The No-Trade Set as a Turnpike: Dissipativity Theory for Portfolios with Proportional Transaction Costs

Many machine-learning trading papers report strong returns while assuming that buying and selling cost nothing. When realistic fees are added, those returns often vanish, because a model that re-chases every forecast trades too often. This proposal brings in a tool from optimal control. The turnpike property says that in a long-horizon optimal plan, the solution quickly moves to a preferred steady operating point, stays there most of the time, and leaves only near the end. Strict dissipativity is the matching condition: there is an "energy-like" accounting function that charges the plan for every moment it spends away from that operating point. We argue that with proportional transaction costs, the turnpike of a portfolio problem is a set, the region of portfolios where trading is not worth its cost. Optimal trajectories reach this set in finite time rather than only approaching it asymptotically. The set's width follows a simple formula in the fee, the discount rate and risk aversion.

This matters for a practical reason. Existing friction-aware methods each carry an unexplained tuning knob. These include the fee multiplier inside single-period convex rebalancing, the weight on a squared-turnover penalty, and the fixed trading interval in dilution schemes. A turnpike analysis predicts what each knob should be. It also predicts when a cheap single-period rule is already optimal and when a genuinely multi-period controller earns its extra cost. The second prediction is the falsifiable heart of the work.

Core Claim

Model the portfolio as a discrete-time control system. The state $w_t \in \mathbb{R}^n$ is the vector of portfolio weights, and the control $u_t = w_{t+1} - w_t$ is the trade. The stage cost is

$$ \ell_t(w,u) = f_t(w) + \sum_i c_i |u_i|, \qquad f_t(w) = -\mu_t^\top w + \tfrac{\rho}{2} w^\top \Sigma w , $$

where $\mu_t$ is the forecast expected return (the "alpha"), $\Sigma$ the return covariance, $\rho$ the risk aversion and $c_i$ the proportional fee on asset $i$. Costs are summed with discount factor $\gamma \in (0,1)$ under a budget constraint.

Our central claim concerns stationary alpha ($\mu_t \equiv \mu$). In that case the long-horizon optimal trajectories obey a set-valued turnpike. The turnpike set is

$$ S_\gamma = \{ w : |\nabla f(w)_i - \nu| \le \kappa_\gamma\, c_i \ \text{for all } i \}, \qquad \kappa_\gamma = \tfrac{1-\gamma}{\gamma}, $$

with $\nu$ the multiplier of the budget constraint. The optimal infinite-horizon policy projects the current portfolio onto $S_\gamma$ and then stops trading. Equivalently, it solves a single-period rebalancing problem with the fee scaled by exactly $\kappa_\gamma$.

The governing dissipation inequality has a linear lower bound in the trade size, from the absolute-value fee, rather than the quadratic one used in smooth theory. This linear bound is why arrival happens in finite time instead of exponentially.

When alpha varies over time with a persistence (half-life) $h$, the set $S_\gamma$ moves. The single-period rule with multiplier $\kappa_\gamma$ is then no longer optimal. Its excess trading cost should grow as $h$ shrinks relative to the time the optimal trajectory needs to reach the moving set. This gives a concrete, testable boundary between regimes. Where alpha persists long relative to the reaching time, a correctly scaled single-period rule suffices. Where alpha is short-lived, multi-period control pays.

A caveat on novelty. The existence of a no-trade region under proportional costs is classical portfolio theory, and so is multi-period convex trading in practice. We do not claim either. What is new is three things. The first is casting the no-trade region as a dissipativity-certified set turnpike. The second is the finite-time arrival result and the explicit $\kappa_\gamma$ with its discounted-stability conditions. The third is the persistence criterion that says when the multi-period controller is worth running.

Relation to Prior Work

The control-theoretic results supply the method, but they all stop at smooth problems with point-valued turnpikes.

Faulwasser et al. prove that strict dissipativity plus reachability yields a turnpike, and the converse. Their results are stated for a steady-state point, and they verify dissipativity on a smooth reactor model using polynomial storage functions.

Trélat and Zhang show that strict strong duality in convex problems implies strict dissipativity. They also introduce measure-turnpikes: the total time spent away from the turnpike set is bounded independently of the horizon. This is the route we reuse. Our problem is convex, so the storage function can be taken linear, $\lambda^\top w$, where $\lambda$ is the steady-state multiplier. Convex duality does not need differentiability, which sidesteps the obstacle described next.

That obstacle comes from Angeli et al.. They show that local controllability is necessary for dissipativity and that dissipativity is robust to perturbations. However, their argument relies on twice-differentiable sensitivity analysis, which the absolute-value fee breaks. Their work therefore constrains our design: we must not lean on second-order arguments.

Discounting is unavoidable in finance. Zanon and Gros show that ordinary dissipativity can fail under discounting. They propose a "strong discounted strict dissipativity" condition that extends the stable range in a linear-quadratic benchmark from $\gamma \ge 0.846$ down to $\gamma = 0.3109$. We adapt this condition to a set-valued turnpike. It is the natural source of the $\gamma$-dependence in $\kappa_\gamma$.

Müller, Grüne and Krügel study when weighted sums of dissipative costs stay dissipative. That question bears directly on our mixture of return, risk and fee terms.

Grüne and Trélat et al. extend turnpikes to periodic orbits. Their work suggests how a slowly moving $S_\gamma$ might be handled. Geshkovski stresses that controllability matters more than open-loop stability, and that turnpike proofs under active constraints are still immature. Portfolio problems sit squarely in that immature territory.

The closest bridge to resource allocation is the restless-bandit model predictive control paper. It uses dissipativity to show that a receding-horizon linear-programming policy is near-optimal, but it assumes reallocation is free. We study exactly the switching cost it omits.

On the finance side, the prior work motivates the problem and supplies our competitors. Boyd et al.'s Markowitz++ adds spread and turnover penalties to rolling single-period convex rebalancing, but chooses the penalty weights empirically. Shen et al.'s doubly regularized portfolio penalizes squared turnover $\|w_t - w_{t-1}\|_2^2$. Our theory predicts that this penalty produces a never-ending stream of small trades, because quadratic penalties give only exponential approach to the target. Uziel et al. freeze positions for a fixed number of days $d$ and need an emergency expert to escape the resulting lag. The proximal subgradient STCO method handles sparsity together with proportional costs. It shows that cost-unaware sparse models collapse to near-zero wealth at a 0.5% fee.

WaveLSFormer and NumHTML report frictionless results. WaveLSFormer reports Sharpe 2.157. NumHTML reports Sharpe 1.62 on earnings-call signals. They illustrate the decoupled signal-then-execute pipeline that our controller would sit behind.

Research Program

Theory. The theoretical work has three parts.

First, we prove the set-turnpike statement for stationary alpha. This means constructing the linear storage function from the steady-state dual multiplier and showing a dissipation inequality relative to $S_\gamma$. The inequality's lower bound should be linear in trade size and in distance to the set. We then derive two consequences. One is finite-time arrival. The other is a horizon-independent bound on time spent outside $S_\gamma$, in the measure-turnpike sense of Trélat and Zhang.

Second, we adapt the Zanon–Gros discounted condition to sets. The aim is to identify the range of $\gamma$, fees and risk aversion for which a finite-horizon receding-horizon controller provably settles into $S_\gamma$ instead of chattering.

Third, for alpha following a mean-reverting process with half-life $h$, we bound the cost gap between the scaled single-period rule and the multi-period optimum as a function of $h$ and the reaching time. A simple heuristic supports the expected form. A signal that decays faster than the controller can profitably trade toward it should be partly ignored. The single-period rule cannot see the decay, so it pays fees chasing it.

We will also try, as a stretch goal, to cover fixed per-trade fees. These are nonconvex, so duality-based storage functions may fail. We will report honestly whether the construction survives.

Algorithms. We build two controllers. The first is a receding-horizon controller that solves a multi-period convex program at each step. Its horizon is set from the predicted reaching time rather than tuned. The second is the closed-form single-period rule with fee multiplier $\kappa_\gamma$. A small convex program computes the dissipativity certificate for a given covariance, fee schedule and discount factor, so practitioners can check stability before deployment.

Baselines

The naive reference point is frictionless mean-variance rebalancing on the same alpha signals, evaluated after fees. It shows how much is lost by ignoring costs. We expect its net wealth to approach zero at a 0.5% fee, as the STCO and Uziel et al. experiments already document.

The strongest single-period competitor is Markowitz++-style rolling convex rebalancing with an absolute-value fee term (Boyd et al.). Its fee multiplier gets the same validation tuning budget as every other method. On S&P 100 with synthetic alpha, that paper reports a Sharpe ratio improved from 4.32 to 4.65 and maximum drawdown cut from 78.9% to 7.0%. This is the decisive competitor, because our theory says a correctly scaled version of it is optimal under stationary alpha. Beating it is therefore not expected in that regime. Matching it with no tuning, using $\kappa_\gamma$ alone, would count as success there. Beating it should happen only where alpha is short-lived. "Beating it" means higher net Sharpe and net wealth at equal or lower turnover, across five or more random seeds of synthetic alpha.

The squared-turnover baseline is the doubly regularized portfolio (Shen et al.), run with its penalty weight tuned on validation. That paper reports a Sharpe ratio of 48.27% on FF100 against 26.97% for equal weighting, and turnover reductions of 10.7% to 66.2% relative to gross-exposure-constrained portfolios.

The dilution baseline is $LSPE_{META}$ (Uziel et al.). It reports cumulative wealth of 9.3E3 on NYSE(N) at a 0.5% commission, against 31.4 for the next-best baseline. Because LSPE combines forecasting with execution, we also run a cleaner ablation: fixed-interval rebalancing every $d$ days on our own signals. This isolates the execution rule from the forecaster.

Two further ablations isolate the key ingredients. The first is the receding-horizon controller with an arbitrary short horizon instead of the reaching-time horizon. The second is the single-period rule with an untuned multiplier of 1 instead of $\kappa_\gamma$.

All methods use identical signals, data splits and fee schedules, and the same number of validation trials. The tuning budget is spent on the competitors' knobs, while our controllers use their theory-derived values.

Why not just…?

Why not just tune the fee multiplier in single-period rebalancing? This is the real threat, and the theory itself says it wins when alpha is stationary. The contribution in that regime is a prediction, not an improvement. We claim the tuned multiplier will land near $\kappa_\gamma$, and that the turnpike controller matches the tuned rule without any tuning. If tuned single-period rebalancing also matches the multi-period controller when alpha half-life is short, our persistence criterion is wrong. The synthetic sweep over $h$ is designed to settle exactly that.

Why not just trade every $d$ days? Fixed dilution ignores how far the portfolio is from the no-trade set. It trades when it need not and waits when a large signal change makes trading worthwhile. Uziel et al. needed a separate emergency expert to cover the lag. We expect fixed-interval rebalancing to show either higher turnover or larger tracking error against the moving $S_\gamma$ at matched net return. The fixed-$d$ ablation measures this directly on NYSE(N), TSE and DJIA.

Why not just penalize squared turnover? A quadratic penalty is smooth, which is convenient. However, our analysis predicts it approaches its target only exponentially and so keeps issuing small trades that each pay the proportional fee. If the doubly regularized baseline shows the same finite-time stopping and the same net performance as the absolute-value controller, this prediction fails. We count trades per period after the initial transient to check it.

Why not just learn the trading policy end to end with reinforcement learning? That approach is possible. However, the file contains no reported friction-aware reinforcement-learning trading result we could compare against. Such a policy would also inherit the stability questions that discounted dissipativity answers. We treat it as future work rather than claim superiority.

Proposed Experiments

Synthetic tests. These come first because the theory makes exact predictions there. We simulate assets with known $\Sigma$ and alpha following a mean-reverting process with half-life $h$, and sweep fee $c$, discount $\gamma$ and $h$. Three checks are decisive.

The first concerns the shape and width of the no-trade set. With stationary alpha, the empirical set where the long-horizon optimum stops trading should coincide with $S_\gamma$. Its width should scale linearly in $c\,(1-\gamma)/\gamma$.

The second concerns finite arrival and horizon-independence. Trajectories under the absolute-value fee should reach $S_\gamma$ in finitely many steps. Time spent outside it should stay constant as the horizon grows, whereas the squared-turnover baseline should keep trading indefinitely.

The third concerns the persistence criterion. The net-cost advantage of the multi-period controller over the scaled single-period rule should be near zero for large $h$ and grow as $h$ falls below the reaching time. A flat curve would refute the main practical claim.

Real data. We use daily data on NYSE(N), TSE, DJIA and MSCI, monthly FF25, FF48 and FF100, and S&P 100 through Qlib. Fees run from 0.25% to 1%. Signals come from two sources: simple momentum and mean-reversion forecasters, and a small learned forecaster. The hourly equities data used by WaveLSFormer is commercial, so we only reproduce its setting if access is available.

Metrics are net cumulative wealth, net Sharpe ratio, turnover, maximum drawdown, and the fraction of periods with any trade. The real-data test of the theory has two parts. First, the scaled single-period rule and the turnpike-horizon controller should match or exceed their tuned counterparts without tuning. Second, measured signal half-life should predict, across datasets, where the multi-period controller helps. We will report the cases where it does not.