acceptodds
Withdrawn from ICLR 2027

AlphaRouter: An Expert Routing Approach for Adaptive Portfolio Selection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Portfolio selection requires adapting asset allocations to changing market conditions while controlling risk and satisfying investment constraints. Existing learning-based methods often predict asset weights directly from market data, making portfolio decisions sensitive to noisy and nonstationary observations. We propose AlphaRouter, an adaptive portfolio selection framework that separates portfolio construction from market-dependent selection. AlphaRouter maintains a set of complementary experts, each capturing distinct historical market patterns to produce a complete portfolio under shared investment constraints. To overcome the high optimization variance of single neural networks in noisy markets, an ensemble of independently trained Transformer routers evaluates expert suitability from recent states, aggregating their calibrated decisions through majority voting to form a robust consensus. A historically best-performing expert serves as a stable reference, and routing adjustments are applied only when router consensus is sufficiently confident, mitigating unstable switching caused by short-term noise. By reducing the learned decision from high-dimensional weight prediction to expert-level selection, AlphaRouter guarantees constraint satisfaction by construction and improves robustness to noisy, nonstationary markets. Experiments on DJIA, HSI, CSI, and CRYPTO show that AlphaRouter achieves the best performance among the compared methods in cumulative wealth, annualized percentage yield, annualized Sharpe ratio, and Calmar ratio.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Adaptive Portfolio Selection, Expert Routing, Transformer, Portfolio Optimization

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_EpEBOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This study is an application-oriented investigation of transformer ensembles. AlphaRouter proposes an adaptive portfolio selection framework that separates portfolio construction from market-dependent selection. Instead of directly predicting high-dimensional asset weights, the method constructs four complementary portfolios using predefined investment strategies and employs an ensemble of five independently trained Transformer routers to select suitable experts based on recent market conditions. To improve stability, AlphaRouter combines majority voting and confidence-based selection with a historically best-performing reference expert, allowing limited portfolio adjustments while reducing unnecessary switching.

Critical strengths

The paper addresses a practical and well-motivated problem by reformulating high-dimensional portfolio weight prediction as selection among predefined feasible portfolios. The proposed framework offers a clear and interpretable design that explicitly considers both market adaptation and allocation stability. In addition, the experimental evaluation is relatively comprehensive, covering four financial markets, 19 baseline methods, multiple performance metrics, and ablation studies. The reported results demonstrate consistently competitive performance across the evaluated markets.

Critical weaknesses

This work appears to be more of an application-oriented study that integrates existing techniques than a proposal of a fundamentally new machine learning methodology. Therefore, it may be more appropriate to evaluate the work primarily in terms of the empirical utility of the proposed design rather than methodological novelty. Although the experiments report competitive performance, it remains unclear whether the proposed Transformer ensemble offers substantial advantages over simpler expert-selection rules. The ablation results suggest that Transformer-based routing can provide additional improvements over the stability reference expert alone. However, it is difficult to determine whether these improvements are statistically significant or consistently reproducible. Therefore, further analyses, such as comparisons with simple rule-based routing methods under the same expert set and stability mechanism, would be helpful in establishing the statistical reliability of the performance gains attributable to the proposed routing approach.

I also have concerns regarding the novelty of the proposed approach, even when considering this work as an application-oriented study. The general idea of dynamically selecting among multiple portfolio strategies is not new, as the authors themselves acknowledge in their discussion of existing mixture-of-experts approaches. Although AlphaRouter introduces a stability-aware routing mechanism, its main components, including Transformer-based expert selection, ensemble voting, and confidence-based fallback, are largely established techniques. The paper would benefit from a clearer explanation of how the proposed framework differs conceptually from existing expert-selection methods and why their combination constitutes a sufficiently novel contribution beyond an effective engineering integration.

Questions for authors
  1. Compared with existing approaches such as Mixture-of-Experts (MoE), hierarchical policy selection, and ensemble gating, what specific problem does AlphaRouter address that has not been adequately addressed by prior methods?

  2. Is there a compelling reason to use a Transformer as the backbone of the proposed method? Could a simpler architecture achieve comparable performance?

  3. In Table 2, several baseline methods appear to produce identical or nearly identical numerical results. For example, CRP, RiskBound, and AlphaMix report exactly the same CW (1.354), APY (0.164), and ASR (1.165) on the DJIA dataset. Similar patterns can also be observed across other datasets. Could the authors explain why these different methods yield such similar results, or alternatively, double-check the accuracy of the reported values?

Reviewer_S1fxOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes an adaptive portfolio selection framework that decouples portfolio construction from market-dependent selection called AlphaRouter. It maintains a bank of four hand-crafted portfolio experts (slow sparse, defensive dense, fast sparse, and adaptive dense) that independently generate complete, feasible portfolios under a capped-simplex constraint. An ensemble of five independently trained Transformer routers evaluates expert suitability from causal market states. Their decisions are aggregated via validation-calibrated majority voting, combined with a confidence threshold, and anchored to a historically dominant reference expert using a static tactical share

Critical strengths

1.Framing portfolio selection as routing among pre-constructed, feasible candidate portfolios instead of unconstrained high-dimensional weight regression is reasonable.

  1. The method achieves promising results across diverse asset classes (U.S., Hong Kong, and Chinese equities, as well as cryptocurrencies) and outperforms several traditional econometrical methods and recent deep learning competitors.

  2. The authors provide ablations (leave-one-expert-out, ensemble size, confidence calibration, stability fallback) and a transaction-cost stress test up to 100 bps, showing that the framework might be applicable to real-world trading circumstances.

Critical weaknesses
  1. The core performance engines are four traditional and hand-crafted heuristic strategies (momentum/trend-following and mean-variance optimization with fixed lookbacks and shrinkage). The "Transformer routers" are nearly inactive for most of the test period, switching only 1 to 3 times across the entire 2-year test horizon (Table 11). Hence the complex Transformer router ensemble may not be actually contributing meaningful intelligence. In Table 10, standalone experts (e.g., fast sparse on CSI, adaptive dense on CRYPTO) already match or exceed many external deep learning SOTAs. The promising results may be driven largely by the intensive engineering and tuning of the underlying heuristic expert bank. The neural router is essentially operating as a very conservative switch that rarely triggers.

  2. The authors state in the abstract that “Existing learning-based methods often predict asset weights directly from market data, making portfolio decisions sensitive to noisy and nonstationary observations. We propose AlphaRouter, an adaptive portfolio selection framework that separates portfolio construction from market-dependent selection”. However, little (if any) additional information is exploited by the proposed AlphaRouter other than the market data itself actually. This can also be seen in the experimental part, where the developing of AlphaRouter is entirely based on the market data sets. It is difficult to see how AlphaRouter is essentially discriminative from market-dependent selection.

  3. The main issue lies in the lack of novelty about the methodology. For example, the term zt,iz_{t,i} in Eq. 6 is only a well-known Sharpe ratio form in the literature. The optimization in Eq. 8 is also a well-known mean-variance portfolio selection model with a long history. Worse still, the authors use covariance shrinkage in Eq. 9 without sufficient explanation. Shrinking the covariance means dropping the co-influence among the assets, which should be explained with persuasive reasons and motivations.

Moreover, there are several ad-hoc settings in the core components of AlphaRouter. For example, the performance score in Eq. 11 consists of four subtracting terms with four tunable hyperparameters. It is unclear why they can be mixed into a composite performance score since they have different orders of magnitude: The turnover in Eq. 25 has the same order of magnitude to the l_1 norm of ww, while the concentration in Eq. 26 has the same order of magnitude to the squared l_2 norm of ww. Worse still, it is unclear why the current form in Eq. 26 can quantify the concentration of a portfolio.

In summary, the current form of the proposed methodology is too ad-hoc and heuristic, which does not meet the standard of ICLR.

Questions for authors

see Weaknesses

Reviewer_nJrbOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

AlphaRouter is a mixture-of-experts approach to sequential portfolio selection. Four predefined experts construct feasible portfolios, while an ensemble of Transformer routers selects a market-dependent expert using recent market statistics and expert returns. The proposed allocation blends the routed expert with a reference expert selected from historical training performance, with confidence filtering intended to limit switching. The framework models portfolio-weight drift and proportional transaction costs. Experiments on DJIA, HSI, CSI, and cryptocurrency data report favorable results against classical portfolio strategies and recent learning-based baselines, together with ablations for routing, reference allocation, and confidence calibration.

Critical strengths
  1. The use of predefined feasible expert portfolios makes it easier to interpret the selected allocations. Convex combinations preserve the stated long-only, fully invested, and asset-cap constraints. The manuscript also considers sequential changes in holdings and transaction costs, although the precise accounting convention could be clarified.
  2. The authors evaluate the approach on four markets and report standalone-expert and reference-only comparisons. These controls help distinguish the strength of the expert bank from the incremental contribution of routing, even though the evidence is not yet decisive.
Critical weaknesses
  1. The central framework combines familiar ingredients: predefined portfolio experts, a learned mixture-of-experts selector, ensemble voting, and a stable reference allocation. The additional routing and blending step appears largely to be another layer of optimization or fine-tuning. The paper establishes feasibility of the blended portfolios but offers little new theory explaining why this routing mechanism should yield systematic or robust improvements in investment performance. This limits its significance for ICLR.
  2. The reported improvements over the reference-only strategy are relatively small, and some cumulative-wealth curves appear to obtain much of their separation during particular market episodes. Strong terminal returns alone do not establish persistent value from adaptive routing. Comparisons with validation-selected static expert mixtures and simpler selectors are missing. More importantly, the supplement implements unanimity-based committee consensus and stateful log-probability switching, whereas Eqs. (15)--(17) describe majority voting with top-two probability-margin gating and reference fallback. The code also standardizes expert scores separately at each date, which is not reflected in the training equations. These discrepancies need to be reconciled before the empirical claims can be assessed reliably.
Questions for authors
  1. What is the principal methodological contribution beyond a conventional mixture-of-experts selector combined with a fixed reference allocation? Is there any theoretical or decision-theoretic argument for the routing rule beyond observed backtest performance?
  2. Can the authors compare against validation-selected static mixtures of the same experts and simpler routing methods under identical costs, blending weights, and tuning budgets? It would also help to quantify the incremental gain over the reference-only portfolio and show how much is attributable to individual market episodes.
  3. Which implementation generated the reported experiments? Eqs. (15)--(17) describe majority voting and top-two probability-margin gating, whereas the supplied code uses unanimity and stateful log-probability switching. Please also explain the date-wise standardization of training scores and reconcile the README's description of the test panel as retrospective development evidence with Section 4.1's final-test-only claim.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.