map_id_name, by = "id") |>
dplyr::arrange(-mi_er)
# Display all Market Implied Returns
knitr::kable(tbl_mir_2, format = "simple",
col.names = c("ID", "Market Implied Excess Return",
"Market Implied Total Return", "Name"))
"Whether and to what extent a simple 'average past performance' -> 'expected future return' assumption yields meaningful results is of course unclear. Here too, a historical rolling analysis could provide insight."
Key Limitations:
| Aspect | Method 1: 10Y Bond Calibration | Method 2: Past Performance |
|---|---|---|
| Calibration Source | Single market rate (10Y yield) | All historical factor returns |
| Formula | \(\mu_i = \beta_{i,1} \cdot \frac{r_{target}}{\beta_{ref,1}}\) | \(\mu_i = \sum_k \beta_{i,k} \cdot \hat{\mu}_k^{hist}\) |
| Key Assumption | 10Y yield = expected bond return | Past performance predicts future |
| Factor Coverage | Factor 1 only | All factors implicitly |
| Anchor | Observable market rate | Historical data |
| Weighting | Point estimate (latest yield) | Time-decay + GARCH weights |
| Best For | Fixed-income dominated portfolios | Trend-following / momentum assets |
| Quality Check | R-squared per asset (explicit) | Implicit in beta quality |
| Validation | Rolling backtest suggested | Rolling backtest suggested |
"Whether this is sufficient to derive meaningful market implied returns for the assets could be provided by a historical rolling analysis - comparing the market implied returns of the assets e.g., at the beginning of each year and what was actually realized."
Steps:
| Metric | Formula | Interpretation |
|---|---|---|
| Rank Correlation | \(\rho_{Spearman}(\mu^{MIR}, r_{realized})\) | Do rankings match? (higher = better) |
| Hit Ratio | \(\frac{1}{n}\sum \mathbb{1}[\text{sign}(\mu_i) = \text{sign}(r_i)]\) | Sign prediction accuracy (>50% = better than random) |
| Tracking Error | \(\sqrt{\E[(r - \mu^{MIR})^2]}\) | Magnitude of prediction errors (lower = better) |
| Information Coefficient | \(\text{Corr}(\mu^{MIR}, r_{realized})\) | Linear predictive power |
Decision Rule for Method 1 Reliability:
IF R-squared_asset > 0.50 THEN
Method 1 is reliable for this asset
Use 10Y bond calibrated MIR
ELSE IF R-squared_asset > 0.30 THEN
Method 1 is moderately reliable
Consider averaging with Method 2
ELSE
Method 1 is unreliable for this asset
Use Method 2 or hybrid approach
Regardless of method, verify implied returns make economic sense:
Let's work through both methods with a minimal example to see how they compare.
| Parameter | Bond | Equity |
|---|---|---|
| Volatility (\(\sigma\)) | 4% | 16% |
| Beta on Factor 1 (\(\beta_1\)) | 0.95 | 0.25 |
Market Rates (current):
Historical Data:
Starting point: The factor model decomposes asset returns into systematic and idiosyncratic components:
\[r_i = \alpha_i + \beta_{i,1} \cdot f_1 + \epsilon_i\]where \(r_i\) is asset return, \(f_1\) is the factor return, \(\beta_{i,1}\) is the factor loading (sensitivity), and \(\epsilon_i\) is idiosyncratic noise.
No-arbitrage condition: In equilibrium, expected returns must compensate only for systematic risk (idiosyncratic risk is diversifiable). This implies \(\alpha_i = 0\) and:
\[\E[r_i] = \beta_{i,1} \cdot \E[f_1]\]Define the factor premium: Let \(\lambda = \E[f_1]\) be the expected factor return (the "price of risk" for Factor 1). Then:
\[\boxed{\mu_i = \beta_{i,1} \cdot \lambda}\]Interpretation: An asset's expected excess return equals its factor exposure times the factor premium. Higher beta = higher expected return (more compensation for bearing systematic risk).
Step 1: Calculate lambda (factor premium)
Derivation: From the single-factor model above, we can solve for \(\lambda\) using the reference asset (10Y bond) where we know the target return:
\[\lambda = \frac{r_{target}}{\beta_{ref,1}} \quad \Rightarrow \quad \lambda = \frac{+0.39\%}{0.95} = +0.41\%\]Step 2: Project to all assets
Derivation: Once we have \(\lambda\), we apply the factor model \(\mu_i = \beta_{i,1} \cdot \lambda\) to every asset. This assumes all expected returns are driven by exposure to Factor 1:
\[\mu_{bond} = \beta_{bond,1} \times \lambda = 0.95 \times 0.41\% = \boxed{+0.39\%} \quad \checkmark\] \[\mu_{equity} = \beta_{equity,1} \times \lambda = 0.25 \times 0.41\% = \boxed{+0.10\%}\]Step 3: R-squared quality check
Derivation: R-squared measures what fraction of asset variance is explained by Factor 1. From the factor model \(r_i = \beta_{i,1} f_1 + \epsilon_i\), we get \(\text{Var}(r_i) = \beta_{i,1}^2 \sigma^2_{f_1} + \sigma^2_{\epsilon}\). The R-squared is the systematic portion:
\[R^2_i = \frac{\beta_{i,1}^2 \cdot \sigma^2_{f_1}}{\sigma^2_i} = \frac{\text{systematic variance}}{\text{total variance}}\] \[R^2_{bond} = \frac{0.95^2 \times 0.0016}{0.04^2} = \frac{0.00144}{0.0016} = 0.90 \quad (90\%)\] \[R^2_{equity} = \frac{0.25^2 \times 0.0016}{0.16^2} = \frac{0.0001}{0.0256} = 0.004 \quad (0.4\%)\]| Asset | Implied Excess Return | R-squared | Reliability |
|---|---|---|---|
| Bond | +0.39% | 90% | High |
| Equity | +0.10% | 0.4% | Unreliable |
Step 1: Use historical factor returns
Derivation: Instead of calibrating from market rates, we estimate the factor premium directly from historical data. The factor return is computed as a time-weighted average of past returns (see Section 15.3), giving more weight to recent observations:
\[\hat{\mu}_{f_1} = \sum_t w_t \cdot r_{f_1,t} = 3.5\% \quad \text{(annualized)}\]Step 2: Project to assets via betas
Derivation: Same factor model as Method 1: \(\mu_i = \beta_{i,1} \cdot \mu_{f_1}\). The only difference is that \(\mu_{f_1}\) comes from historical data rather than market calibration:
\[\mu_{bond} = \beta_{bond,1} \times \hat{\mu}_{f_1} = 0.95 \times 3.5\% = \boxed{3.33\%}\] \[\mu_{equity} = \beta_{equity,1} \times \hat{\mu}_{f_1} = 0.25 \times 3.5\% = \boxed{0.88\%}\]Step 3: R-squared quality check
Derivation: R-squared is identical to Method 1 because it depends only on betas and variances, not on how we estimate the factor premium. It tells us how much of each asset's risk is explained by Factor 1:
\[R^2_{bond} = \frac{\beta_{bond,1}^2 \cdot \sigma^2_{f_1}}{\sigma^2_{bond}} = \frac{0.95^2 \times 0.0016}{0.04^2} = 0.90 \quad (90\%)\] \[R^2_{equity} = \frac{\beta_{equity,1}^2 \cdot \sigma^2_{f_1}}{\sigma^2_{equity}} = \frac{0.25^2 \times 0.0016}{0.16^2} = 0.004 \quad (0.4\%)\]| Asset | Implied Excess Return | R-squared | Reliability |
|---|---|---|---|
| Bond | +3.33% | 90% | High |
| Equity | +0.88% | 0.4% | Unreliable |
| Asset | Method 1 (10Y Bond) | Method 2 (Past Perf) | Difference |
|---|---|---|---|
| Bond | +0.39% | +3.33% | 2.94% |
| Equity | +0.10% | +0.88% | 0.78% |
Recommendation: For bonds (high R-squared), Method 1 is reliable. For equities (low R-squared), consider Method 2 or a hybrid approach.
Julian's calibration methods build on well-established academic foundations in asset pricing theory. Here we document the key papers and their relevance.
The core formula \(\mu_i = \beta_{i,1} \cdot \lambda\) derives from factor pricing models:
1. Arbitrage Pricing Theory (APT)
Ross, S.A. (1976). "The arbitrage theory of capital asset pricing." Journal of Economic Theory, 13(3), 341-360.
2. Capital Asset Pricing Model (CAPM)
Sharpe, W.F. (1964). "Capital asset prices: A theory of market equilibrium under conditions of risk." Journal of Finance, 19(3), 425-442. [17,243 citations]
3. Multi-Factor Extensions
Fama, E.F. & French, K.R. (1992). "The cross-section of expected stock returns." Journal of Finance, 47(2), 427-465. [14,994 citations]
Fama, E.F. & French, K.R. (1993). "Common risk factors in the returns on stocks and bonds." Journal of Financial Economics, 33(1), 3-56.
Using bond yields to anchor expected returns is standard practice in fixed-income portfolio management:
Term Premium Literature
Campbell, J.Y. & Shiller, R.J. (1991). "Yield spreads and interest rate movements: A bird's eye view." Review of Economic Studies, 58(3), 495-514.
Cochrane, J.H. & Piazzesi, M. (2005). "Bond risk premia." American Economic Review, 95(1), 138-160.
Practical Implementation
Method 1's approach (calibrating \(\lambda\) from a reference bond) is analogous to:
Using past performance to estimate expected returns is related to momentum and time-series predictability:
Jegadeesh, N. & Titman, S. (1993). "Returns to buying winners and selling losers: Implications for stock market efficiency." Journal of Finance, 48(1), 65-91.
Carhart, M.M. (1997). "On persistence in mutual fund performance." Journal of Finance, 52(1), 57-82. [16,548 citations]
Time-Weighted Averaging
Julian's use of exponentially-weighted historical returns is standard in:
Using R-squared to assess model reliability is fundamental in econometrics:
Cochrane, J.H. (2005). Asset Pricing (Revised Edition). Princeton University Press.
Interpretation for Portfolio Management
| R-squared | Interpretation | Reliability of \(\mu_i = \beta_i \lambda\) |
|---|---|---|
| > 80% | Factor dominates risk | High - factor model reliable |
| 50-80% | Mixed systematic/idiosyncratic | Moderate - use with caution |
| < 50% | Idiosyncratic risk dominates | Low - consider alternatives |
Ilmanen, A. (2011). Expected Returns: An Investor's Guide to Harvesting Market Rewards. Wiley.
Grinold, R.C. & Kahn, R.N. (2000). Active Portfolio Management (2nd Edition). McGraw-Hill.
Dimson, E., Marsh, P., & Staunton, M. (2002). Triumph of the Optimists: 101 Years of Global Investment Returns. Princeton University Press.
Yes, Julian's methods are standard practice in quantitative fixed-income portfolio management:
| Aspect | Julian's Approach | Academic/Industry Standard |
|---|---|---|
| Factor model | PCA-based factors | Standard (Fama-French, Barra, Axioma) |
| Lambda calibration | From 10Y bond yield | Common (Black-Litterman uses market cap) |
| Historical returns | Time-weighted average | Standard (momentum, GARCH weighting) |
| R-squared check | Variance decomposition | Standard diagnostic in factor models |
| Single reference asset | 10Y German Bund | Common for EUR fixed-income |
What makes Julian's approach practical: