Paper: arXiv 2609.29887
Authors: Yi-Chen Liu, Chung-Han Hsieh
Abstract
This paper investigates cost-sensitive online window size selection for portfolio management under changing market conditions. Specifically, we propose a two-level framework that constructs portfolios using candidate window sizes and dynamically aggregates them through online learning. By treating candidate window sizes as ``experts,’’ we dynamically update their aggregation weights using turnover-inclusive losses. Moreover, we derive finite-horizon cost-sensitive tracking-regret bounds that account for turnover of the aggregated portfolio, with static regret as a special case. Under bounded losses and cost rates, suitably tuned Fixed Share achieves asymptotically no tracking regret for sublinear switching budgets, with Hedge covering the static case.
Complexity vs Empirical Score
- Math Complexity: 8.5/10
- Empirical Rigor: 4.0/10
- Quadrant: Lab Rats — high math complexity, limited empirical testing
Why this score: The paper presents a highly mathematical framework for cost-sensitive online window selection with strong theoretical guarantees. While the theoretical rigor is high, the excerpt provided does not detail any empirical validation or backtesting, which is crucial for quant finance. The novelty lies in integrating turnover costs into regret bounds for dynamic window selection.
Research Flowchart
flowchart TD
A["Research Goal: Cost-Sensitive Online Window Size Selection for Portfolio Management"] --> B{"Key Methodology: Two-Level Framework"};
B -- Level 1: Portfolio Construction --> C["Candidate Window Sizes as 'Experts'"];
B -- Level 2: Dynamic Aggregation --> D["Online Learning: Update Aggregation Weights with Turnover-Inclusive Losses"];
C & D --> E["Computational Process: Derive Finite-Horizon Cost-Sensitive Tracking-Regret Bounds"];
E --> F{"Key Findings: Asymptotically No Tracking Regret (Fixed Share), Static Case Covered (Hedge)"};
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 two‑level framework that builds portfolios from candidate window sizes and dynamically aggregates them via online learning with turnover‑inclusive losses.
Source sentence
Specifically, we propose a two-level framework that constructs portfolios using candidate window sizes and dynamically aggregates them through online learning.
Assumes 5
- The per‑round prediction loss is bounded: a ≤ ℓ(w, y) ≤ b for all (w, y). statistical
Source sentence
Assume further that a ≤ℓ(w, y) ≤b for all (w, y) ∈W × Y, where a < b are fixed finite constants.
- Transaction cost rates are non‑negative at each time step. execution
Source sentence
Let c(t) ≥0 denote the transaction cost rate.
- A switching budget K (0…T‑1) limits the number of expert changes. model
Source sentence
Given a time horizon T ≥1 and a switching budget K ∈{0, . . . , T −1}, our goal is to select q(t) ∈∆n before y(t) is revealed...
- The maximum ℓ1 distance between any two probability‑simplex vectors is 2, used to bound combined loss. statistical
Source sentence
Because the original loss is bounded in [a, b] and the maximum l1 distance between any two vectors on the probability simplex is 2, the combined loss at time t lies in the interval ˜ℓj(t) ∈[a, b+2c(t)].
Show 1 more
- Synthetic data generation assumes returns follow a linear autoregressive model with a regime shift at halfway point. other
Source sentence
the market undergoes a sharp regime shift precisely halfway through the dataset: for the first 1,000 days, the returns are driven by a jt = 63, and jt = 5 for the remaining days.
Data
2021-03 – 2024-02 · DJIA stocks · equities · daily
- Synthetic regime‑shift dataset synthetic unknown
Source sentence
− µ + ǫ(t) ... for the first 1,000 days, the returns are driven by a jt = 63, and jt = 5 for the remaining days.
- DJIA market-prices public
Source sentence
Table 1: Summary of the empirical studies with transaction cost rate c = 0.001. DJIA: 2021/03/04–2024/02/25; ...
- S&P 500 market-prices public
Source sentence
Table 1: Summary of the empirical studies with transaction cost rate c = 0.001.
Costs
Modeled
A flat transaction cost rate c is incorporated into the loss; experiments use c = 0.001.
Source sentence
Define the cost-sensitive loss of expert j at time t as their prediction loss plus their incurred transaction cost: ˜ℓj(t) := ℓ(wj(t), y(t)) + c(t)∥wj(t) −wj(t −1)∥1.
Finds 2
total return 38.1% · max drawdown 12.9%
- Fixed Share achieves a cost‑sensitive tracking‑regret bound that scales with the transaction‑cost rate and switching budget, recovering Hedge’s static‑regret bound as a special case. theoretical
Source sentence
We derive a cost-sensitive tracking-regret bound for Fixed Share, with the static-regret bound for Hedge recovered as a special case.
- In synthetic experiments, Fixed Share attains lower cumulative return than Hedge but still tracks the regime shift with bounded regret. performance
Source sentence
Figure 1 shows the cost-sensitive tracking regret defined in P…
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 15 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 (76)
- On Unified Adaptive Portfolio Management — → 2307.03391
- Literature review: Machine learning techniques applied to financial market prediction (2019)
- On-Line Portfolio Selection with Moving Average Reversion (2012) — link
- Approximation to Bayes Risk in Repeated Play (1957)
- Prediction, Learning, and Games (2006)
- Tests for Parameter Instability and Structural Change with Unknown Change Point (1993)
- Empirical Asset Pricing via Machine Learning (2020)
- On the optimality of the Hedge algorithm in the stochastic regime (2019)
and 68 more
- Portfolio Selection (1952)
- Foundations of Portfolio Theory (1991)
- Convex optimization (2004)
- Online convex optimization and no-regret learning: Algorithms, guarantees and applications (2018) — link
- CVXPY: A Python-Embedded Modeling Language for Convex Optimization (2016)
- An analog of the minimax theorem for vector payoffs. (1956)
- Blackwell Approachability and No-Regret Learning are Equivalent (2011)
- Tracking a Small Set of Experts by Mixing Past Posteriors (2001) — link
- Dynamic Regret Reduces to Kernelized Static Regret (2025) — link
- Learning with kernels: support vector machines, regularization, optimization, and beyond (2002)
- Autoregressive Conditional Heteroscedasticity with Estimates of the Variance of United Kingdom Inflation (1982)
- Online Learning: A Modern Introduction Using Convex Optimization (2026) — link
- Does the Stock Market Overreact? (1985)
- Adaptive sliding window normalization (2025) — link
- Learning from Time-Changing Data with Adaptive Windowing (2007) — link
- Detecting Concept Drift with Support Vector Machines (2000)
- Experience with a learning personal assistant (1994) — link
- Adaptive information filtering: Learning in the presence of concept drifts (1998)
- The Weighted Majority Algorithm (1994) — link
- Big Data: Deep Learning for financial sentiment analysis (2018)
- Machine learning in finance: the case of deep learning for option pricing (2017)
- Advances in financial machine learning (2018)
- Identifying Optimal Window Size Configurations for Big Data Time Series Forecasting (2024) — link
- Determining the impact of window length on time series forecasting using deep learning (2019) — link
- Forecasting price movements using technical indicators: Investigating the impact of varying input window length (2017) — link
- Selection of Market Window Size in Portfolio Strategies (2018) — link
- Machine learning in the Chinese stock market (2022) — link
- Machine learning in finance (2020)
- International financial adjustment (2007)
- Out-of-Sample Forecast Tests Robust to the Choice of Window Size (2012)
- Out-of-Sample Exchange Rate Predictability with Taylor Rule Fundamentals (2009)
- Why has US inflation become harder to forecast? (2007)
- Selection of estimation window in the presence of breaks (2007)
- Forecast Combination Across Estimation Windows (2011) — link
- Optimal forecasts in the presence of structural breaks (2013)
- Learning in the presence of concept drift and hidden contexts (1996)
- Fractal Market Analysis: Applying Chaos Theory to Investment and Economics (1994)
- Rolling window selection for out-of-sample forecasting with time-varying parameters (2017)
- MLP-Based Learnable Window Size for Bitcoin Price Prediction (2022)
- Adaptive sliding window normalization (2025)
- Time-Varying Window Length for Correlation Forecasts (2017)
- Topic: No-Regret and Coarse Correlated Equilibria (2020)
- Risk-Sensitive Online Learning (2006)
- Efficient learning algorithms for changing environments (2009) — link
- Strongly Adaptive Online Learning (2015)
- Online Portfolio Selection: A survey (2014) — link
- Technical Note - A Robust Perspective on Transaction Costs in Portfolio Optimization (2018)
- Trading costs (2018)
- Measuring and modeling execution cost and risk (2006)
- A Closer Look at Adaptive Regret (2016)
- Equilibria under Dynamic Benchmark Consistency in Non-Stationary Multi-Agent Systems (2025) — link
- Tracking the Best Expert (1998)
- A New Interpretation of Information Rate (1956)
- Universal portfolios (1991)
- On-Line Portfolio Selection using Multiplicative Updates (1998)
- Universal Portfolio Selection (1998)
- Stochastic Nonstationary Optimization for Finding Universal Portfolios (2000)
- PAMR: Passive Aggressive Mean Reversion Strategy for Portfolio Selection (2012)
- Confidence weighted mean reversion strategy for online portfolio selection (2013)
- Optimal Portfolio Selection and Dynamic Benchmark Tracking (2005)
- On-line Portfolio Selection Using Stochastic Programming (2003)
- Mirror Descent Meets Fixed Share (and Feels No Regret) (2012)
- Tracking the Best Linear Predictor (2001)
- Relative Loss bounds for Multidimensional Regression Problems (1997)
- The Relaxation Method of Finding the Common Point of Convex Sets and Its Application to the Solution of Problems in Convex Programming (1967)
- On Information and Sufficiency (1951)
- Convex Optimization: Algorithms and Complexity (2015)
- Investment Science (2013)
1 of 76 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.