💰 quantitative finance

Portfolio Preference Elicitation in Institutional Crossing Markets

This paper proposes and validates a hybrid preference elicitation mechanism for institutional crossing markets that combines price-directed demand queries with value verification to effectively navigate hidden-information problems in nonseparable portfolio spaces, demonstrating that such a combined approach significantly outperforms single-method designs in recovering social welfare while highlighting the trade-off between security-level and factor-based package representations based on disclosure costs.

Original authors: Yoontae Hwang

Published 2026-05-21

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yoontae Hwang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a massive, high-stakes game of Poker, but instead of playing for chips, institutional investors (like big pension funds) are trying to trade entire portfolios of stocks at once.

The problem is that these investors don't just want to buy or sell one stock; they want to move a whole hand of cards. Maybe they need to sell Apple to buy Microsoft to balance their risk, or sell a tech stock to buy a utility stock to hedge against a market crash. The value of the trade depends entirely on the combination of stocks, not just the individual pieces.

However, the "marketplace" (the crossing platform) usually only asks about one stock at a time. It's like asking a poker player, "Do you want to bet on the Ace of Spades?" when the player is actually trying to figure out if they have a winning hand. If the platform only looks at single stocks, it misses the big picture and might make bad trades that hurt the investors.

This paper proposes a new way for these platforms to talk to investors. It treats the process like a detective solving a mystery with a limited number of clues.

The Detective's Toolkit: Search vs. Verification

The paper argues that finding the best trades requires two different types of detective work, and you need both to win:

The "Search" (Demand Queries):
- The Analogy: Imagine a treasure hunter using a metal detector. They sweep the beach (the market) to find where the treasure might be buried. They don't dig yet; they just find the "hot spots."
- In the Paper: The platform asks investors, "If the price of these stocks were X, what would you buy or sell?" This helps the platform find the general areas where investors want to trade. It's fast and covers a lot of ground, but it's not precise. It tells the platform where to look, but not exactly how much the trade is worth.
The "Verification" (Value Queries):
- The Analogy: Once the metal detector beeps, the treasure hunter digs a hole and pulls out a gold bar. They weigh it on a scale to know its exact value.
- In the Paper: The platform picks a specific trade package found during the "search" phase and asks, "Exactly how much profit (surplus) would you make if we did this specific trade?" This gives the exact number needed to make a final decision.

The Big Discovery:
If you only do "Search" (ask about prices), you find good spots but don't know the exact value, so you can't be sure the trade is worth it. If you only do "Verification" (ask for exact values), you might be digging holes in the wrong places because you didn't know where to look.

The Paper's Result: The best strategy is a Hybrid. Use the metal detector to find the hot spots, then dig and weigh the gold. When they tested this, the hybrid method recovered 88% of the possible value, whereas using only one method recovered less than 50%.

The "Bridge" Step: Don't Forget What You Found

There is a special step called the "Bridge Value Query."

The Analogy: Imagine the treasure hunter finds a great spot during the search phase and marks it. Before they start digging for new spots, they go back and weigh the gold at that first spot right now. This ensures that even if they find a slightly better spot later, they haven't lost the value of the first one.
In the Paper: After the "search" phase finds a good temporary deal, the platform asks for the exact value of that deal immediately. This locks it in as a "safe option" so the system doesn't accidentally discard a good trade while looking for a perfect one.

The "Menu" Problem: How to Describe the Trade

The paper also asks: How should we ask the questions? Should we list every single stock in the portfolio, or should we use a simpler description?

Security-Level Packages (The "Exact Menu"):
- The Analogy: Ordering a meal by listing every single ingredient: "I want 3.2 grams of salt, 14.5 grams of flour, and 2 eggs."
- When it works: If it's cheap and safe to reveal exactly what you want, this is the most efficient way. It gets the best deal (97% efficiency).
Factor-Completed Baskets (The "Cheat Sheet"):
- The Analogy: Ordering a meal by saying, "I want a 'Healthy Breakfast' with a little extra protein." You don't list the exact grams; you describe the type of meal, and the chef fills in the rest.
- When it works: If revealing your exact ingredients is dangerous (like revealing your secret trading strategy to competitors), this method is better. It hides the specific details while still getting a good deal. The paper shows that if the "cost" of revealing secrets is high, this "cheat sheet" method actually results in a better overall outcome than the exact menu.

The Bottom Line

This paper solves a communication puzzle for big investors. It shows that:

You need both a map and a scale. You need to search for where trades are possible and verify their exact value. Doing just one is inefficient.
The "Bridge" is crucial. You must lock in the value of your best find before moving on.
The "Menu" matters. If you are worried about leaking your secrets, it's better to describe your trade in broad categories (like "Tech Growth") rather than listing every single stock, even if the broad description is slightly less precise.

By using this "Search + Verify + Bridge" approach, platforms can help investors trade their complex portfolios much more effectively, recovering nearly 95% of the value they could get if the platform knew everything about them perfectly.

Technical Summary: Portfolio Preference Elicitation in Institutional Crossing Markets

1. Problem Definition

Institutional investors frequently trade portfolios rather than isolated securities, valuing trades based on nonseparable portfolio objectives such as risk reduction, benchmark deviation, and hedge completion. However, standard hidden liquidity discovery mechanisms (e.g., dark pools, crossing networks) typically elicit information about individual securities. This creates a fundamental mismatch: the object of value is a signed portfolio trade, but the information unit elicited is a single security order.

The core problem is hidden-information preference elicitation in a continuous, signed portfolio space. A platform must identify mutually beneficial trades without observing full demand. The challenge lies in balancing two conflicting requirements:

Search: Locating high-value regions in a high-dimensional, nonseparable portfolio space.
Verification: Obtaining exact welfare values to certify allocations, as demand responses alone provide only conservative lower bounds on welfare.

The paper frames this as a query-based mechanism design problem where the platform must decide what limited portfolio information to elicit (via demand and value queries) and how to represent the packages (via security-level vs. factor-completed baskets) to maximize welfare under communication constraints.

2. Methodology and Architecture

2.1 Model Framework

The authors model institutional crossing as a signed multi-asset allocation problem.

Participants: $n$ investors with private surplus functions over signed trade vectors $d \in \mathbb{R}^m$ .
Valuation: The surplus is modeled as a quadratic function: $v_i(d) = \theta_i^\top d - \frac{1}{2}d^\top H_i d$ , where $H_i$ captures curvature from risk, liquidity, and mandates.
Nonseparability: Proposition 1 establishes that trade legs are complementary or substitutable based on cross-terms in the curvature matrix ( $g^\top H_i h$ ). This implies that a trade's value depends on the existing portfolio position and other legs in the package.
Welfare Benchmark: The full-information benchmark ( $W^\star$ ) maximizes total surplus minus residual external execution costs ( $\Psi$ ).

2.2 Elicitation Architecture

The proposed mechanism separates search from verification using a hybrid query protocol:

Demand Queries: The platform posts price vectors $p^\ell$ . Participants respond with optimal trades $d^\ell_i$ maximizing $v_i(d) - p^{\ell\top}d$ . These queries use price-directed responses to locate active regions of the portfolio space.
Surrogate Search: A structural surrogate model ( $\hat{v}_i$ ) is updated using demand responses and projected moments to guide the selection of subsequent query prices.
Bridge Value Query: After the demand query phase, the platform identifies an interim allocation ( $a^{dq}$ ). A "bridge" query asks participants to report the exact surplus of this specific allocation. This step converts the demand-discovered allocation into an exact-valued candidate, preserving it as a feasible option for the final allocation.
Value Queries: The platform issues targeted queries for specific packages (proposed by the surrogate model) to obtain exact surplus values.
Report-Based Allocation: The final allocation is chosen from the finite set of elicited reports (demand lower bounds and exact value reports) by solving a winner determination problem. The surrogate model guides which packages are queried but does not determine the final welfare; the final welfare is determined strictly by the elicited reports.

2.3 Package Representations

The paper compares two representations for value query proposals:

Security Level (SL): Uses demand-discovered securities as coordinates. This preserves exact security-level expressiveness.
Factor Completed (FC): Combines factor exposure targets with a residual sleeve over demand-discovered securities. This uses a lower-dimensional factor guide ( $F$ ) to summarize risk exposures, completing the portfolio via a minimum-cost map ( $C(u, \eta)$ ).

3. Key Contributions

Framing Hidden Liquidity as Portfolio Elicitation: The paper shifts the design focus from single-order matching to signed portfolio trade elicitation, recognizing that optimal trades depend on current positions and portfolio curvature.
Hybrid Demand-Value Architecture: It develops and analyzes a protocol that combines price-directed demand queries (for search) with value queries (for verification). The analysis clarifies that demand queries locate promising regions but are insufficient for welfare accounting without verification, while value queries are ineffective without targeted search.
Package Representation as a Design Variable: The paper treats the choice between security-level and factor-completed baskets as an economic design choice. It distinguishes between unadjusted efficiency (welfare before disclosure costs) and disclosure-adjusted welfare (welfare penalized by the informativeness of pretrade messages).

4. Experimental Results

The authors conduct computational experiments using equity panels from the S&P 500, KOSPI 200, Nikkei 225, and DAX.

4.1 Search and Verification Complementarity

Hybrid vs. Single-Mode: Under a limited query budget (18 queries), hybrid elicitation recovers 88.01% of the full-information welfare. In contrast, value-only elicitation recovers only 47.84%, and demand-only elicitation recovers 49.56%.
Scaling: As the communication budget expands to 48 queries, hybrid elicitation approaches 95.29%, while value-only remains stagnant around 49.86%.
Mechanism: Demand queries successfully locate high-value regions (primitive valuation of demand-discovered candidates reaches 96.71%), but without the bridge value query and subsequent verification, the conservative lower-bound accounting prevents these candidates from being fully utilized in the allocation rule.

4.2 Package Representation and Disclosure Costs

Unadjusted Efficiency: When pretrade disclosure is inexpensive, Security Level (SL) packages are the most efficient, achieving 97.25% relative welfare. This outperforms Same Securities (SN) crossing (94.69%) and Factor Completed (FC) baskets (88.85%).
Disclosure-Adjusted Frontier: When pretrade message informativeness is costly, FC becomes superior. The paper identifies a "break-even" disclosure cost ( $c^\star$ ) where the reduction in information leakage from using factor baskets compensates for their lower unadjusted efficiency. For example, at a leakage cost of $c=0.20$ , FC becomes the welfare leader.
Liquidity Regimes: In "thin" contra-side liquidity regimes (where direct opposing interest in the same security is scarce), SN crossing performance drops significantly (to ~~53%), while portfolio packages (SL and FC) maintain high welfare (~~72-74%) by combining legs that offset each other through the portfolio curvature matrix.

4.3 Robustness

Bridge Value Query: Including the bridge value query improves welfare by 5.61 to 7.03 percentage points under tight budgets, confirming the importance of preserving the demand-discovered interim allocation as an exact candidate.
Factor Completion: The FC construction (combining factors with a residual sleeve) significantly outperforms incomplete basket alternatives (e.g., Low Rank or No Completion), demonstrating that the residual sleeve is necessary to capture participant-specific active demand.

5. Significance and Claims

The paper claims to characterize portfolio crossing as a selective verification problem. Its primary significance lies in demonstrating that:

Search and Verification are Distinct but Complementary: Neither demand queries nor value queries alone are sufficient for efficient hidden liquidity discovery. The hybrid protocol is necessary to navigate the high-dimensional portfolio space and certify welfare.
Representation is a Trade-off: There is no universal "best" package representation. Security-level packages maximize unadjusted efficiency, while factor-completed baskets offer a disclosure-adjusted welfare frontier when pretrade information leakage is costly.
Design Implications: For hidden liquidity platforms, the choice of package representation (SL vs. FC) should be driven by the cost of pretrade disclosure and the thickness of the contra-side liquidity, rather than a fixed assumption of indivisibility or single-security focus.

The authors conclude that their results provide a design framework for platforms to balance the need for nonseparable value discovery with the constraints of information leakage, moving beyond simple order matching to sophisticated portfolio preference elicitation.