Learning the Efficient Frontier

Summary

The authors propose a novel deep learning method, NeuralEF, to tackle large-scale constraint convex optimization problems by formulating them as sequence-to-sequence (SEQ2SEQ) problems. The key idea is to learn the complex relationship using an attention mechanism between financial conditions and optimal portfolio allocation, which is often a time-consuming task in portfolio management. They demonstrate the efficacy of NeuralEF on the Efficient Frontier problem in finance, a problem known for its non-continuous and highly variable nature. In the experiments, the authors utilize a synthetic dataset with varying conditions. The results illustrate NeuralEF's ability to effectively provide accurate portfolio allocations, all the while significantly reducing the computational resources required for large-scale simulations.

Strengths

The originality and advancement of this paper lie in the proposed NeuralEF approach, which formulates a convex optimization problem as a SEQ2SEQ problem. The paper contributes to the development of the field by applying SEQ2SEQ modeling, originally a tool used in natural language processing, to a completely different domain of portfolio optimization. The concept of learning the mapping from financial conditions to portfolio allocation through deep learning has the potential to be applied to a variety of other optimization problems.

Weaknesses

- This work seems to operate under the assumption that expected returns and volatilities are known with certainty. This might be a limiting factor when considering the application of this model in real-world scenarios, where such accurate estimation is challenging. - The study lacks a rigorous evaluation on real-world financial data. While the synthetic data-based experiments might provide insights into the model's capabilities, the applicability of the model to real-world scenarios remains unclear without empirical validation on real datasets. - It's also worth noting that modern portfolio approaches, such as risk-based portfolios, often rely less on expected returns due to the inherent difficulty in their estimation. Without comparative studies against such methods, it might be challenging to evaluate the practicality and superiority of NeuralEF.

Questions

- Could the authors clarify how the NeuralEF model handles uncertainties in expected returns and volatilities? These are critical aspects in the field of asset management, and their treatment could significantly impact the model's performance and applicability. - Could the authors consider evaluating the model using real-world financial data? Such evaluations could give more insight into how the model would perform in practice and help to assess its real-world applicability. - How would the proposed model compare with other modern portfolio strategies, such as risk-based portfolios, that don't rely heavily on expected returns due to the difficulty in their accurate estimation?

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

2 fair

Presentation

2 fair

Contribution

2 fair

Limitations

- The paper does not address the assumptions made about the certainty of expected returns and volatilities. These assumptions could significantly impact the model's performance and applicability in real-world settings where such values are typically estimated with uncertainty. - The practical utility of the proposed model remains somewhat uncertain without empirical validation using real-world financial data and without comparison with other portfolio approaches like risk-based portfolios, which are often used due to their reduced reliance on expected returns.

Summary

The paper presents a sequence-to-sequence (SEQ2SEQ) deep learning model, called NeuralEF, to solve the Efficient Frontier (EF) problem in portfolio optimization. The authors reformulate the EF problem as a SEQ2SEQ task and train a deep neural network (DNN) to approximate the convex optimizer. The proposed method aims to provide a fast and accurate approximation of the EF while handling variable-length inputs and robustly respecting linear constraints. A Dynamic Greedy Allocation Rebalancing (DGAR) module is introduced to ensure that the model's forecasts remain within the feasible domain and respect the constraints.

Strengths

The paper proposes a novel deep learning-based approach to tackle the EF problem, leveraging the powerful capabilities of SEQ2SEQ models. The DGAR module is an interesting addition that ensures the feasibility of the solutions generated by the DNN.

Weaknesses

Lack of innovation in employing sequence-to-sequence models for convex optimization problems: While the paper presents the NeuralEF method as a novel approach for solving convex optimization problems using sequence-to-sequence models, it is important to acknowledge that existing research has already explored similar ideas. For instance, sequence-to-sequence models have been applied to combinatorial optimization problems, and transformers have been used for embedding[1,2]. To strengthen the paper, the authors should discuss the novelty of their approach in light of these existing works and emphasize the unique contributions of their method to the field. Inadequate demonstration of the effectiveness of the DGAR module: The effectiveness of the Dynamic Greedy Allocation Rebalancing (DGAR) module, which is a key component of the NeuralEF method, is not thoroughly demonstrated or compared with the model's performance without it. This raises questions about the impact of the DGAR module on the model's performance and its ability to handle discontinuous behavior and diverse constraint types. To address this weakness, the authors should provide a more detailed analysis of the effectiveness of the DGAR module, including an ablation study or comparison with the model without DGAR. References: [1] Pomo: Policy optimization with multiple optima for reinforcement learning. Advances in Neural Information Processing Systems 33 (2020) [2] Learning to iteratively solve routing problems with dual-aspect collaborative transformer. Advances in Neural Information Processing Systems 34 (2021)

Questions

1 The formatting template used in this paper is incorrect; it should follow the NeurIPS 2023 guidelines instead. 2 Insufficient justification for the claimed speedup and practical significance: The paper claims a significant speedup of the NeuralEF method compared to the baseline optimizer, but the justification for this speedup is not thoroughly provided. Furthermore, the practical significance of the speedup is not well addressed, considering the potential trade-offs between solution quality and training requirements. The authors claim that their NeuralEF method is 396 times faster than the baseline optimization method. However, if this speedup takes into account the 25x acceleration achieved by using a GPU, as mentioned in line 116, the actual speedup of the NeuralEF method compared to the baseline technique is approximately 15.84 times (396 / 25). It is important to consider whether this speedup is significant enough to justify the use of the proposed method over traditional techniques, given that this nework is already trained using 1 biliion samples. Moreover, there are concerns about the solution quality of the NeuralEF method, which might be worse than the baseline method when calculating the speedup. 3 Lack of comprehensive comparison with other optimization techniques: As shown in Table 3, the accuracy for Asset case 12 is below 84% across all three metrics. Without any comparison to other methods, it is difficult to determine if this level of accuracy is good or not. It also difficult to determine if it is sufficient for practical applications.

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

2 fair

Limitations

Inadequate demonstration of the effectiveness of the DGAR module: The effectiveness of the Dynamic Greedy Allocation Rebalancing (DGAR) module, which is a key component of the NeuralEF method, is not thoroughly demonstrated or compared with the model's performance without it. This raises questions about the impact of the DGAR module on the model's performance and its ability to handle discontinuous behavior and diverse constraint types. To address this weakness, the authors should provide a more detailed analysis of the effectiveness of the DGAR module, including an ablation study or comparison with the model without DGAR. Poor out-of-domain generalization: The NeuralEF method exhibits poor performance when applied to out-of-domain examples, which limits its potential applicability in real-world scenarios.

Summary

This paper considers the problem of solving efficient frontier (EF), a fundamental resource allocation problem where one has to find an optimal portfolio maximizing a reward at a given level of risk. Traditionally, this optimization is solved by quadratic optimization techniques. This paper introduces NeuralEF, a sequence-to-squence formulation, attempting to make fast neural approximations of EF. The purpose is to robustly forecast the result of the EF convex optimization problem with respect to heterogeneous linear constraints and variable number of optimization inputs. By experimental results, the proposed NeuralEF is a viable solution to accelerate large-scale simulation.

Strengths

- The paper is well-written with clear notations and definitions. - In the problem of solving for EF, obtaining robust and smooth results is a challenge given its discontinuous nature. The proposed method is shed light on how data-driven methods can help in smoothing the result in an organic way. - r - The methodology, especially the mathematical programming techniques, including formations in eq. 4-10 and the solver implementation details in Sec 4.2 are clear.

Weaknesses

- Out-of-sample result is not discussed. It would be interesting to check the out-of-sample results in a distributional manner, to highlight the behavior bias of the proposed method in practical cases. - Out-of-domain generalization is briefly discussed in Sec. 4.3. However, EF are dependent on correlation regime and shock, therefore it makes sense to discuss the solution accuracy in different regimes.

Questions

- What behavior bias characteristics does the proposed NeuralEF have? - In Sec. 4.1, what does the solution accuracy look like in different regimes?

Rating

8: Strong Accept: Technically strong paper, with novel ideas, excellent impact on at least one area, or high-to-excellent impact on multiple areas, with excellent evaluation, resources, and reproducibility, and no unaddressed ethical considerations.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

no concern.

Summary

In this paper, the authors propose NeuralEF, a deep neural network (DNN) approach that approximates what is known as the ``efficient frontier'' (EF) in economics. To this end, they use a stacked transformer encoder architecture, as well as pre-processing steps (such as ordering assets) and a greedy algorithm (DGAR) that uses dynamic programming to come up with an asset allocation. The main idea behind NeuralEF is to formulate the EF optimization problem as a sequence-to-sequence (SEQ2SEQ) problem. This allows them to use self attention inside a large language model (LLM) framework. The authors claim that this allows them to understand the relationships between the optimization inputs. They provide experimental results using NeuralEF on a convex optimization problem from generated data.

Strengths

1. The paper is on an interesting topic: applying self-attention and LLMs to economics problems. This is important because many economic problems can not be solved in high dimension with standard tools and analysis from Economics. 2. Despite using a DNN, the authors are conscientious of their carbon footprint and compute resource usage. This makes their proposed approach more environmental as well as accessible to those with limited compute resources. 3. The authors wrote a vectorized implementation to solve multiple EF problems at once. Their proposed methodology is computationally efficient. 4. The authors augment existing DNN approaches by introducing separate modules that perform pre-processing and a dynamic greedy allocation module to respect constraints.

Weaknesses

1. Some claims by the authors seem not to be substantiated. For example, they claim that their approach allows them to understand the relationships between the optimization inputs. Yet, there is little discussion of these relationships in the paper, and it is also unclear how their experimental results shed light in this area. The claim that their approach is robust to discontinuities resulting from large switches in assets appears also unsubstantiated. 2. NeuralEF has many moving parts and thus is difficult to understand and to use. It is unclear what the contribution of each component is. There are some details missing in the explanation of how NeuralEF works (e.g. how the initial allocation is chosen). 3. Beyond being able to leverage recent advances in LLMs, it is unclear why a SEQ2SEQ encoding of the problem makes sense in the context of solving the EF problem. Intuitively, multiple portfolios are equally efficient and there is no natural ordering of the assets in general. 4. The significance of the experimental results are unclear to me because the data generation process does not appear to be explained in the paper.

Questions

1. What is an intuitive explanation of why it makes sense to sequentialize the assets? There is no natural ordering in a portfolio of assets and so, would you need to worry about making your solutions invariant to reordering? 2. What is the contribution of the main components of NeuralEF to the overall performance? For example, how much does the DGAR component affect the accuracy attained and the compute time needed? How much does the preprocessing contribute to the overall performance? 3. How does DGAR scale to larger number of assets? Even with $O(N\log N)$ complexity, this part seems costly with a sizeable portfolio. 4. Why include a class constraint when DGAR seems to just ignore it? 5. Does the result of NeuralEF depend on the initial asset allocation chosen? 6. How is the data generated? What is the function used to generate the data? What kind of noise is added to the data? 7. Why does the training set need to be so big (i.e. a billion examples)? 8. Why is the data generated for the experiments convex? Are there convexity assumptions or requirements for NeuralEF? 9. How does NeuralEF and the results in the paper contribute to understanding the relationship between optimization inputs to the EF problem? 10. The authors claim that discontinuities due to switching between assets doesn't impact the resulting portfolio return. Is this robustness to discontinuities also observed for the resulting volatility of the portfolio?

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

2 fair

Presentation

3 good

Contribution

2 fair

Limitations

Yes