Estimation Risk and Asset Allocation

Definition

Estimation risk refers to the distortion in forecasts and portfolio decisions that arises from substituting a sample estimate for an unknown true parameter. In the long-horizon context, estimation uncertainty in the mean return compounds over the investment horizon, producing biases and efficiency losses that do not vanish unless both $T$ (sample length) and $T/H$ (sample-to-horizon ratio) grow large.

Key Ideas

The standard practice of compounding the arithmetic sample mean $\hat\mu$ over $H$ periods produces an estimate of expected terminal wealth that is upward-biased by $e^{\sigma^2 H/T}$ — from Jensen's inequality applied to the estimation error in $\hat\mu$ .
The geometric mean is unbiased only when $H = T$ ; for $H \neq T$ it is biased in either direction.
All common estimators nest in the class $C = e^{(\hat\mu + k \cdot \frac{1}{2}\sigma^2)H}$ ; the minimum mean squared error (MSE) weight is $k_M = 1 - \tfrac{3}{2}H/T$ , always below the unbiased weight $k_U = 1 - H/T$ .
The unbiased estimator is surprisingly imprecise: for $H/T > 0.2$ the geometric estimator, though biased, has lower root-mean-square error (RMSE) than the unbiased one.
Parameter uncertainty in $\hat\mu$ , properly incorporated into expected utility, decreases the optimal risky-asset allocation as $H/T$ grows — the opposite of conventional "stocks for the long run" advice.

How It Works

The Compounding Bias

Let log-returns $r_t \overset{iid}{\sim} \mathcal{N}(\mu, \sigma^2)$ . The true expected portfolio value after $H$ periods is: $E(V_H) = e^{\mu H + \frac{1}{2}\sigma^2 H}$

The maximum likelihood (ML) estimator substitutes $\hat\mu$ (sample mean over $T$ periods, $\hat\mu = \mu + \sigma\varepsilon/\sqrt{T}$ , $\varepsilon \sim \mathcal{N}(0,1)$ ): $A = e^{(\hat\mu + \frac{1}{2}\sigma^2)H} \implies E(A) = E(V_H) \cdot e^{\sigma^2 H/T}$

The upward bias factor $e^{\sigma^2 H/T}$ grows with $H/T$ and vanishes only when $H/T \to 0$ .

The Estimator Family

Estimator	$k$	Bias	RMSE ranking
Arithmetic / ML ( $A$ )	$1$	$+e^{\sigma^2 H/T}$	Worst
Geometric ( $G$ )	$0$	Zero only if $H=T$	2nd best for $H \approx T/3$
Unbiased ( $U$ )	$1 - H/T$	0	3rd (worse than $G$ for large $H/T$ )
Min-MSE ( $M$ )	$1 - \tfrac{3}{2}H/T$	Non-zero	Best at all horizons

For $H > T/3$ the ordering is $M < G < U < A$ . At $T = 75$ , $H = 40$ : $A \to \$120$ , $U \to \$80$ , $M \to \$25$ per $\$1$ invested.

The Allocation Paradox

With power utility (risk aversion $\gamma > 1$ ), integrating out the posterior $\hat a \sim \mathcal{N}(a, \sigma^2/T)$ yields: $w^* = \frac{\hat a - r_0}{\gamma \sigma^2 \left(1 + \dfrac{H}{T}\right)}$

The denominator increases with $H/T$ , so the optimal risky-asset weight decreases as the horizon lengthens. For $T = 30$ , $H = 40$ , $\gamma = 2$ : $w^*$ falls from 87% (no parameter uncertainty) to 53%. Log-utility ( $\gamma = 1$ ) is the exception: linearity in $a$ makes the allocation horizon-independent.

Intuition: Estimation error in $\hat a$ gets compounded $H$ times in the utility calculation. The longer the horizon, the more this uncertainty is amplified, making the risky asset effectively riskier from the investor's perspective.

Robustness

Serial correlation: Enters through $F_T/F_H$ (ratio of variance-of-sum per period for past vs. future); for realistic moving average (MA(4)) autocorrelation in S&P 500 annual returns, the correction is a second-order effect.
Heteroskedasticity: Averages out for long horizons if variance is stationary.
Better $\mu$ estimates: A more precise prior/model (e.g., dividend discount model) enters simply as an increase in effective sample size $T$ .

Why It Matters

Compounds the "declining equity premium" evidence (Fama-French 2002; Jagannathan et al. 2000): even after lowering the per-period estimate, a further downward correction is needed for long-horizon compounding.
Challenges the canonical recommendation that long-horizon investors should hold more equities (Samuelson 1969, Merton 1969 known-parameter case; Campbell-Viceira 1999 predictability extension).
Relevant to pension funds, endowments, and any institution projecting long-term portfolio values: arithmetic mean compounding used by Ibbotson Associates is worst-case for long horizons.

Open Questions

The paper assumes a one-shot allocation; dynamic rebalancing and learning (Barberis 2000; Brennan 1998) may partially offset the effect.
The result requires $\gamma > 1$ ; log-utility investors are unaffected. The size of the correction is sensitive to the assumed risk aversion.
Extension to multi-asset (mean-variance) setting with estimation error in the full covariance matrix is not covered here.