← Blog/blog/target-cap-does-not-cap-trading

The turnover cap that does not cap trading

A sophisticated return model can generate 10,000 joint scenarios and still lose the practical argument to one line of portfolio accounting. Mensah and coauthors compare historical windows with ARMA–GARCH marginals joined by a Student-t copula for 30 exchange-traded funds, or ETFs. The conditional model changes the estimated opportunity set, but the trading rule often changes realized outcomes more.

The quiet reason is that the experiment constrains changes between two target portfolios. Markets move the actual holdings away from yesterday's target before the next trade. Rebalancing from those drifted holdings is a different quantity—and it is not bounded by the target-change cap.

01

Two distances, one portfolio

One-way turnover is half the absolute change in portfolio weights; the factor one-half avoids counting every sale and its matching purchase twice. The paper calls the distance from the old target to the new target τ. After returns occur, yesterday's holdings drift to new pre-trade weights. The distance from those drifted weights to the new target is implementation turnover κ.

The TypeScript below is the code the page runs. Python and C++ are line-for-line translations of the same normalization and distance.

def drift_weights(weights, returns):
    wealth = sum(w * (1 + r) for w, r in zip(weights, returns))
    return [w * (1 + r) / wealth for w, r in zip(weights, returns)]

def one_way_turnover(first, second):
    return sum(abs(b - a) for a, b in zip(first, second)) / 2

def turnover_gap(previous, target, returns):
    drifted = drift_weights(previous, returns)
    target_change = one_way_turnover(previous, target)
    required_trade = one_way_turnover(drifted, target)
    return target_change, required_trade
02

A hand-checkable counterexample

Start 50/50 in two assets. Let one rise 10% while the other falls 10%. Wealth is unchanged, but the holdings drift to 55/45. If the desired target is also 55/45, the paper's baseline cap moves the official target only to 0.500132/0.499868. Target adjustment is exactly 0.0132%, while the trade from drifted holdings back to that capped target is 4.9868%.

The sharper example keeps the target at 50/50. Then target adjustment is zero, yet the same return shock requires 5.0% one-way turnover. A cap on one distance says nothing finite about their ratio when the denominator is zero.

implementation turnovertarget adjustment
Deterministic two-asset illustration. The desired target is 55/45, the previous target is 50/50, and the baseline paper cap is applied. Return shocks are symmetric; these are computed toy values, not paper-reported performance.
Implementation turnover divided by the baseline target-adjustment cap in the same deterministic example. The widening multiple comes only from return-driven drift.
03

What the paper actually found

Across 946 out-of-sample sessions, historical inputs produced higher observed reward–risk ratios in most matched comparisons. But none of 48 paired differences remained significant after Holm or Benjamini–Hochberg multiple-testing correction. That is evidence against a broad superiority claim, not proof that the methods are equivalent.

The sensitivity experiment is more revealing. Replacing the tiny baseline cap with a 1% daily target-adjustment allowance changed some annualized geometric returns by several percentage points and greatly increased implementation turnover. Changing the conditional simulation from 10,000 scenarios to 5,000 or 20,000 moved the reported annualized returns by at most 0.023 percentage points; changing its seed moved them by at most 0.026 points.

Portfolio1% cap − baseline returnImplementation TO: base → 1%
Historical LO MC99-6.57 pp41.7% → 89.3%
Historical LS TC99+4.13 pp46.1% → 346.1%
Dynamic LO TC99+2.32 pp43.4% → 277.9%
Dynamic LS MV-8.90 pp41.7% → 189.6%
Dynamic LS TC99+2.84 pp43.4% → 323.9%
Selected paper-reported values from Appendix Table A10. LO/LS mean long-only/limited long-short; MV is minimum variance; TC99 is tangent CVaR at 99%; MC99 is minimum CVaR at 99%.
04

What not to conclude

The study uses adjusted close-to-close returns and does not observe execution prices, market impact, short financing, or the initial formation trade. Its proportional costs are illustrative. The authors also note that targets estimated after a close are attributed to the next close; this does not establish that the portfolio could trade at the estimation close. The 30 ETFs diversify tickers, not necessarily independent economic risks, and provider adjustment factors were not reconstructed from raw corporate actions.

Nor does failure to reject 48 adjusted comparisons prove equal forecasting value. A restrictive policy can compress two genuinely different forecasts into nearly identical holdings. Model evaluation and policy evaluation answer different questions.

05

What to probe next

Put the constraint on implementation turnover κ, or report both τ and κ every day. Simulate executable timing with bid–ask spreads, nonlinear market impact, short borrow, and delayed fills. Then vary the cap on a common grid and test whether either input method adds value at the same realized trading budget. The site's linear-model page is a useful route into mean–variance objectives, while the neural-network page shows how a more flexible return model can be plugged into the same decision layer.

References

  1. Mensah, A. A., Jaffri, A. M., Appiah, N., Lindquist, W. B., Rachev, S. T., & Fabozzi, F. J. (2026). Dynamic Scenario-Based Portfolio Optimization for Multi-Asset ETFs. arXiv:2611.16540
  2. Markowitz, H. (1952). Portfolio Selection. The Journal of Finance 7(1)
  3. Rockafellar, R. T., & Uryasev, S. (2000). Optimization of Conditional Value-at-Risk. Journal of Risk 2(3)