Jacquier-Kane-Marcus (2004) Optimal Estimation of the Risk Premium for the Long Run and Asset Allocation

Summary

When a sample estimate of the mean log-return is compounded over a long horizon $H$ , Jensen's inequality applied to estimation error produces upward-biased forecasts of portfolio value — a bias that grows with $H/T$ . The paper derives an analytical minimum-mean-squared-error (MSE) estimator that compounds at a rate strictly below the arithmetic, unbiased, and (for $H > T/3$ ) geometric estimators. Applied to asset allocation, properly incorporating parameter uncertainty reverses the conventional wisdom: longer investment horizons require lower allocations to risky assets, not higher.

Key Claims

The arithmetic (maximum likelihood, ML) estimator of $E(V_H)$ is biased upward by factor $e^{\sigma^2 H/T}$ ; the geometric estimator is unbiased only when $H = T$ .
All estimators of interest belong to the class $C = e^{(\hat\mu + k \cdot \frac{1}{2}\sigma^2)H}$ : $k=1$ (arithmetic), $k=0$ (geometric), $k_U = 1 - H/T$ (unbiased), $k_M = 1 - \tfrac{3}{2}H/T$ (minimum MSE).
The minimum-MSE estimator $M$ is always lower than unbiased $U$ ; for $H > T/3$ , the ordering is $M < G < U < A$ . Geometric beats unbiased at large horizons because $U$ has such high variance that its unbiasedness is not worth the cost.
At $T=75$ , $H=40$ : $A$ forecasts $\$1 \to \$120$ ; $U \to \$80$ ; $M \to \$25$ . root mean squared error (RMSE) of $A$ is 2.5× $E(V_H)$ — "catastrophic precision."
With power utility (risk aversion $\gamma > 1$ ), integrating parameter uncertainty into expected utility yields the optimal risky-asset weight $w^* = \frac{\hat a - r_0}{\gamma \sigma^2 (1 + H/T)}$ , decreasing in $H/T$ . For $T=30$ , $H=40$ , $\gamma=2$ : weight falls from 87% (conventional) to 53%.
Serial correlation and heteroskedasticity are second-order corrections for long horizons; the $H/T$ ratio drives results.

Concepts Introduced or Extended

Estimation Risk and Asset Allocation

Entities Mentioned

Quotes

"Longer investment horizons require lower, not higher, allocations to risky assets."

"The catastrophic lack of precision of A, the relatively disappointing imprecision of U, and strong performance of the geometric estimator in the middle range of investment horizons are the striking features of [the simulation]."

My Take

A rare case where a clean analytical result overturns textbook advice. The key insight is that $H/T$ — not $H$ alone — governs both the bias and the asset allocation, so practitioners using long historical samples for short-to-medium horizons are closer to optimal than academics using arithmetic averages. The paper does not address dynamic rebalancing or learning, but the static result is stark and under-appreciated. Competes with Barberis (2000) and Campbell-Viceira (1999) on long-horizon allocation but reaches the opposite qualitative conclusion via a different channel.