A mean-reverting spread is a weighted basket of assets whose combined price keeps snapping back toward an average. Statistical arbitrage trades these spreads, and its best formulations balance two things. The spread must revert fast enough to complete many round trips, and it must swing widely enough that each trade clears transaction costs. Every such formulation buys some assets and sells others short. That rules them out for mutual funds, pension mandates, and short-restricted markets such as Chinese A-shares, where most capital actually sits.
This note starts from a simple observation. A long-only portfolio $w$ held against a benchmark $b$ already contains a hidden long-short spread, its active weights $a = w - b$. On that spread, "short selling" is underweighting, which is capped at each asset's benchmark weight. Because the cap is an ordinary linear bound, the reversion-versus-variance problem can be posed and solved in the long-only setting. That cap also changes how the strategy behaves as we feed it more history. We expect the bound to act as built-in shrinkage, so the long-only spread should reach its best out-of-sample performance with less data, and fall off sooner when older, regime-shifted data is added.
Core Claim
Existing mean-reverting portfolio design needs genuine shorts. Mean-Reverting Portfolio With Budget Constraint imposes dollar-neutral or net-budget constraints with unbounded negative weights. Optimal Mean-Reverting Portfolio With Leverage Constraint for Statistical Arbitrage in Finance bounds leverage by $\|w\|_1 \le L$. Finding Moving-Band Statistical Arbitrages via Convex-Concave Optimization maximizes spread variation while keeping the spread inside a band around its moving average, again under an $\ell_1$ leverage limit. Long-only methods take a different route. PAMR and OLMAR exploit reversion by rebalancing on the simplex. They never construct a stationary object and never control its variance against costs. Nobody has posed the reversion-variance trade-off with the long-only constraint built into the problem itself.
We pose it as follows. Let $p_t \in \mathbb{R}^N$ be asset prices on day $t$, let $b$ be a fixed long-only benchmark (equal-weight or capitalization-weight), and let $a$ be the active weights. The active spread price is $s_t = a^\top p_t$. We maximize its variation $\sum_t (s_t - s_{t-1})^2$, as in the moving-band formulation, subject to three requirements. First, $s_t$ must stay within a band of half-width one around its trailing moving average. Second, the active weights must be dollar-neutral, $\mathbf{1}^\top a = 0$. Third, they must satisfy the long-only bound
$$a_i \ge -b_i \quad \text{for every asset } i,$$
together with a tracking budget $\|a\|_1 \le L$. The first line of this formulation is exactly the Johansson–Schmelzer–Boyd problem. What is new is the lower bound, which ties the short side of the spread to benchmark weights. The investor's holdings $w = b + a$ are then automatically non-negative and sum to one.
The bound has an asymmetric effect. An asset can be bought without limit, but it can only be sold down to zero. The spread's short side is therefore drawn from heavy benchmark names, and assets with small benchmark weight can serve only as long legs.
Our empirical claim concerns how much history the strategy should be fit on. Let $T$ be the lookback length used to fit $a$, and let $S(T)$ be out-of-sample Sharpe after costs. The data-scaling literature finds concave, saturating utility in data volume, of the form $1 - e^{-\lambda D}$ (Optimal Pricing of Internet of Things). It also finds that extra data can actively hurt (What Makes Good Data for Alignment?). For non-stationary prices, we expect $S(T)$ to rise and then fall. We predict three things. The capped long-only spread will reach its peak at a smaller $T$ than the unconstrained dollar-neutral spread. Its peak will be lower, but its curve will be flatter on both sides. Its weights will jump less between adjacent refits than a naive simplex optimizer's.
The third prediction speaks to A New Spin on Optimal Portfolios and Ecological Equilibria. That paper shows that long-only mean-variance problems develop exponentially many near-equal solutions with almost no assets in common, so tiny estimation errors flip the chosen portfolio. It leaves open whether more data cures this. We test that question directly on a reversion objective. We also separate the case where non-negativity is applied to the whole portfolio from the case where it is applied only to active weights around a benchmark. We expect the benchmark anchor to break much of the degeneracy, because the optimizer no longer chooses among many equivalent corners of the simplex.
Research Program
The solver adapts the convex-concave procedure of Johansson et al.. At each step the convex variation objective is linearized and the problem is solved as a linear program. The procedure handles the new bound without modification, because each subproblem remains a linear program.
For a second objective family, we add the same bound to the predictability and portmanteau criteria. These are measures of how much of the spread's movement is explained by its past, and the minimum-variance threshold of Mean Reversion with a Variance Threshold applies to them. We solve these with the successive convex approximation and alternating-direction machinery of Zhao & Palomar (2019), where box constraints enter the projection step. One caveat: the semidefinite relaxation that Cuturi and d'Aspremont prove exact without bounds generally loses exactness once inequality bounds are added. We will report the relaxation gap rather than assume it is zero.
Execution uses band-crossing entries and exits. Proportional costs are swept from 0 to 50 basis points. Turnover is charged on the full holdings $w$, not just on $a$, because the investor actually trades $w$.
The scaling study refits every strategy on rolling lookbacks from roughly three months to ten or more years. We record four outcomes: net out-of-sample Sharpe, maximum drawdown, the fraction of lower bounds that are active, and weight stability between consecutive refits, measured as cosine overlap. We fit saturating and peaked curves to $S(T)$ and read off each method's peak lookback. To show the proposal is more than lookback tuning, we include a curated-window variant in the spirit of Capturing the Temporal Dependence of Training Data Influence. It drops historical segments whose removal does not change the fitted spread's validation reversion statistic.
Baselines
Naive reference points. These are the benchmark $b$ itself and buy-and-hold, run on identical universes and cost settings. Any active spread that fails to beat $b$ after costs adds nothing.
Strongest long-only online competitor. This is OLMAR, which reverts to a multi-day moving average on the simplex and reports cumulative wealth of $3.68 \times 10^{16}$ on NYSE (O), with tolerance to roughly 0.25–0.5% costs. We pair it with Online Lazy Updates, which builds an $\ell_1$ turnover penalty into the update and reports wealth of 901.00 (costless) on its 263-stock S&P 500 set. These are the strongest existing methods that obey the same constraint. We tune their window and penalty on the same validation folds, with the same grid size as our band width and $L$. Beating them means higher net Sharpe and lower drawdown at costs of 10 basis points and above. Their raw costless wealth is driven by compounding and is not the target.
Strongest unconstrained reference. This is the moving-band method of Johansson et al.. On CRSP 2010–2023 it reports a Sharpe ratio of 0.84, a 12% maximum drawdown, and 79% profitable portfolios, using genuine shorts. We rerun it ourselves on our universes. It serves as an upper bound: the target is to retain a large fraction of its net Sharpe while staying long-only, not to exceed it.
Short-restricted reference. On A-shares we compare against Quantformer, whose long-only factor QF_Month_1 reports a Sharpe ratio of 0.915. The setups differ in rebalancing frequency and stock universe, so we treat this number as context, not as a head-to-head result. No reported result was found on Lacuna for any long-only mean-reverting spread method on these benchmarks.
Key ablation. We solve the unconstrained dollar-neutral spread, then make it long-only after the fact. Negative holdings in $b + a$ are clipped and the result is renormalized; separately, it is projected onto the simplex. This isolates whether building the bound into the optimization matters.
Why not just…?
Why not just run OLMAR or PAMR, which already exploit mean reversion long-only? They react to single-day or moving-average reversion of individual prices. They never check that the combination they hold actually reverts, and they never ensure it swings enough to pay costs. Their own papers report weakness on trending markets: PAMR underperforms on the Dow Jones data (PAMR), and OLMAR documents single-period reversion crashing on that same market. If OLMAR matches our method's net Sharpe at realistic costs, the spread construction adds nothing, and we will say so. The cost sweep settles this question.
Why not just solve the usual long-short spread and clip it to long-only? This is the most serious threat, because it is trivial to implement. Clipping changes the spread weights after the fact, so the resulting portfolio carries no guarantee of stationarity or of meeting the variance floor. Its effect should be largest when many bounds would be active, which happens in capitalization-weighted universes with long tails of small names. The ablation reports performance as a function of the fraction of active bounds. If clipping is equivalent everywhere, the formulation reduces to a heuristic and the contribution shrinks to the scaling study.
Why not just use the longest history available, or pick the lookback by validation? The saturation and degradation results cited above, together with regime breaks noted in Zhao & Palomar (2018), argue against blindly using everything. Validated lookback selection is a fair competitor, so every method receives it. Our claim is narrower: the peak lookback and the flatness of $S(T)$ differ systematically between capped and uncapped spreads. That is a measurable difference, not something lookback selection can remove.
Why not just allow shorting, as leverage-enabled decoders do? RAT escapes the simplex by permitting negative weights. The target investors cannot do that, and DeepTrader notes that short restrictions limit hedging on CSI100. The unconstrained baseline measures exactly what the restriction costs.
Proposed Experiments
We use four datasets. CRSP 2010–2023 allows direct comparison with Johansson et al. NYSE (O) and the 263-stock S&P 500 set allow comparison with OLMAR and OLU. The Chinese A-share universe from 2010–2023 is a market where short restrictions actually bind. Each runs under both equal-weight and capitalization-weight benchmarks.
Three checks decide the outcome. The first asks whether the capped spread retains at least half of the unconstrained moving-band strategy's net Sharpe while beating OLMAR and OLU at costs of 10 basis points or more. The second asks whether building in the bound beats post-hoc clipping by a margin that grows with the fraction of active bounds. The third asks whether $S(T)$ for the capped spread peaks at a shorter lookback than for the unconstrained spread, with higher weight overlap between refits than full-portfolio simplex optimization achieves.
If the third check holds, it answers the open question from Garnier-Brun et al. empirically. Adding history does not fix long-only instability, but anchoring the active weights to a benchmark substantially does. If it fails, that failure is itself informative: it would mean that, for reversion objectives, longer history rather than anchoring is the remedy.
References
- Mean-Reverting Portfolio With Budget Constraint
- Optimal Mean-Reverting Portfolio With Leverage Constraint for Statistical Arbitrage in Finance
- Finding moving-band statistical arbitrages via convex–concave optimization
- PAMR: Passive aggressive mean reversion strategy for portfolio selection
- On-Line Portfolio Selection with Moving Average Reversion
- Optimal Pricing of Internet of Things: A Machine Learning Approach
- What Makes Good Data for Alignment? A Comprehensive Study of Automatic Data Selection in Instruction Tuning
- A New Spin on Optimal Portfolios and Ecological Equilibria
- Mean Reversion with a Variance Threshold
- Capturing the Temporal Dependence of Training Data Influence
- Online Lazy Updates for Portfolio Selection with Transaction Costs
- Quantformer: From attention to profit with a quantitative transformer trading strategy
- Relation-Aware Transformer for Portfolio Policy Learning
- DeepTrader: A Deep Reinforcement Learning Approach for Risk-Return Balanced Portfolio Management with Market Conditions Embedding