Paper: arXiv 2406.00655

Authors: Andrzej Cichocki, Sergio Cruces, Auxiliadora Sarmiento, Toshihisa Tanaka

Abstract

This paper introduces a novel family of generalized exponentiated gradient (EG) updates derived from an Alpha-Beta divergence regularization function. Collectively referred to as EGAB, the proposed updates belong to the category of multiplicative gradient algorithms for positive data and demonstrate considerable flexibility by controlling iteration behavior and performance through three hyperparameters: \(α\), \(β\), and the learning rate \(η\). To enforce a unit \(l_1\) norm constraint for nonnegative weight vectors within generalized EGAB algorithms, we develop two slightly distinct approaches. One method exploits scale-invariant loss functions, while the other relies on gradient projections onto the feasible domain. As an illustration of their applicability, we evaluate the proposed updates in addressing the online portfolio selection problem (OLPS) using gradient-based methods. Here, they not only offer a unified perspective on the search directions of various OLPS algorithms (including the standard exponentiated gradient and diverse mean-reversion strategies), but also facilitate smooth interpolation and extension of these updates due to the flexibility in hyperparameter selection. Simulation results confirm that the adaptability of these generalized gradient updates can effectively enhance the performance for some portfolios, particularly in scenarios involving transaction costs.

Complexity vs Empirical Score

  • Math Complexity: 8.5/10
  • Empirical Rigor: 6.5/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper introduces a novel generalized family of exponentiated gradient algorithms derived from Alpha-Beta divergence, involving advanced mathematical formulations and derivations. It also demonstrates empirical validation through extensive simulation experiments in online portfolio selection, including transaction costs and hyperparameter optimization, making it both mathematically dense and backtest-ready.

Research Flowchart

  flowchart TD
  A["Research Goal<br>Develop flexible EG updates via<br>Alpha-Beta divergence for OLPS"] --> B["Derive EGAB Update<br>Family from Alpha-Beta<br>Divergence Regularization"]
  B --> C{"Enforce Unit l1 Norm?"}
  C -->|"Method 1"| D["Scale-Invariant<br>Loss Functions"]
  C -->|"Method 2"| E["Gradient Projection<br>onto Feasible Domain"]
  D --> F["Apply to Online Portfolio<br>Selection with Gradient Methods"]
  E --> F
  F --> G["Simulation on<br>Portfolio Data"]
  G --> H["Key Findings<br>Unified view of OLPS algorithms<br>Hyperparameters enable interpolation<br>Improved performance with transaction costs"]

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper proposes a novel family of generalized exponentiated gradient (EG) updates (EGAB) based on an Alpha‑Beta divergence regularizer for online portfolio selection.

Source sentence
This paper introduces a novel family of generalized exponentiated gradient (EG) updates derived from an Alpha-Beta divergence regularization function.

Assumes 5

  • Cumulative return at the end of each period is fully reinvested and no additional wealth can be added. execution
    Source sentence
    We assume that the cumulative return obtained at the end of the t-th period (e.g., one day) will be completely reinvested at the beginning of the t+1-th period and no additional wealth can be taken into the portfolio.
  • In the absence of transaction costs, wealth evolves as the product of portfolio returns. execution
    Source sentence
    In the absence of transaction costs, the final cumulative wealth is given by the expression
  • Initial portfolio is uniformly allocated across all assets. model
    Source sentence
    At the beginning, we assume that the portfolio is uniformly allocated, i.e., w_0 = u = 1/N.
  • During training, the next period’s price‑relative vector is estimated by the current observed vector when no preprocessing is used. data
    Source sentence
    During training and in absence of a relative‑prices preprocessing, we estimate
Show 1 more
  • Weight vectors are non‑negative and sum to one. model
    Source sentence
    The portfolio, which reflects the investment decision in the t-th period, is denoted by a weight vector w_t

Data

  • NYSE-N · OLPS toolbox market-prices public
    Source sentence
    The first four datasets (NYSE-N, NYSE-O, MSCI, and TSE) are commonly utilized in OLPS literature for evaluation, as documented in references [32]- [?], and are available in the OLPS toolbox [37].
  • NYSE-O · OLPS toolbox market-prices public
    Source sentence
    The first four datasets (NYSE-N, NYSE-O, MSCI, and TSE) are commonly utilized in OLPS literature for evaluation, as documented in references [32]- [?], and are available in the OLPS toolbox [37].
  • MSCI · OLPS toolbox market-prices public
    Source sentence
    The first four datasets (NYSE-N, NYSE-O, MSCI, and TSE) are commonly utilized in OLPS literature for evaluation, as documented in references [32]- [?], and are available in the OLPS toolbox [37].
  • TSE · OLPS toolbox market-prices public
    Source sentence
    The first four datasets (NYSE-N, NYSE-O, MSCI, and TSE) are commonly utilized in OLPS literature for evaluation, as documented in references [32]- [?], and are available in the OLPS toolbox [37].
  • DOW30 market-prices public
    2005-01 to 2021-12 · 30 largest US industrial companies
    Source sentence
    Another dataset under consideration is the historical dataset of the Dow Jones Industrial Average (DOW30), which tracks the top 30 biggest US companies, primarily in the industrial sector. This dataset spans from the beginning of January 2005 to the end of December 2021, encompassing a total of T = 4280 trading days.
  • SP-IT (information technology sector of S&P 500) market-prices public
    Information technology stocks in S&P 500
    Source sentence
    Additionally, we enriched the dataset collection by incorporating specific sectors of the Standard & Poor’s 500 index. These sectors include: information technology (SP-IT), consumer staples (SP-CS), communication services (SP-CSe), financials (SP-F), industrials (SP-I), real estate (SP-RE), energy (SP-E), healthcare (

Costs

Modeled Costs are part of the reported headline results.

Commission rates of 0%, 0.025% and 0.25% are incorporated; the EGAB criterion explicitly accounts for transaction costs.

Source sentence
Within this application, we propose an improved criterion that accounts for the existence of transaction costs, and show the effectiveness of learning algorithmic hyperparameters and data preprocessing options through a validation set with a past sample fraction.

Finds 4

  • The generalized EGAB updates improve portfolio performance, especially when transaction costs are present. performance
    Source sentence
    Simulation results confirm that the adaptability of these generalized gradient updates can effectively enhance the performance for some portfolios, particularly in scenarios involving transaction costs.
  • EGAB-P achieves the best performance due to simplex projection enhancing sparsity. performance
    Source sentence
    However, the best performance is obtained for EGAB-P, because the simplex projection of the updated weights enhances the sparsity in comparison with the scalar normalization used by EGAB-N.
  • EGAB algorithms show greater resilience to increasing commission rates compared to other OLPS methods. performance
    Source sentence
    the EGAB algorithms exhibit a greater resilience with respect to the increase of commissions, mainly, because their optimized criteria already have them into account.
  • Turnover of EGAB variants decreases as transaction costs increase. performance
    Source sentence
    One can observe how the EG`, EGAB-N, and EGAB-P algorithms progressively lower their turnover as the transaction costs increase for higher commission rates.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 22 related papers). Coverage is partial, so this is neither replication nor refutation.

Code

No repository by the authors found in the paper.

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.

Cited References (45)

  1. What are the Numbers that Experiments Provide (1994)
  2. On‐Line Portfolio Selection Using Multiplicative Updates (1998) — link
  3. An Analysis of Transformations (1964) — link
  4. Log-Euclidean metric learning on symmetric positive definite Manifold with application to image set classification (2015)
  5. Extended SMART Algorithms for Non-negative Matrix Factorization (2006) — link
  6. Exponentiated Gradient Exploration for Active Learning (2016) — link
  7. Tracking the Best Expert (1998) — link
  8. Moving average reversion strategy for on-line portfolio selection (2015) — link
and 37 more
  1. Additive versus exponentiated gradient updates for linear prediction (1995) — link
  2. Online portfolio selection (2014) — link
  3. Generalized Alpha-Beta Divergences and Their Application to Robust Nonnegative Matrix Factorization (2011) — link
  4. Exponentiated Gradient versus Gradient Descent for Linear Predictors (1997) — link
  5. PAMR: Passive aggressive mean reversion strategy for portfolio selection (2012) — link
  6. Explicit update vs implicit update (2008) — link
  7. Robust and efficient estimation by minimising a density power divergence (1998) — link
  8. Convergence of exponentiated gradient algorithms (2001) — link
  9. Universal portfolios with side information (1996) — link
  10. Robust Blind Source Separation by Beta Divergence (2002) — link
  11. Families of Alpha- Beta- and Gamma- Divergences: Flexible and Robust Measures of Similarities (2010) — link
  12. OLPS: a toolbox for on-line portfolio selection (2016)
  13. Robust Median Reversion Strategy for Online Portfolio Selection (2016) — link
  14. Nonnegative Matrix and Tensor Factorizations (2009) — link
  15. Information Geometry and Its Applications (2016) — link
  16. Alpha-Beta Divergence For Variational Inference (2018) — link
  17. Convergence of the Exponentiated Gradient Method with Armijo Line Search (2018) — link
  18. Exponentiated Gradient Meets Gradient Descent (2019) — link
  19. Winnowing with Gradient Descent (2020)
  20. Reparameterizing Mirror Descent as Gradient Descent (2020) — link
  21. Aggregating exponential gradient expert advice for online portfolio selection (2020) — link
  22. Exponentiated Gradient Reweighting for Robust Training Under Label Noise and Beyond (2021) — link
  23. Online learning: A comprehensive survey (2021) — link
  24. Exponential Gradient with Momentum for Online Portfolio Selection (2021) — link
  25. Adaptive online portfolio strategy based on exponential gradient updates (2021) — link
  26. Stochastic Mirror Descent: Convergence Analysis and Adaptive Variants via the Mirror Stochastic Polyak Stepsize (2021) — link
  27. α β -GAN: Robust generative adversarial networks (2022) — link
  28. Step-size Adaptation Using Exponentiated Gradient Updates (2022) — link
  29. Universal Portfolios (1991) — link
  30. A Portfolio Selection Strategy Based on the Peak Price Involving Randomness (2023) — link
  31. Competitive Online Strategy Based on Improved Exponential Gradient Expert and Aggregating Method (2023) — link
  32. Inverse Reinforcement Learning with Constraint Recovery (2023) — link
  33. OEC: an online ensemble classifier for mining data streams with noisy labels (2023) — link
  34. A quantum online portfolio optimization algorithm (2024) — link
  35. Passive Aggressive Ensemble for Online Portfolio Selection (2024) — link
  36. Brain-like learning with exponentiated gradients (2024) — link
  37. Boosting with Tempered Exponential Measures (2023) — link

0 of 45 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.