One Basket, Many Experts: Decorrelated Re-ranking for Financial Asset Recommendation

Ryuki Unno, Keigo Sakurai, Koshi Watanabe, Keisuke Maeda, Takahiro Ogawa, Miki HaseyamaarXiv 2026 · arXiv

TL;DR

Researchers have developed DSER (Decorrelated-Satellite Expert Re-ranking), a simple, training-free method to improve financial asset recommendations by trading off a bit of user relevance for significantly better risk-adjusted returns. By replacing just a few popular, highly correlated assets in a recommended list with decorrelated "satellites," providers can offer portfolios that perform better without losing the interest of retail investors.

What Problem Does This Paper Address?

Most recommendation systems (like those for movies or products) are optimized for relevance: how likely is a user to "click" or "buy"? In Financial Asset Recommendation (FAR), relevance is crucial for user engagement, but it’s only half the story. The other half is portfolio performance: how well do those recommended assets actually perform as a group?

The authors identify a major conflict: the most "relevant" assets—those that popular experts and typical users are already buying—tend to be a handful of crowded, highly correlated names. If a recommender only suggests these, the user ends up with a poorly diversified "basket" where all assets move in lockstep. This caps the potential risk-adjusted return (the Sharpe ratio). Existing solutions often try to bake diversification into complex, "black-box" machine learning models, making it hard for financial providers to explain or audit the trade-offs they are making.

Methods

The authors propose DSER, a plug-in re-ranker that sits on top of an existing "expert-following" baseline. The process is elegantly simple and governed by a single integer "knob," $m$:

  1. The Core: Take a top-$k$ list of assets recommended by following high-performing investors. Keep the top $k-m$ items as a "relevance core" to ensure the user still sees familiar, popular assets.
  2. The Satellites: Replace the remaining $m$ slots with "satellites."
  3. Greedy Selection: These satellites are chosen one by one from a pool of liquid assets. The algorithm picks the asset with the lowest mean Pearson correlation to the assets already in the basket.
Figure 1: The DSER pipeline. An expert-following ranking is split into a fixed relevance core (the top 𝑘−𝑚items) and 𝑚satellite slots. DSER re-ranks only the satellites, filling each slot with the asset of minimal correlation to the current basket (arg min 𝜌). The knob 𝑚sets how much of the list is decorrelated.

Because $m$ is a simple integer (e.g., replace 2 out of 10 slots), the trade-off between relevance and performance is transparent and auditable.

Datasets

The study uses FAR-Trans, currently the only public dataset that pairs real retail brokerage transactions with daily price data (covering ~29,000 customers and 321 assets from 2018 to 2022). To ensure realistic results, the authors curated a subset called U92, consisting of the 92 most liquid, exchange-listed equities, preventing the algorithm from "cheating" by picking illiquid stocks with artificially low correlation.

Key Insights and Ideas

  • The Diversification Budget: The "aha moment" is that reducing correlation creates a "budget." A less correlated basket is naturally less volatile. Instead of just taking that lower volatility, DSER uses the budget to admit higher-volatility assets that offer higher returns. The result is a basket with flat volatility but significantly higher total returns.
  • Asymmetric Trade-offs: The cost in relevance (measured by nDCG) is consistent across different market phases. However, the performance gain is phase-dependent—it's a massive win during rallies but can be neutral or slightly negative during specific rebounds or drawdowns.
  • Simplicity Wins: The authors compared their integer "knob" to a continuous parameter version (similar to Maximal Marginal Relevance or MMR). They found that the hard integer threshold actually performs better in the "deployable band" where providers operate, as it concentrates the "relevance hit" in the lowest-ranked slots.
Figure 4: Integer 𝑚vs. continuous 𝜆(U92, 𝑛=10). DSER (stars) and a 𝜆-swept MMR with the same criterion (line) coincide at the endpoints, and within the deployable band (shaded, nDCG≥0.34) DSER lies above MMR at matched relevance.

Key Results

In head-to-head comparisons against a wide array of baselines (including collaborative filtering and graph neural networks), DSER achieved the best balance of relevance and performance.

  • The Sweet Spot: Setting $m=2$ (replacing 2 slots in a 10-asset list) improved the Sharpe Ratio (SR) by +0.27 and Return on Investment (ROI) by +0.031, at a modest relevance cost of only -3.6% nDCG.
  • Performance Peak: At $m=4$, the Sharpe Ratio jumped from 1.794 to 2.119.
  • Combined Metric: Using a weighted score that balances both axes, DSER ($m=2$ and $m=4$) outperformed all other baselines, including specialized hybrid models like MVECF and PfoTGNRec.
The relevance-performance landscape. DSER (red stars) occupies the optimal top-right frontier.

Limitations and Future Work

  • Single Dataset: While FAR-Trans is the standard, the results are currently limited to this one dataset.
  • Capacity Constraints: The "satellite" assets are less liquid than the "core." The authors show the gains survive trading costs for individual retail portfolios (up to ~€100k), but if a massive broker recommended the same satellites to millions of users, the market impact would likely erode the gains.
  • Offline Evaluation: The relevance metrics are "offline"—they measure what users bought without seeing the DSER list. An online A/B test is needed to see if users actually accept these decorrelated suggestions.
  • Personalization: Future work could involve personalizing the satellites based on a user's existing portfolio rather than just the recommended basket.

Summary updated

Abstract

Financial asset recommendation for retail investors is typically optimized for relevance, that is, how well a recommendation matches what each user will buy. From the service provider's point of view, the value of a recommendation also depends on portfolio performance: the risk-adjusted return the recommended assets (the basket) deliver, which drives user engagement and fee revenue. The two goals are rarely achieved jointly, because the recommendations that best match user behavior concentrate on a few popular assets whose returns are highly correlated, which caps the risk-adjusted return a provider can offer. We propose Decorrelated-Satellite Expert Re-ranking (\textbf{DSER}), which retains the most relevant recommendations from an expert-following list and replaces the rest with return-decorrelated alternatives. The number of replaced slots is a single integer~$m$ (a knob) that the provider sets, so the relevance--performance trade-off becomes an explicit, auditable choice rather than a property hidden inside a learned model. In experiments on the public FAR-Trans retail-brokerage dataset, under the primary protocol, replacing a few slots improves risk-adjusted return at a small relevance cost, and the gain survives realistic trading costs at typical retail portfolio sizes. Minimizing correlation would lower basket volatility, but in our data the least-correlated assets are more volatile and higher-return, so their volatility offsets that reduction while their return drives the gain, not lower volatility. The resulting trade-off is asymmetric, with the relevance cost consistent across market phases while the gain is phase-dependent. The provider sets $m$ to read off exactly how much relevance each operating point costs.
BibTeX
@misc{unno2026one,
  author = {Ryuki Unno and Keigo Sakurai and Koshi Watanabe and Keisuke Maeda and Takahiro Ogawa and Miki Haseyama},
  title = {{One Basket, Many Experts: Decorrelated Re-ranking for Financial Asset Recommendation}},
  year = {2026},
  eprint = {2611.05952},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.05952},
}
Figure