5.8.4 Turnover Control & the Multi-Period Trade-Off

Basic Prerequisites: 03 · Transaction-Cost Models.

---

1. Intuition & Practical Objective

Turnover is the rate at which you convert an alpha model into trading. Everything in this folder lives on this axis: a brilliant signal rebalanced daily at $300\%$ annual turnover can lose money, while a modest signal rebalanced quarterly can compound. This page answers the two questions a practitioner actually faces:

1. How much may I trade? Two controls: penalize turnover with a linear price $\lambda$ (an economic cost), or budget it with a hard constraint $\Vert w-w_0{\Vert }_1\le \tau$ (a risk control). The first requires you to know your cost; the second does not, and is therefore what risk desks prefer.
2. How fast should I converge? The target is a moving object - alphas revise every period. The great insight of the dynamic-trading literature (Gârleanu & Pedersen 2013) is that you should trade partially toward an “aim” portfolio, not jump to the current target, because a jump buys a noisy estimate you will want to undo next period.

The one-sentence essence. “Turnover is a dial, and the optimal setting is where the marginal alpha of one more unit of trading equals the marginal cost of executing it - in a static problem that is a penalty $\lambda$; across periods it is a fraction of the way to a long-horizon aim.”

---

2. Mathematical Ground Truth & Derivations

2.1 Two ways to control turnover

Penalty form. Add $\lambda \Vert w-w_0{\Vert }_1$ to the objective:

$$\underset{w\in \mathcal{C}}{\max }{\mu }^{\top }w-\frac{\delta }{2}{w}^{\top }\Sigma w-\lambda \Vert w-w_0{\Vert }_1.$$

By the envelope logic of 02, $\lambda$ is the model’s assumed cost per unit of trading. Choosing $\lambda$ therefore means choosing a cost forecast - and §3 shows that the certainty-equivalent-optimal $\lambda$ lands exactly on the true cost, which is the cleanest possible statement of “price your trades honestly.”

Budget form. Instead cap the ${\ell }_1$ distance:

$$\underset{w\in \mathcal{C}}{\max }{\mu }^{\top }w-\frac{\delta }{2}{w}^{\top }\Sigma w\text{s.t.}\Vert w-w_0{\Vert }_1\le \tau .$$

This is convex, needs no cost estimate, and is the form a risk committee is happy to mandate (“no more than $20\%$ turnover per rebalance”). The price of the budget is its shadow price - exactly the $\lambda$ that the penalty form would have used.

The trade-off frontier. Sweep either control and plot the pairs $(\text{turnover},\text{net alpha})$:

$$\text{net }{\alpha }_{\text{ann}}(\lambda )=\underset{\text{gross}}{\underbrace{12{\mu }^{\top }w(\lambda )}}-\underset{\text{realized cost}}{\underbrace{12c_{\text{true}}\Vert w(\lambda )-w_0{\Vert }_1}}.$$

2.2 The multi-period problem and partial adjustment

Over $H$ periods, ignoring the path is a mistake. The canonical tracking problem is

$$\underset{\{w_t\}}{\min }\sum _{t=1}^H[\frac{\rho }{2}(w_t-w^{\ast }{)}^{\top }\Sigma (w_t-w^{\ast })+\frac{\kappa }{2}(w_t-w_{t-1}{)}^{\top }\Lambda (w_t-w_{t-1})],$$

a tracking-error penalty on $w_t$ plus a cost penalty on the move. In the scalar case this is a tridiagonal linear system whose solution is a monotone ramp to the target, and its asymptotic behaviour is geometric:

$$w_t-w^{\ast }\approx G^t(w_0-w^{\ast }),G=1+\frac{a}{2}-\sqrt{a+\frac{a^2}{4}},a=\frac{\rho {\sigma }^2}{\kappa \eta }.$$

• Cheap trading ($\kappa$ small, $a$ large) ⇒ $G\to 0$: jump to the target at once.
• Expensive trading ($\kappa$ large, $a$ small) ⇒ $G\to 1$: creep toward it; over a finite horizon you may never arrive.

The Gârleanu–Pedersen aim (preview of 06). With predictable returns and proportional costs, the optimal policy is

$$\boxed{x_t=x_{t-1}+(I+\kappa \Sigma {)}^{-1}({\text{aim}}_t-x_{t-1})},\text{aim}=(I+\kappa \Sigma {)}^{-1}(\delta \Sigma {)}^{-1}\mu ,$$

i.e. trade a matrix fraction $(I+\kappa \Sigma {)}^{-1}$ of the gap between the aim portfolio and the current book - not all of it. The matrix is non-diagonal, so a signal in one name moves several weights: cross-asset cost coupling (verified in 06).

---

3. Computational Implementation - the frontier and the ramp

(A) sweep the linear penalty $\lambda$ and report turnover, gross alpha, realized net alpha at a true cost of 10 bp, net Sharpe and certainty-equivalent utility. (B) solve the multi-period tracking problem and compare the optimal ramp to “all at once” and “equal slices.” numpy + scipy.

Three verified readings:

• (A) The utility-optimal $\lambda$ is exactly the true cost. Certainty-equivalent utility peaks at $0.15554$ over $\lambda \in [0.00075,0.0010]$, and $0.0010$ is the true 10 bp cost used to compute realized net alpha. Miss it low ($\lambda =0$) and you overtrade: turnover $0.5932$, net Sharpe $1.1621$, utility $0.15153$. Miss it high ($\lambda =0.004$) and you undertrade: turnover drops to $0.0952$ but gross alpha falls to $0.1731$ and utility to $0.14952$. The frontier is a hump, and its peak is the honest cost.
• (A) Net Sharpe and utility disagree - know which you are optimizing. Net Sharpe rises monotonically with $\lambda$ ($1.1621\to 1.4427$) because shrinking risk raises the ratio, while utility peaks in the middle because it also counts the alpha you gave up. A desk judged on Sharpe will always over-penalize turnover; a desk judged on dollars will not. State the objective before tuning the dial.
• (B) The optimal path is a geometric ramp, and “all-at-once” is only right when trading is cheap. With $\kappa =1$ the optimum (objective $0.3090$) beats both all-at-once ($0.5000$) and equal slices ($1.1562$). As impact grows to $\kappa =16$ the ramp flattens dramatically ($w_1=0.2135$ vs $0.618$ at $\kappa =1$) and the advantage over all-at-once explodes ($1.7083$ vs $8.0000$): at high cost, front-loading is catastrophic. Note the path never reaches the target within $H=8$ at $\kappa =16$ - the horizon is a real constraint, and “finish rebalancing” is a choice, not an identity.

---

4. Failure Modes & First-Principles Breakdowns

1. Tuning $\lambda$ by feel instead of by cost. §3 shows the penalty is a cost forecast; setting it “to be safe” makes you skip good trades. The correct discipline is to set $\lambda$ equal to your measured marginal cost (or to the shadow price of a turnover budget you have committed to).
2. The turnover budget hides regime change. A tight $\tau$ glues the portfolio to a stale book exactly when the world has moved (a regime shift looks identical to noise in a one-period problem). Budgets are about inputs you can see; robustness about regimes is a different tool (Robust · 05).
3. Single-period myopia. Re-solving “optimally” every period with a fresh target front-loads trades and re-trades the same signal repeatedly - the turnover spiral of 05. The fix is the partial-adjustment / aim policy, not a bigger penalty.
4. Ignoring the trading horizon. The ramp of §3(B) shows the target may be unreachable in $H$ periods at high cost. Pretending you will “finish the rebalance” next week when the cost model says otherwise is a forecast error, and it is a systematic one.

---

5. References

• Gârleanu, Nicolae & Pedersen, Lasse Heje (2013), Dynamic Trading with Predictable Returns and Transaction Costs, Journal of Finance 68(6):2309–2340
• Boyd, Busseti, Diamond, Kahn, Koh, Nystrup & Speth (2017), Multi-Period Trading via Convex Optimization, Foundations and Trends in Optimization 3(1):1–72
• Lobo, Fazel & Boyd (2007), Annals of OR 152:341–365
• Grinold & Kahn (2000), Active Portfolio Management, 2nd ed.
• Clarke, de Silva & Thorley (2002), FAJ 58(5):48–66

---

6. Connected Graph Bridges

• Back: 03 · Transaction-Cost Models
• Execution bridge: Optimal Execution (Almgren–Chriss) (the within-day version of the ramp) · POV
• Siblings: Robust · Constraints as Robustness (turnover as a $w$-space constraint) · Covariance Shrinkage & RMT (churn from noisy inputs)
• Continue: 05 · Failure Modes & Practice · 06 · Advanced Extensions
• Hub: Index Hub