PaperlayerView exact revision on arXiv

Source-linked AI summary

FactorBench: A Portfolio-Aware Benchmark for Automated Factor Mining

Zhuohan Wang, Carmine Ventre

arXiv:2610.06947v1q-fin.PMcs.LG

TL;DR

It remains unclear whether advances across automated factor-mining paradigms produce more generalizable, distinct, and economically useful signals. FactorBench compares roughly 5,000 factors from nine methods across five equity markets using a shared evaluation contract and portfolio-aware tests. It finds no consistent winner across methods or evaluation stages.

  • Problem

    It remains unclear whether newer factor-mining paradigms yield more generalizable, distinct, and economically useful signals, amid method-specific evaluation pipelines and discovery scores that serve only as intermediate proxies.

    Source: §1 · p. 1

  • Method

    FactorBench evaluates roughly 5,000 factors from nine methods across five equity markets under a shared contract, tracing factor validity and generalization through pool distinctness to after-cost portfolio performance.

    Source: §1 · p. 2, §6 · p. 9

  • Results

    No factor-mining paradigm consistently dominates: discovery performance can deteriorate out of sample, larger pools need not contain broader information, and long-only advantages do not consistently carry to long–short portfolios.

    Source: §6 · pp. 9–10, §5 · p. 9

  • Takeaways & Limitations

    Progress in factor mining is multidimensional and cannot be established by a search objective or isolated predictive score alone.

    Source: §6 · p. 10

    Conclusions apply to evaluated implementations under this shared task: five large-cap markets, one daily 20-day target, one principal chronological split, and three random seeds per automated method.

    Source: §5 · p. 9

Abstract

from arXiv · show

Factor mining seeks to discover signals from financial data that predict future asset returns and guide portfolio construction. Automated factor mining now spans genetic programming, reinforcement learning, generative models, and large language model agents. Yet it remains unclear whether advances across these paradigms yield more generalizable, distinct, and economically useful financial signals. We introduce FactorBench, a portfolio-aware benchmark comparing roughly five thousand mined factors from nine automated mining methods across five equity markets. A shared data and evaluation contract supports both symbolic expressions and executable Python factors, connecting heterogeneous discovery algorithms to common signal combination and portfolio construction procedures. FactorBench traces the outputs of mining systems across three levels: factor validity, temporal generalization, and predictiveness beyond measured risk and style exposures; within- and across-method pool distinctness, including similarity to the benchmark Alpha101; and composite-signal quality and after-cost long-only and long--short portfolio performance. After systematically assessing whether advances in factor mining translate into signal quality and portfolio performance, FactorBench finds that no paradigm consistently dominates.

1 INTRODUCTION

Factor mining faces weak, non-stationary signals and inconsistent evaluation across methods, leaving unclear whether newer search approaches yield practically useful discoveries. FactorBench compares mining systems across markets and evaluation stages, finding no consistent downstream progression.

1 INTRODUCTION

  • Financial signals are weak and non-stationary, and repeated searches over historical observations can find patterns that do not persist.Method-specific data, objectives, and backtesting pipelines also make published results difficult to compare; discovery scores are only intermediate proxies for downstream value.Source: §1 · p. 1
  • FactorBench evaluates roughly 5,000 factors from 9 methods spanning genetic programming, reinforcement learning, generative search, and LLM agents across 5 equity markets.Its shared data and evaluation contract accommodates symbolic expressions and executable code while preserving each method’s native discovery process.Source: §1 · p. 2
  • FactorBench evaluates progress through generated factors, delivered pools, and the portfolios those signals support, connecting predictiveness to distinctness and investment value.Discovery performance is treated as an intermediate result rather than the endpoint.Source: §1 · p. 2
  • Across markets and evaluation stages, advances in search do not consistently produce more generalizable, distinct, or economically useful signals than earlier approaches.Relative advantages change from discovery to out-of-sample evaluation, from individual factors to pools, and from signals to portfolios.Source: §1 · p. 2

2 RELATED WORK

Automated factor discovery has progressed from symbolic search toward generative systems and LLM-driven research workflows. Existing benchmarks assess individual capabilities or generated factors, while FactorBench evaluates the path from heterogeneous discovery methods to downstream investment decisions.

2 RELATED WORK

  • Genetic programming explores symbolic expressions through evolutionary operations, while later methods use sequential decision-making, collection-level objectives, or grammar-guided search.AlphaGen optimizes a new factor’s marginal contribution to an existing collection (Yu et al., 2023), AlphaQCM uses distributional reinforcement learning, and AlphaCFG combines an alpha-oriented grammar with Monte Carlo tree search (Yang et al., 2026).Source: §2 · p. 2
  • Generative methods emphasize diverse candidate collections, while LLM agents extend discovery to hypothesis generation, executable implementation, feedback, and candidate revision.Examples include AlphaForge (Shi et al., 2025), AlphaSAGE (Chen et al., 2026), R&D-Agent (Li et al., 2026), AlphaAgent (Tang et al., 2025), and QuantaAlpha (Han et al., 2026).Source: §2 · p. 2
  • Prior factor benchmarks assess LLM capabilities, generated-factor properties, or executable strategies, whereas FactorBench evaluates whether discovery advances translate into distinct signals and downstream portfolio value.Financial-language benchmarks assess capabilities such as document understanding and forecasting, while investment platforms standardize data, market environments, and backtesting.Source: §2 · pp. 2–3

3 FACTORBENCH: BENCHMARK DESIGN

FactorBench makes heterogeneous mining systems comparable while retaining their native discovery processes. It uses a shared market and temporal setting, maps symbolic and executable outputs into common signals, and evaluates factors, pools, and portfolios.

3.1 BENCHMARK OVERVIEW

  • The benchmark’s four-block workflow connects a shared setting and heterogeneous mining systems to common signal execution and portfolio-aware evaluation.Methods retain their native algorithms and objectives, while resource use and delivered output are recorded.Source: §3.1 · pp. 3–4

3.2 BENCHMARK SETTING AND FACTOR PRODUCTION

  • All methods use the same five equity-market panels and predict subsequent 20-trading-day returns under a chronological 2017–2021 training, 2022–2023 validation, and 2024–2025 test split.The test period is reserved for final trading evaluation.Source: §3.2 · p. 4
  • The comparison covers nine methods across genetic programming, reinforcement-learning and guided symbolic search, generative search, and LLM agents, plus an implementable Alpha101 reference.Eight methods produce symbolic expressions and R&D-Agent produces executable Python code; both representations pass through the shared execution contract.Source: §3.2 · p. 4

3.3 EVALUATION PROTOCOLS

  • Factor evaluation checks executability and scorable dates, then measures period-specific IC with orientation fixed from training and predictiveness after adjustment for five measured exposures.A factor–market instance needs at least 100 scorable dates in each split; residual IC measures predictiveness after removing market beta, volatility, momentum, short-term reversal, and liquidity.Source: §3.3 · p. 4
  • Pool distinctness measures within-method redundancy, across-method signal-space convergence, and similarity to Alpha101 using cross-sectional signal correlations, primarily Spearman.The appendix also examines temporal stability and robustness to using Pearson correlation.Source: §3.3 · pp. 4–5
  • Portfolio evaluation selects valid, less-redundant factors using pre-test evidence, combines them by three weighting approaches, and tests long-only and dollar-neutral long–short portfolios after costs.Composite evaluation before portfolio construction separates information retained through selection and combination from performance after portfolio constraints, turnover, and trading costs.Source: §3.3 · p. 5

4 EXPERIMENTAL EVALUATION

FactorBench evaluates mined factors from individual validity and temporal generalization through pool distinctness to composite-signal and portfolio outcomes. Across these levels, method rankings do not align consistently.

4.1 FACTOR-LEVEL EVALUATION

  • 4,838 of 4,914 candidate factors execute successfully (98.5%), and 4,698 (95.6%) meet the common scorability requirement across all three splits.The nine methods completed three-seed runs in each market; delivered factor counts ranged from 107 to 1,619.Source: §4.1 · p. 5, Table 1 · p. 6
  • GP, RL, and generative methods often lead training IC, but much of that advantage disappears in validation and generally does not recover in testing.LLM methods start with lower training IC, so their smaller train-to-test decline does not mean higher held-out predictiveness; repeated training-objective optimization may increase adaptive overfitting.Source: §4.1 · pp. 5–6, Figure 2 · p. 6
  • Most method–market points fall below the raw-versus-neutralized IC parity line, with reductions especially consistent for GP, AlphaForge, and AlphaGen.GP shows the largest overall decline, and R&D-Agent’s neutralized IC becomes negative in some markets.Source: §4.1 · p. 6, Figure 3 · p. 7
  • Exposure overlap rankings remain stable across training, validation, and testing, with GP and R&D-Agent consistently showing the highest overlap.The figure measures overlap through adjusted R2 from regressing factors on exposures; the paper attributes GP’s overlap to rediscovery from shared price-and-volume inputs and suggests pretrained financial knowledge may shape R&D-Agent’s overlap.Source: §4.1 · p. 6, Figure 4 · p. 7

4.2 POOL-LEVEL EVALUATION

  • AlphaGen and AlphaQCM show the greatest within-pool redundancy, while GP pools are also highly correlated in several markets.Internal diversity varies by market; AlphaSAGE is highly redundant in CSI 300 but substantially more diverse elsewhere.Source: §4.2 · p. 7, Figure 5 · p. 8
  • Non-LLM methods tend to resemble one another, the three LLM systems resemble each other, and LLM-generated pools are typically closer to Alpha101 than non-LLM pools.The strongest between-method overlap is between AlphaGen and AlphaQCM; greater internal diversity does not necessarily mean greater distinctness from familiar factor families.Source: §4.2 · p. 7, Figure 5 · p. 8

4.3 PORTFOLIO-LEVEL EVALUATION

  • Long-only portfolios all earn positive net returns of 22.4%–31.2% and Sharpe ratios of 1.03–1.31, while long–short returns are smaller and R&D-Agent and AlphaQCM lead long–short Sharpe.R&D-Agent’s composite IC falls from 0.0253 to 0.0011 after exposure neutralization, whereas AlphaQCM retains most of its raw IC.Source: Table 2 · p. 8, §4.3 · pp. 8–9

5 DISCUSSION AND LIMITATIONS

Across factor, pool, and portfolio evaluations, no factor-mining paradigm consistently dominates within FactorBench. The authors limit that conclusion to the evaluated implementations and benchmark setting.

5 DISCUSSION AND LIMITATIONS

  • At the factor level, stronger discovery-period performance does not consistently persist through temporal testing or adjustment for measured exposures.Source: §5 · p. 9
  • At the pool and portfolio levels, neither distinctness nor method rankings improve monotonically across paradigms, and no paradigm dominates across the full evaluation chain.Source: §5 · p. 9
  • The conclusions cover evaluated implementations in five large-cap equity markets, one daily 20-day target, one principal chronological split, and three seeds per automated method—not broader paradigms’ fundamental capabilities.Portfolio tests also use static universes and simplified cost and borrowing assumptions without fully modeling survivorship, capacity, market impact, or borrow availability.Source: §5 · p. 9

6 CONCLUSIONS

FactorBench evaluates mined factors through shared data and execution contracts, tracing outputs from factor quality to portfolio performance. Across the evaluated methods, results do not identify a clear winner, and search performance alone does not establish progress.

6 CONCLUSIONS

  • FactorBench evaluates approximately 5,000 factors from nine automated mining methods across five equity markets while preserving each method’s native discovery process.Its shared data, target, temporal-split, and execution contract supports symbolic and executable factors.Source: §6 · p. 9
  • The three-level evaluation follows factors from validity and temporal generalization through exposure-neutralized predictiveness and pool distinctness to after-cost long-only and long–short performance.The findings frame factor-mining progress as multidimensional rather than established by a search objective or isolated predictive score alone.Source: §6 · pp. 9–10
  • The results do not identify a clear winner among genetic programming, reinforcement-learning, generative, and LLM-based approaches; stronger discovery-period performance often deteriorates out of sample.Larger pools do not necessarily contain broader information, and much composite predictiveness diminishes after measured exposures are removed; positive long-only outcomes do not consistently translate to long–short advantage.Source: §6 · pp. 9–10

A DATA AND PROCESSING PROTOCOL

FactorBench applies a common daily-data protocol to five fixed large-cap equity universes, using shared input fields and a 20-trading-day forward-return target. Fixed constituents provide a consistent universe but do not reconstruct historical membership and may introduce survivorship bias.

A DATA AND PROCESSING PROTOCOL

  • The five market panels use fixed universes totaling 782 securities, including a fixed 272-stock CSI 300 subset.The markets are the S&P 100, Hang Seng Index, CSI 300, FTSE 100, and Nikkei 225.Source: A DATA AND PROCESSING PROTOCOL · p. 12
  • Mining methods receive adjusted daily open, high, low, close, and volume, plus a VWAP proxy and 20-day average daily volume.The VWAP proxy is (high + low + close)/3; average daily volume requires at least five observations.Source: A DATA AND PROCESSING PROTOCOL · p. 13
  • The common data layer applies no additional winsorization, cross-sectional normalization, or general missing-value imputation.When required, these transformations occur within the mining method or evaluation procedure.Source: A DATA AND PROCESSING PROTOCOL · p. 14
  • The prediction target is each stock’s 20-trading-day forward close-to-close return, measured on the market’s trading calendar and withheld from factor inputs.The close is adjusted, and the target is stored separately from the input fields.Source: A DATA AND PROCESSING PROTOCOL · p. 14
  • Observations span January 2016 through December 2025, with fixed constituent sets throughout; the design may introduce survivorship bias because it does not reconstruct historical index membership.Evaluation runs from January 2017 through October 2025, using 2016 as warm-up history.Source: A DATA AND PROCESSING PROTOCOL · p. 14

B.1 COMMON SETUP AND IMPLEMENTATION SUMMARY

The benchmark evaluates methods under shared market panels, target, and chronological splits while retaining native discovery procedures. It records implementation, resource, and output details, but runtime is not a controlled measure of computational efficiency.

B.1 COMMON SETUP AND IMPLEMENTATION SUMMARY

  • All nine methods use the same market panels, input fields, 20-day target, and chronological split, with three repetitions per market and 135 runs overall.Training covers 2017–2021, validation 2022–2023, and testing January 2024–October 2025; LLM runs are nondeterministic.Source: §B.1 · p. 14
  • Methods retain native search algorithms, objectives, expression spaces, and pool construction, while test outcomes are excluded from factor acceptance and downstream construction.Training and validation data may support search, parameter estimation, feedback, or model selection as required by each method.Source: §B.1 · p. 14
  • Released implementations are recorded with fixed upstream commits and run-level provenance, while search budgets preserve source specifications where feasible rather than equalizing compute.The budgets are not intended to equalize candidate evaluations, generated factors, or compute units.Source: §B.1 · p. 14
  • The three LLM methods use the same Qwen3.8-27B endpoint, so differences arise from their workflows, prompts, feedback, and search budgets rather than the underlying model.R&D-Agent generates Python factors, whereas AlphaAgent and QuantaAlpha generate expressions in their respective domain-specific languages.Source: §B.1 · p. 14
  • Adapters serialize symbolic expressions and Python factors as date-by-asset signals, and failed or unscorable outputs remain in the validity analysis.Scorable outputs require sufficient observations to compute IC on at least 100 dates in every temporal split.Source: §B.1 · pp. 14–15
  • Recorded wall-clock time describes practical resource footprint in this setting, not a controlled computational-efficiency comparison.Methods use different software stacks, hardware resources, and external API calls.Source: §B.1 · p. 15

B.2 METHOD-SPECIFIC IMPLEMENTATIONS

FactorBench connects heterogeneous mining methods to shared data and execution procedures while retaining their native search processes. Its validity analysis counts all delivered outputs but evaluates predictive performance only on factors that meet a common scorability threshold.

B.2 METHOD-SPECIFIC IMPLEMENTATIONS

  • The GP baseline searches AlphaGen expressions, while AlphaGen uses a PPO policy and AlphaCFG uses grammar-constrained Monte Carlo tree search.GP outputs are checked for parseability and cross-sectional dispersion; AlphaGen retains its native target and training-based pool updates, while AlphaCFG retains its grammar, tree search, and pool construction.Source: §B.2 · pp. 15–16
  • AlphaQCM, AlphaForge, and AlphaSAGE use distinct search designs, with FactorBench adapting their data interfaces and exporting their factor pools for common evaluation.AlphaForge's target is changed to the shared 20-day return and its loop is capped at 50 rounds; AlphaSAGE's native target already matches, and AlphaQCM's validation results guide trajectory selection while test data remain unavailable to search.Source: §B.2 · pp. 16–17
  • The LLM methods retain their released prompts and feedback workflows, but FactorBench runners supply shared-panel evaluation and keep test data unavailable during search.R&D-Agent generates executable programs and receives training and validation IC and RankIC; AlphaAgent and QuantaAlpha receive validation feedback for their respective discovery or trajectory-selection workflows.Source: §B.2 · p. 17
  • A shared adapter handles market panels and output recording while methods retain native search procedures; Alpha101 is evaluated as a fixed reference, and Alpha Jungle and CogAlpha are excluded from reported experiments.The latter methods lack official source code, and their paper-scale configurations require orders of magnitude more LLM calls than can be reproduced under the common endpoint; reduced versions would confound quality with search budget.Source: §B.2 · pp. 15, 17, 18

C FACTOR-LEVEL EVALUATION

  • The paired temporal analysis uses the same factors in each split, requiring finite ICs on at least 100 dates in training, validation, and testing.Predictive performance is evaluated only for scorable factors; all delivered factors remain in validity reporting, and factor signs are fixed on training data for validation and test.Source: §C.1 · pp. 18–19, C FACTOR-LEVEL EVALUATION · p. 18, Figure 6 · p. 19

C.1 FACTOR VALIDITY

  • 4,914 factors were delivered; 4,838 (98.5%) executed successfully and 4,698 (95.6%) met the common scorability requirement.Among Alpha101's 71 unique applicable formulas, all executed successfully and 70 were scorable across all five markets.Source: §C.1 · p. 18
  • All 76 execution failures came from LLM-based pipelines, while numerical degeneracy was a principal attrition source for several symbolic methods.AlphaCFG, AlphaForge, and AlphaSAGE had 11, 25, and 47 IC-incomplete factors, respectively; QuantaAlpha had 33 execution failures and 27 executable factors without an IC in every split.Source: §C.1 · p. 18

C.2 FACTOR GENERALIZATION

Factor predictiveness often weakens outside the discovery period, and measured style exposures account for part of it. These patterns vary by method and market, with no consistent held-out advantage across methods.

C.2 FACTOR GENERALIZATION

  • GP, reinforcement-learning, and generative methods often have higher training RankIC than validation or test RankIC, while no method leads consistently across all five markets.LLM-based methods generally start with lower training RankIC, so their smaller absolute changes do not amount to systematically higher held-out RankIC.Source: §C.2 · p. 19
  • ICIR separation between training and later periods remains visible for many methods, but several retain positive test ICIR in individual cases.ICIR is descriptive here: overlapping 20-day return targets make daily IC observations serially dependent, and interquartile ranges describe factor dispersion rather than statistical uncertainty.Source: §C.2 · p. 19, Figure 7 · p. 20
  • Positive train-to-test degradation is widespread across IC, RankIC, and ICIR, especially for GP, reinforcement-learning, and generative methods in several markets.Exceptions in the S&P 100 and Nikkei 225 show that degradation is not universal at the method–market level.Source: §C.2 · pp. 19–20, Figure 8 · p. 20

C.3 FACTOR NEUTRALIZATION BY STYLES

  • Neutralization reduces validation predictiveness for most method–market points, though attenuation varies and occasionally IC increases.In several markets, raw validation IC is already near zero or negative, limiting evidence of profitable predictiveness before or after neutralization.Source: Figure 9 · p. 21, §C.3 · p. 21
  • The five price- and volume-derived controls are not an exhaustive risk model, so positive neutralized IC indicates predictiveness beyond this control set, not exposure-free alpha.A positive ΔIC indicates a decline in predictiveness after removing measured exposures; it is not a percentage of alpha explained.Source: §C.3 · p. 21
  • Style-overlap effects persist in held-out testing, but their magnitude differs across markets; GP and R&D-Agent show comparatively large IC reductions in several markets.The analysis controls for market beta, volatility, momentum, short-term reversal, and liquidity, while leaving the return target unchanged and using matched observations.Source: §C.3 · pp. 20–21, Figure 10 · p. 22

D POOL-LEVEL EVALUATION

Factor pools differ in redundancy within and across methods, and these patterns persist across temporal splits and correlation definitions. Similarity does not itself establish shared provenance or independent economic value.

D POOL-LEVEL EVALUATION

  • Earlier formula-based methods tend to resemble one another, especially AlphaGen and AlphaQCM; LLM-method similarities are generally smaller and most pronounced within the LLM group.R&D-Agent and AlphaAgent nevertheless have greater similarity to Alpha101 than most formula-based methods in several markets, and these relationships change little across temporal splits.Source: D POOL-LEVEL EVALUATION · p. 22
  • Pearson correlation preserves the qualitative redundancy and between-method patterns, while similarities are lower than Spearman for several pools, particularly AlphaGen, AlphaQCM, and AlphaSAGE.Spearman captures similarity in stock rankings; Pearson additionally requires similarity in relative signal magnitudes.Source: D POOL-LEVEL EVALUATION · p. 22, Figure 14 · p. 24
  • The similarity measure averages factor-pair cross-sectional correlations over dates and reports absolute values, with summaries within pools, between methods, and against Alpha101.Within-method and Alpha101 comparisons are reported by market; between-method entries aggregate eligible run-pair comparisons across markets.Source: D POOL-LEVEL EVALUATION · p. 22, Figure 11 · p. 23
  • AlphaGen and AlphaQCM show consistently high within-method similarity over validation and test, while GP pools are also relatively redundant in several markets.AlphaForge, AlphaCFG, and most LLM-based pools are generally more internally distinct, though the pattern depends on market.Source: D POOL-LEVEL EVALUATION · p. 22, Figure 12 · p. 23
  • Pool distinctness does not improve monotonically across mining paradigms: some methods deliver narrow signal spaces, while others produce smaller, more heterogeneous pools.Stable patterns across temporal splits and correlation measures suggest that the structures reflect mining-system search spaces and inductive biases rather than only temporary training relationships.Source: D POOL-LEVEL EVALUATION · pp. 22–23
  • Signal correlation measures similarity in realized outputs, not common provenance; high similarity does not establish reproduction, and low similarity does not establish independent predictive or economic value.Similar rankings can imply similar trades even when raw signal magnitudes are not linearly proportional.Source: D POOL-LEVEL EVALUATION · p. 23

E PORTFOLIO-LEVEL EVALUATION

Portfolio results vary substantially across markets and methods, and no method establishes a stable global lead. Style-neutralized long–short results further distinguish residual predictiveness from returns associated with measured exposures.

E PORTFOLIO-LEVEL EVALUATION

  • Selection ranks eligible factors by validation IC after fixing direction on training data, then excludes candidates correlated at 0.8 or higher with any previously selected factor.The procedure uses no test-period predictive metric and imposes no positive-validation-IC requirement or pool-size limit.Source: E PORTFOLIO-LEVEL EVALUATION · pp. 24–25
  • Under Pearson correlation, within-method redundancy remains visible, and several LLM-based methods show comparatively greater overlap with Alpha101.This pattern is reported under an alternative correlation definition.Source: Figure 15 · p. 25

E.1 DEFAULT PERFORMANCE AND VARIATION

  • Long-only Sharpe ratios occupy a narrow range, while cross-market variation often exceeds differences between method means; positive CAGR alone does not establish alpha for fully invested portfolios.The reported comparison requires the equal-weight universe to distinguish stock-selection value from common market returns.Source: §E.1 · p. 25, Table 5 · p. 25
  • Cross-market dispersion prevents a reliable global ranking: long–short Sharpe is positive for every method in the S&P 100, while several methods turn negative in other markets.The Nikkei 225 favors several formula-based methods, so pooled averages combine materially different outcomes.Source: §E.1 · p. 26, Table 7 · p. 26

E.2 WEALTH DYNAMICS AND EXPOSURE DEPENDENCE

  • Long-only wealth paths reflect underlying market returns, whereas long–short gains concentrate in the S&P 100 and Nikkei 225, with losses in several other markets.The figure separates gross performance, after-cost performance, and after-cost performance from style-neutralized composites.Source: §E.2 · pp. 26–27, Figure 16 · p. 27
  • After style neutralization, AlphaQCM alone remains profitable among automated long–short methods: net CAGR 3.7% to 1.3%, Sharpe 0.668 to 0.283, and turnover 4.58 to 6.02.All other automated methods have negative style-neutralized long–short CAGR and Sharpe; AlphaQCM also has the lowest style-neutralized turnover among automated methods.Source: §E.2 · p. 27

E.3 SENSITIVITY TO PORTFOLIO CONSTRUCTION

Portfolio outcomes depend on trading assumptions and signal-combination choices. Long–short and style-neutralized results are especially fragile, and no single combination rule dominates across methods.

E.3 SENSITIVITY TO PORTFOLIO CONSTRUCTION

  • Long-only CAGR declines steadily as costs rise but remains positive, while most long–short methods approach or fall below zero.Longer rebalance intervals generally improve net performance by reducing turnover and aligning holding periods with the 20-day prediction target.Source: §E.3 · pp. 28–29, Figure 18 · p. 29
  • Long–short outcomes are more sensitive than long-only outcomes to costs, rebalancing frequency, and signal combination.The sensitivity figures vary one design choice at a time; these tests do not select the best test-period configuration.Source: §E.3 · p. 27, Figure 19 · p. 29
  • Most methods produce negative net CAGR for style-neutralized long–short portfolios under default combination and cost assumptions; slower rebalancing mitigates losses but does not yield a common positive result.Long-only portfolios retain positive absolute returns because they remain fully invested, though their advantage relative to raw composites narrows.Source: Figure 21 · p. 29, §E.3 · p. 30
  • No signal-combination rule dominates across methods: rankings change, and some long–short outcomes change sign, between equal weighting, IC weighting, and ridge estimation.The style-neutralized sensitivity panels report market-aggregated net CAGR for long-only and long–short portfolios.Source: §E.3 · p. 29, Figure 20 · p. 29, Figure 21 · p. 29
Loading 2610.06947v1…