Sparse Portfolios That Directly Optimize Drawdown Ratios

Investors often judge a portfolio by ratio metrics such as the Calmar ratio (average return divided by the worst peak-to-trough loss, called maximum drawdown) or the Sortino ratio (average return divided by downside deviation). They also want the portfolio to hold only a few assets, to keep trading and custody costs down. Both requirements break standard tools. Portfolio weights live on the probability simplex: they are non-negative and sum to one. On that set the usual $\ell_1$ sparsity penalty equals one everywhere, so it cannot push any weight to zero. The ratio metrics have kinks and are not concave, so gradient methods built for smooth tracking error do not apply. Current sparse-simplex methods therefore optimize a smooth proxy, such as squared tracking error, conditional value-at-risk (CVaR, the average loss in the worst tail of days), or linear transaction costs, and then only report drawdown after the fact.

This note proposes to optimize the ratio itself under a hard limit of $K$ assets. The approach rests on one structural observation. Drawdown measured on cumulative simple returns is a maximum of linear functions of the weights, so it is convex. A ratio of linear return to convex risk can be turned into a short sequence of "return minus $\lambda$ times risk" problems using Dinkelbach's method, and this sequence finds the best ratio on any feasible set, including the non-convex $K$-sparse simplex. All non-convexity then sits in a single inner problem: choosing a $K$-asset support for a convex, non-smooth objective. Once a support is fixed, that problem is a linear program that can be solved exactly. The resulting algorithm, which we call Fractional Sparse Simplex Ascent (FSSA), alternates between choosing $\lambda$, choosing the support, and solving the linear program.

Core Claim

Let $r_\tau \in \mathbb{R}^N$ be the vector of asset returns on day $\tau$ over a training window of $T$ days. Let $w$ be the portfolio weights, and let $\mu$ be the mean return vector. Define the additive drawdown as

$$D(w) = \max_{0 \le s \le t \le T} \sum_{\tau=s+1}^{t} \left(-r_\tau^\top w\right),$$

the largest cumulative loss over any contiguous stretch of days. $D$ is a pointwise maximum of linear functions of $w$, so it is convex and piecewise linear. It can also be written with $O(T)$ auxiliary "running peak" variables and linear constraints. We want to maximize the Calmar-type ratio $\rho(w) = \mu^\top w / D(w)$ over $\Delta_K = \{w \ge 0, \mathbf{1}^\top w = 1, \|w\|_0 \le K\}$, where $\|w\|_0$ counts the nonzero weights. Dinkelbach's method defines

$$F(\lambda) = \max_{w \in \Delta_K} \; \mu^\top w - \lambda D(w).$$

The optimal ratio $\lambda^\ast$ is the unique root of $F(\lambda) = 0$, and the updates $\lambda \leftarrow \rho(w_\lambda)$ converge to it. This follows from a standard fractional-programming argument that uses only the positivity of $D$ and exact inner maximization, not convexity of the feasible set. Sortino- and Omega-type ratios fit the same template, because downside deviation and expected shortfall below a threshold are also convex in $w$.

The claim has three parts. First, a solver built on this reduction can reach, or come very close to, the certified best in-sample ratio found by an exact mixed-integer solver on small universes. Second, it scales to universes where the mixed-integer solver cannot. Third, and most importantly, optimizing drawdown ratios directly produces a different and better out-of-sample support than CVaR- or tracking-error-based sparse methods at the same $K$. If the third part fails, the whole program is not worth doing.

Relation to Prior Work

The closest works fall into two groups.

The first group solves sparse-simplex portfolio problems with non-smooth but convex objectives. Autonomous Sparse Mean-CVaR Portfolio Optimization minimizes CVaR with a continuous "tailed" approximation of the cardinality count, solved by a proximal alternating method. A Linearly Convergent Proximal Subgradient Algorithm for Sparse Portfolio Optimization with Transaction Cost writes the $K$-sparse constraint as an exact difference-of-convex penalty and handles piecewise-linear transaction costs. Both motivate this proposal by showing that exact or near-exact sparsity beats $\ell_1$-style relaxation on real market data. Neither addresses path-dependent drawdown or any ratio objective. Both need a penalty scale tuned by hand, whereas FSSA enforces $K$ directly and has no penalty scale to tune.

The second group handles sparsity differently but optimizes smooth tracking losses. DCC: Differentiable Cardinality Constraints for Partial Index Tracking uses a steep sigmoid for the cardinality count. Index Tracking with Cardinality Constraints: A Stochastic Neural Networks Approach samples supports with a straight-through Gumbel-Softmax estimator. Partial Index Tracking: A Meta-Learning Approach differentiates through a quadratic program and explicitly lists drawdown-type objectives as future work. These papers constrain our design. Composing a steep sigmoid or a low-temperature relaxation with a kinked metric yields gradients that are zero almost everywhere or have very high variance, so FSSA uses no differentiable cardinality surrogate.

The support-selection step borrows its method from Sparse projections onto the simplex. That paper shows that Euclidean projection onto the $K$-sparse standard simplex keeps the $K$ largest coordinates and projects them onto the smaller simplex. The NP-hardness result in Learning Sparse Distributions using Iterative Hard Thresholding applies to projection onto more general sparse distribution sets. We therefore use the exact top-$K$ projection and do not claim more than it provides.

The fractional reformulation mirrors a result in Principled Algorithms for Optimizing Generalized Metrics in Binary Classification. That paper shows that ratio metrics such as $F_\beta$ reduce to finding a scalar $\lambda^\ast$ and minimizing a weighted cost-sensitive loss. We apply the same idea to a portfolio setting with a non-convex sparse feasible set, which that paper never considers.

Research Program

The inner step for a fixed $\lambda$ has three stages.

First, the solver searches for a support. It runs projected subgradient ascent on $\mu^\top w - \lambda D(w)$. The subgradient of $D$ is the negated sum of returns over the currently worst drawdown stretch. Each iterate is projected onto $\Delta_K$ with the exact top-$K$ projection. A smoothed variant replaces the maximum in $D$ with a log-sum-exp at temperature $\eta$, which gives more stable support signals early on.

Second, the solver polishes the weights. It freezes the support and solves the resulting linear program exactly, using the running-peak formulation of $D$.

Third, the solver refines the support. It tries single-asset swaps ranked by subgradient magnitude and accepts a swap only when the linear-program value improves.

The outer Dinkelbach loop warm-starts each inner step from the previous support. We expect under ten outer iterations. We will analyze two properties. The first is the conditions under which the outer loop still converges monotonically when the inner step is only locally optimal: the ratio is monotone if each inner solution has $F$ at least that of the previous iterate, which the warm start and accept-if-improving swaps guarantee. The second is how far FSSA falls short of the exact optimum, measured against mixed-integer certificates on small universes. We will also measure support stability across rebalancing dates, since turnover is a known weakness of greedy selection.

Baselines

The naive reference points are two portfolios. The first is the equal-weight $1/N$ portfolio, which Autonomous Sparse Mean-CVaR Portfolio Optimization used as a baseline. The second is a $K$-asset market-capitalization or highest-individual-Calmar selection followed by the exact linear program on that support. These show whether any joint optimization is needed.

The strongest existing method is ASMCVaR. It is the best-performing sparse, non-smooth, simplex-constrained method in the file, and CVaR is the most natural proxy for tail and drawdown risk. The file reports only that on FF100 it achieved nearly 10x the final wealth of Mean-CVaR and Mean-CVaR-$\ell_1$, the highest Sharpe ratios on all six French datasets, and over 89% asset overlap across sparsity targets. We found no reported Calmar or maximum-drawdown number on Lacuna for ASMCVaR or any other method on these datasets, so we will rerun it ourselves. The second strong baseline is STCO from A Linearly Convergent Proximal Subgradient Algorithm for Sparse Portfolio Optimization with Transaction Cost. Its sparsity behavior is notable: it averaged about 1.05 assets on NYSE(O) under a bound of $K=10$, which we will track as a possible degenerate-concentration failure mode.

To make the comparison fair, all methods see the same rolling training windows, rebalancing dates, transaction-cost model (0.25% and 0.5%), and realized cardinality $K$. Competitor penalty scales are tuned on the same validation windows with the same budget of solver calls that FSSA spends on its temperature $\eta$ and swap budget.

The ablations isolate the key ingredients. The first optimizes return minus $\lambda D$ with a fixed $\lambda$ chosen by validation instead of Dinkelbach updates. The second removes linear-program polishing. The third replaces drawdown with CVaR inside FSSA, which separates the effect of the objective from the effect of the solver.

Beating ASMCVaR means a statistically significant improvement in out-of-sample Calmar ratio across walk-forward windows, with no worse out-of-sample Sortino ratio or cost-adjusted final wealth, on a majority of the ten benchmarks.

Why not just…?

Why not just minimize CVaR with ASMCVaR and let drawdown follow? This is the most serious threat. CVaR penalizes bad individual days, while drawdown penalizes runs of moderate losses, and a portfolio can have low CVaR and still suffer long drawdowns. We have no evidence yet that the difference matters out of sample, and that is exactly what the CVaR-inside-FSSA ablation tests. If FSSA with CVaR matches FSSA with drawdown on out-of-sample Calmar, the headline claim is false and the contribution shrinks to a faster solver.

Why not just sweep $\lambda$ on the scalarized objective and pick the best by validation? This alternative plausibly works as well out of sample, because Dinkelbach's method only locates the in-sample optimum more efficiently. We treat it as an ablation rather than dismiss it. The expected gain from Dinkelbach is roughly 3–5x fewer inner solves than a 20-point grid, plus a certified in-sample ratio. If the grid wins out of sample, we will report that the in-sample fractional optimum overfits.

Why not just select $K$ assets heuristically and then optimize? On tracking tasks, two-stage selection loses badly: backward selection reached MAE 7.5871 versus 3.5385 for DCCfpp on S&P 100 at $K=25$ in DCC: Differentiable Cardinality Constraints for Partial Index Tracking. Drawdown depends on how asset returns co-move over time, so assets that each look good on their own may not combine well. The naive baseline tests whether this intuition holds for drawdown.

Why not just use a mixed-integer solver? For drawdown ratios this is a genuine alternative, because the running-peak formulation makes each Dinkelbach step a mixed-integer linear program with $N$ binary variables. On NYSE(O), NYSE(N), MSCI and FF25 we expect an exact solver to finish, and we will use it as the ground-truth oracle rather than as a straw man. FSSA is justified only where the exact solver times out, on FF100, TSE and S&P 500, or where frequent rebalancing multiplies the number of solves. We will report the crossover point in $N$.

Proposed Experiments

The first set of experiments covers the six Kenneth French datasets from Autonomous Sparse Mean-CVaR Portfolio Optimization and NYSE(O), NYSE(N), TSE and MSCI from A Linearly Convergent Proximal Subgradient Algorithm for Sparse Portfolio Optimization with Transaction Cost. We use walk-forward rebalancing, $K \in \{5, 10, 15, 20\}$, and both cost levels. The metrics are out-of-sample Calmar, Sortino, maximum drawdown, cost-adjusted final wealth, turnover and realized cardinality. Significance is assessed with a block bootstrap over rebalancing windows.

The second set tests scaling on the S&P 500 daily data used by DCC: Differentiable Cardinality Constraints for Partial Index Tracking. Here we report runtime and in-sample ratio against a time-limited exact solver.

Three checks decide the outcome. On small universes, the FSSA in-sample ratio should fall within 2% of the mixed-integer optimum. If it does not, the support search is inadequate. FSSA with drawdown should beat both FSSA with CVaR and ASMCVaR on out-of-sample Calmar at matched $K$. If it does not, direct optimization of the metric adds nothing. Finally, the support overlap between FSSA and ASMCVaR should be well below the 89% self-overlap ASMCVaR reports. If the supports are nearly identical, any performance gap is coming from weight polishing and not from choosing different assets.