PaperlayerView exact revision on arXiv

Source-linked AI summary

Online Portfolio Selection: A Survey

Bin Li, Steven C. H. Hoi

arXiv:1212.2129v2q-fin.CPcs.AIcs.CEq-fin.PM

TL;DR

This survey frames online portfolio selection as a sequential decision problem and organizes existing strategies into five major groups. It also connects these algorithms with Capital Growth theory and identifies open challenges for future research.

  • Problem

    Online portfolio selection is an interdisciplinary computational-finance problem whose existing algorithms and research challenges require structural understanding.

  • Method

    The article surveys algorithmic strategies, categorizes them into five groups, and relates them to Capital Growth theory.

  • Results

    The survey presents Benchmarks, Follow-the-Winner, Follow-the-Loser, Pattern-Matching, and Meta-Learning approaches, then outlines open challenges.

  • Takeaways & Limitations

    The survey provides a framework for understanding existing online portfolio-selection methods and their underlying trading ideas.

  • Takeaways & Limitations

    The article focuses on algorithmic aspects and structural understanding rather than theoretical details.

Abstract

from arXiv · show

Online portfolio selection is a fundamental problem in computational finance, which has been extensively studied across several research communities, including finance, statistics, artificial intelligence, machine learning, and data mining, etc. This article aims to provide a comprehensive survey and a structural understanding of published online portfolio selection techniques. From an online machine learning perspective, we first formulate online portfolio selection as a sequential decision problem, and then survey a variety of state-of-the-art approaches, which are grouped into several major categories, including benchmarks, "Follow-the-Winner" approaches, "Follow-the-Loser" approaches, "Pattern-Matching" based approaches, and "Meta-Learning Algorithms". In addition to the problem formulation and related algorithms, we also discuss the relationship of these algorithms with the Capital Growth theory in order to better understand the similarities and differences of their underlying trading ideas. This article aims to provide a timely and comprehensive survey for both machine learning and data mining researchers in academia and quantitative portfolio managers in the financial industry to help them understand the state-of-the-art and facilitate their research and practical applications. We also discuss some open issues and evaluate some emerging new trends for future research directions.

1. INTRODUCTION

Online portfolio selection is framed as a sequential wealth-allocation problem suited to multi-period Capital Growth Theory. The survey organizes existing methods, connects them to their trading ideas, and defines a scope centered on algorithmic approaches rather than detailed theory.

  • Motivation: Capital Growth Theory provides the theoretical perspective most closely fitted to the online, multiple-period setting.It targets portfolio growth rate, or expected log return, rather than a single-period mean-risk trade-off.
  • Motivation: Online portfolio selection allocates wealth sequentially across assets to maximize cumulative wealth or expected log return.The task differs from single-period Mean Variance portfolio selection because decisions and market information arrive over multiple periods.
  • Survey organization: The survey classifies state-of-the-art algorithms into Benchmarks, Follow-the-Winner, Follow-the-Loser, Pattern-Matching, and Meta-Learning approaches.Its classification is presented as a structural overview of online portfolio selection techniques.
  • Scope: The survey emphasizes algorithmic aspects and structural understanding while skimming theoretical details and excluding most single-period Mean Variance studies.It also distinguishes online portfolio selection from financial-engineering work focused on forecasting time series or trading individual stocks.
  • Survey organization: The article addresses a literature gap by surveying categories that earlier surveys covered only partially or no longer represented the state of the art.Earlier work emphasized competitive analysis or Pattern-Matching approaches without comprehensively covering the broader set of categories.

2. PROBLEM SETTING

The problem setting models online portfolio selection as repeated allocation across assets using sequential price relatives and portfolio rebalancing. Wealth compounds through period returns, while an extended formulation accounts for proportional transaction costs under idealized market assumptions.

  • Online portfolio model: Each price-relative vector records asset price changes, and the portfolio earns the inner-product period return b_t · x_t.The ith price relative is the ratio of the current closing price to the previous closing price for asset i.
  • Online portfolio model: Cumulative wealth grows multiplicatively across periods because each period’s return is reinvested.The strategy is scored by final cumulative wealth after the sequence of trading periods.
  • Online portfolio model: At each period, a portfolio manager chooses a portfolio vector from the simplex using previously observed market information.Portfolio entries are nonnegative and sum to one, so the model excludes margin and short selling.
  • Online portfolio model: The online framework repeatedly computes a portfolio, reveals the market price relative, records the period return, and updates the selection rule.This procedure supports backtesting different online portfolio selection algorithms.
  • Assumptions: The basic model assumes no transaction costs, unlimited liquidity at closing prices, and no market impact from portfolio strategies.These assumptions simplify the market interaction represented in the baseline formulation.
  • Transaction costs: The survey models transaction costs but does not introduce strategies that directly solve transaction-cost issues.This marks a boundary between the formulation’s evaluation model and the survey’s algorithm coverage.
  • Transaction costs: Under proportional transaction costs, rebalancing reduces wealth through a transaction-cost factor determined by the old portfolio, price relatives, and new portfolio.Buying and selling use separate rates, and the factor lies between 1−γ_s/(1+γ_b) and 1.

3. ONLINE PORTFOLIO SELECTION APPROACHES

Online portfolio selection algorithms use explicit portfolio-update schemes within a sequential learning framework. The survey groups them by trading or weight-transfer direction and includes benchmark methods alongside several strategy families.

  • General framework: Online portfolio selection algorithms formulate the task as sequential learning and derive an explicit portfolio update for each period.Their updates implicitly assume price-relative predictions and optimize the resulting portfolio.
  • General framework: The survey first presents benchmark algorithms, then classifies later methods by the direction in which portfolio weights are transferred.This organization follows the categories summarized in its classification table.
  • Follow-the-Winner: Follow-the-Winner methods increase relative weights for more successful experts or stocks based on historical performance.The category is defined by transferring weight toward prior winners.

3.1. Benchmarks

The survey presents Buy-And-Hold, Best Stock, and Constant Rebalanced Portfolio strategies as benchmarks for online portfolio selection. It contrasts their wealth accumulation, highlighting CRP’s volatility-pumping behavior, universality as a target, and transaction-cost limitation.

  • Buy-And-Hold Strategy: Buy-And-Hold invests using an initial portfolio and does not rebalance, allowing holdings to change with market fluctuations.Its final wealth is the initial-portfolio-weighted average of individual assets’ final wealth.
  • Best Stock Strategy: The Best Stock strategy is a hindsight Buy-And-Hold strategy that invests all capital in the best-performing stock.
  • Constant Rebalanced Portfolio: Constant Rebalanced Portfolio rebalances to a fixed portfolio every period, with BCRP choosing the optimal fixed portfolio offline from the complete market sequence.Uniform CRP uses equal portfolio weights, while BCRP is a hindsight strategy.
  • Constant Rebalanced Portfolio: On a synthetic market with no long-run asset growth, BAH wealth remains roughly 1 while CRP wealth is roughly 9^2 and increases exponentially.CRP achieves this by rebalancing every period and exploiting underlying volatility, known as volatility pumping.
  • Constant Rebalanced Portfolio: Frequent CRP transactions can incur high transaction costs, motivating Semi-CRP strategies that rebalance only on selected periods.
  • Universality: Universal algorithms asymptotically approach BCRP’s exponential growth rate for arbitrary sequences of price relatives.The survey identifies universality as a desired theoretical property for online portfolio selection.

3.2. Follow-the-Winner Approaches

Follow-the-Winner methods increase weights for successful experts or stocks, with Universal Portfolios aggregating constant-rebalanced strategies and related methods pursuing favorable regret, efficiency, or model-adaptive guarantees.

  • Follow-the-Winner Approaches: Follow-the-Winner methods increase relative weights of successful experts or stocks and often target the best constant-rebalanced portfolio (BCRP).
  • 3.2.1. Universal Portfolios: Universal Portfolio assigns capital across parameterized constant-rebalanced portfolios over the valid portfolio simplex, then pools their wealth.This construction is analogous to a Fund of Funds and uses historical manager performance as weights.
  • 3.2.1. Universal Portfolios: Cover’s Universal Portfolio asymptotically approaches BCRP’s exponential growth rate under suitable smoothness conditions, with regret O(m log n).
  • 3.2.1. Universal Portfolios: Weighted Universal Portfolios retain the same regret scale as Cover’s method while achieving a better constant term.
  • 3.2.1. Universal Portfolios: Follow-the-Winner research extends Universal Portfolios through side information, transaction-cost models, alternative expert classes, and connections to prediction, compression, and mean-variance theory.
  • Follow-the-Winner Approaches: Other surveyed methods improve computational efficiency or adapt to market structure, including linear-time updates, GBM-linked regret bounds, and regime-switching strategies.Some approaches report O(m) running time, regret O(m log(Q + m)), or empirical outperformance on Cover’s two-stock pairs; regime-switching methods face distributional concerns.
  • 3.2.5. Aggregating-type Algorithms: Regime-switching portfolio methods may better reflect non-i.i.d. markets, but their distributional assumptions can be poorly matched to observed market behavior.

3.3. Follow-the-Loser Approaches

Follow-the-Loser strategies transfer wealth toward poorer-performing assets by exploiting mean reversion. The section covers Anticor, PAMR, CWMR, OLMAR, and RMR, emphasizing their empirical performance alongside theoretical and practical limitations.

  • Mean reversion: Follow-the-Loser strategies transfer weights from better-performing stocks to poorer-performing stocks under a mean reversion assumption.CRP embodies this principle by rebalancing toward its initial allocation after relative performance differences emerge.
  • Anticor: Anticor uses logarithmic price relatives from two market windows and cross-correlation statistics to determine mean-reversion transfers.Transfers reflect positive cross-correlation and negative autocorrelation, then are normalized to remain in the simplex domain.
  • Anticor: Anticor empirically outperformed contemporaneous strategies but lacked a useful universal regret bound because its mean-reversion design was heuristic.The survey identifies systematic learning of mean reversion as a desired direction beyond Anticor’s heuristic approach.
  • PAMR: PAMR defines an ε-insensitive loss that remains zero below a threshold and aggressively updates the portfolio when the threshold is exceeded.Its closed-form update transfers wealth from good-performing stocks toward poorer-performing ones, followed by simplex projection.
  • PAMR: PAMR achieved significant performance gains, showed parameter robustness, and ran with linear update time in back tests.These properties supported its reported practicability for large-scale real-world applications, although meaningful regret bounds remained difficult.
  • OLMAR and RMR: OLMAR replaces single-period price prediction with multiple-period moving-average reversion, addressing cases where PAMR and CWMR failed.It uses simple moving averages and passive-aggressive online learning; the survey reports the best results among existing algorithms at that time.
  • OLMAR and RMR: RMR uses an L1-median estimator for expected price relatives and applies an OLMAR-like portfolio optimization procedure.The survey reports that RMR outperformed the state of the art on most datasets.

3.4. Pattern-Matching based Approaches

Pattern-Matching approaches identify historically similar market situations and optimize a portfolio from the associated price relatives. They combine nonparametric sample-selection methods with utility-based portfolio optimization, often incorporating aggregation.

  • Overview: Pattern-Matching approaches use historical similarities to predict future price relatives and seek universal consistency for stationary ergodic markets.They provide an alternative to approaches centered exclusively on winners or losers.
  • Overview: Their pipeline first selects a similarity set C of historical price relatives, then optimizes a portfolio using those samples and assigned probabilities.The two stages are combined into specific online algorithms, with an additional aggregation step.
  • Sample selection: Similarity is generally measured by comparing preceding market windows represented as concatenated vectors.The procedure iterates over historical price relatives and retains those whose preceding windows satisfy the selected similarity criterion.
  • Sample selection: Sample-selection variants use Euclidean-distance thresholds, nearest neighbors, correlation coefficients, or discretized histogram partitions.Threshold parameters control the number of selected samples or the similarity criterion in several variants.
  • Portfolio optimization: Portfolio optimization may use log utility, semi-log-optimal utility, Markowitz-type mean-variance utility, or transaction-cost-adjusted utility.If the similarity set is empty, the algorithms can use a uniform or previous portfolio.
  • Portfolio optimization: Semi-log-optimal strategies approximate logarithmic utility to reduce computational demands, while retaining a theoretically analyzed universal-consistency framework.The survey attributes the universal-consistency proof to Vajda’s analysis.

3.5. Meta-Learning Algorithms

Meta-Learning Algorithms combine multiple portfolio experts and are closely related to expert learning. They broaden the strategy space by aggregating experts from one or more strategy classes and adapting their weights over time.

  • Framework: Meta-Learning Algorithms combine several base experts, either from the same strategy class or from different classes.The framework is directly applicable to a fund of funds, where assets are delegated to other funds.
  • Framework: Meta-learning can combine algorithms with and without theoretical guarantees, potentially extending properties such as universality to the whole system.The survey presents this as a way to support a broader application area.
  • Algorithms: The Aggregating Algorithm updates expert weights using a learning rate, a prior distribution, and each expert’s loss and action.Its formulation accommodates a measurable set of experts and period-by-period weight updates.
  • Algorithms: Fast Universalization evenly distributes wealth among base experts, lets them operate independently, and pools their wealth.It extends Universal Portfolios from parameterized constant-rebalanced portfolios to a wider class of investment strategies.
  • Algorithms: Discrete buy-and-hold combinations aggregate experts such as Anticor and PAMR across finite sets of window sizes or mean-reversion thresholds.These combinations provide finite-parameter alternatives to universalization in continuous parameter spaces.
  • Algorithms: OGU and ONU are meta-optimization extensions of EG and ONS that can achieve the growth rate of the optimal convex combination of underlying experts.The survey notes that a universal base expert can transfer this property to the combined system.
  • Algorithms: FLH handles changing environments by dynamically adding and dropping finite experts according to performance while weighting active experts with a meta-learning algorithm.The framework can incorporate universal base experts such as ONS.

4. CONNECTION WITH CAPITAL GROWTH THEORY

The survey connects online portfolio algorithms with Capital Growth Theory by interpreting their implicit predictions and portfolio objectives. It distinguishes universal-portfolio connections from approaches that explicitly optimize expected log return under predicted scenarios.

  • Capital Growth Theory: Capital Growth Theory maximizes expected log return through prediction and portfolio optimization.Predictions assign possible price-relative combinations probabilities, after which the portfolio maximizing expected log return is selected through convex optimization.
  • Universal Portfolio Connection: Universal portfolio methods connect to Capital Growth Theory by targeting exponential growth relative to the hindsight BCRP strategy.BCRP is optimal in an i.i.d. market, making it a special case of the Kelly criterion’s target.
  • Universal Portfolio Connection: Universal Portfolios, Exponential Gradient, Follow the Leader, and Follow the Regularized Leader have daily-average regret that asymptotically approaches zero.The survey states that these methods can achieve BCRP’s exponential growth rate, which is CGT-optimal in an i.i.d. market.
  • Explicit Capital Growth Connection: A second connection rewrites algorithms in Capital Growth form using predicted price-relative scenarios and their probabilities.Some algorithms, including EG and ONS, fit both CGT connections, combining a formulation in Kelly’s form with motivations based on universal portfolio selection.
  • Explicit Capital Growth Connection: Single-scenario methods such as EG, PAMR, CWMR, OLMAR, and RMR optimize portfolios under an implicit or explicit certain prediction.PAMR, CWMR, and OLMAR omit log utility because adding it preserves the basic idea but creates a convexity issue.
  • Explicit Capital Growth Connection: Pattern-Matching approaches assign uniform probability to similar historical scenarios and maximize expected log return over the resulting similarity set.This produces an optimal fixed-fraction portfolio, although BS, BM, and BGV use different portfolio-optimization procedures.
  • Trading Principles: The survey also classifies trading ideas as momentum, mean reversion, or other prediction principles.EG, FTL, and FTRL are identified with momentum-style assumptions about subsequent winners and losers.

5. CHALLENGES AND FUTURE DIRECTIONS

The survey identifies prediction and portfolio optimization as the two central stages where online portfolio selection still faces open challenges. Future work should improve price-relative prediction, incorporate broader information and market effects, and address risk and market realism.

  • Accurate Prediction: Existing prediction assumptions can perform well but remain imperfect, motivating more subtle models of future price-relative distributions.The survey frames prediction accuracy as a central challenge in the prediction stage.
  • Accurate Prediction: Pattern recognition remains difficult because similar samples are hard to identify, probabilities are usually uniform, and existing windows share fixed lengths and intervals.The survey highlights varying-timing pattern discovery as an additional direction.
  • Accurate Prediction: Autocorrelations in returns, including positive and negative dependencies across assets and horizons, remain underused for predicting future price relatives.The survey separately notes that absolute and squared returns exhibit stronger autocorrelations and nonlinear relationships than simple returns.
  • Accurate Prediction: Calendar effects and side information such as volume, fundamentals, and expert opinions are identified as underexploited sources for prediction.The survey notes that existing algorithms largely focus on price relatives or prices alone.
  • Portfolio Optimization: Capital Growth Theory can maximize cumulative wealth but may incur high risk, creating a need for effective online return-risk trade-offs.The survey notes that simply adding variance may not efficiently balance return and risk.
  • Portfolio Optimization: Optimal f may reduce the risk of Kelly’s approach but requires difficult drawdown estimation before incorporation into online portfolio selection.The survey presents this as an open challenge rather than an established solution.
  • Practical Constraints: Future models should consider short selling, margin, transaction costs, and market liquidity beyond current long-only simplex-constrained formulations.The survey states that liquidity has not been included in algorithm formulations and that practical transaction-cost extensions remain difficult to convey algorithmically.

6. CONCLUSIONS

The article surveys online portfolio selection as a sequential decision problem, organizes existing methods into five categories, and connects their trading ideas to Capital Growth Theory. It concludes that many research problems remain open.

  • Contributions: The survey formulates online portfolio selection as sequential decision learning and categorizes algorithms into five major groups.The groups are Benchmarks, Follow-the-Winner, Follow-the-Loser, Pattern-Matching based approaches, and Meta-Learning algorithms.
  • Contributions: It connects the surveyed algorithms to Capital Growth Theory to clarify the essence of their underlying trading ideas.The survey also outlines open challenges for further investigation.
  • Open Problems: Many open research problems remain unsolved despite the number of proposed online portfolio selection algorithms.The authors intend the survey to support understanding of the state of the art and inspire further study.
Loading 1212.2129v2…