PaperlayerView exact revision on arXiv

Source-linked AI summary

AlphaLedger: Portfolio-Grounded On-Policy Distillation for Regime-Aware Alpha Mining

Xiang Li, Zikai Wei, Yiyan Qi, Wanyun Zhou, Xiang Liu, Penglei Sun, Zhenheng Tang, Yongqi Zhang, Jian Guo, Xiaowen Chu

arXiv:2610.23744v1cs.CE

TL;DR

Alpha mining needs supervision aligned with a factor’s incremental after-cost portfolio value, although that value is assessed only after sequential generation is complete; deployment also faces changing stock evidence and factor effectiveness. AlphaLedger distills portfolio-grounded teacher guidance into a local generator and adapts factor use with matured, state-matched evidence. In the common CSI500 universe, its annualized return is 31.6% versus 22.0% for the strongest strategy baseline.

  • Problem

    A factor’s incremental value depends on its contribution to an existing ensemble after trading costs, but is measured only after generation, while stock evidence and factor effectiveness can change.

    Source: §1 · p. 2

  • Method

    EC-OPD distills a frozen teacher’s legal-action distributions on student trajectories using completed-factor portfolio evidence, while RRG uses matured state-matched returns with shrinkage and bounded deviations from equal weighting.

    Source: §1 · p. 2, §3.1 · p. 3, §3.5 · p. 5

  • Results

    31.6% annualized return for AlphaLedger versus 22.0% for GPT-5, the strongest strategy baseline by annualized return, in the common CSI500 universe.

    Source: §4.3 · p. 6

  • Takeaways & Limitations

    The reported experiments show competitive performance in EviPool and CSI500, with component ablations and repeated-generation evaluation.

    Source: §5 · p. 9

    The student is trained once with a frozen teacher and the screened factor library remains fixed during testing; five seeds assess sampling variability from one checkpoint, not training-run uncertainty or significance across market periods.

    Source: Limitations · p. 9

Abstract

from arXiv · show

Hosted LLMs enable automated alpha mining but incur recurring API costs and expose research context to external providers. Training a local generator requires supervision aligned with portfolio utility: a predictive factor may add little after redundancy and trading costs. This incremental value is assessed only after the factor program is complete, whereas generation proceeds sequentially. Deployment must also adapt to evolving stock-specific evidence and changes in factor effectiveness across market regimes. We propose AlphaLedger to address these challenges. Economic-context on-policy distillation (EC-OPD) conditions a frozen teacher on completed factors and their cost-adjusted marginal portfolio contributions, then distills its legal-action distributions along the student's own trajectories. The deployed generator requires neither teacher calls nor outcome inputs. EviPool constructs a dynamic stock universe from lagged multi-source evidence, while a regime-residual gate (RRG) reweights factors using matured, state-matched returns with shrinkage and bounded deviations from equal weighting. On Chinese A-shares, AlphaLedger achieves a 55.4\% annualized net return, a Sharpe ratio of 2.06, and a maximum drawdown of 10.8\%. In a common CSI500 universe, its annualized net return exceeds the strongest baseline by 43.6\% (31.6\% vs.\ 22.0\%), with a 13.5\% higher Sharpe ratio.

1 Introduction

AlphaLedger trains a local factor generator using portfolio evidence and adapts deployment to stock evidence and market regimes. On Chinese A-shares, it reports a 55.4% annualized net return, while its CSI500 return exceeds the strongest baseline.

1 Introduction

  • Local generation needs portfolio-aware supervision because a predictive factor may duplicate existing signals or raise trading costs, while its incremental value is measured only after program completion.Generation proceeds through intermediate choices, which a terminal scalar reward does not directly use to express action preferences.Source: §1 · p. 2
  • EC-OPD conditions a frozen teacher on completed student-generated factors and their cost-adjusted marginal portfolio contributions, then distills legal next-action distributions along student trajectories.The student does not receive portfolio outcomes as inputs.Source: §1 · p. 2
  • EviPool uses lagged multi-source evidence to construct a dynamic candidate universe, while RRG weights factors using matured returns from matching market states, shrinkage, and bounded deviations from equal weighting.The evaluation compares seven baseline families under a shared trading protocol and includes component ablations and repeated generations.Source: §1 · p. 2
  • 55.4% annualized net return, Sharpe ratio 2.06, and maximum drawdown 10.8% on Chinese A-shares; in CSI500, annualized return is 31.6% versus GPT-5’s 22.0%.The CSI500 Sharpe ratios are 1.43 for AlphaLedger and 1.26 for GPT-5.Source: §1 · p. 2

2 Related Work

Prior alpha-mining work spans symbolic search, LLM research agents, and post-training, while investment-universe studies commonly use index membership or data coverage. The paper frames portfolio-informed action supervision and evidence-based stock eligibility as gaps in these lines of work.

2.1 Automated Alpha Mining

  • Existing alpha-mining approaches range from symbolic search and LLM research agents to local generator post-training, but scalar rewards do not provide action-level supervision informed by marginal net portfolio contribution.The passage also distinguishes prompt-based refinement, which keeps experience outside model parameters.Source: §2.1 · p. 2
  • Figure 1 contrasts API-based alpha mining with AlphaLedger’s local generator trained from portfolio evidence and deployment using an evidence-driven universe and adaptive weights.Source: Figure 1 · p. 3

2.2 Investment Universes

  • Prior studies select candidates mainly by index membership or data coverage, leaving the role of newly available stock-specific evidence in universe construction largely unexamined.The cited settings range across Chinese and U.S. indices, broader A-share stocks, and six U.S. stocks.Source: §2.2 · p. 3

3 Method

AlphaLedger separates stock eligibility, factor generation, and factor weighting, using point-in-time information to adapt each decision. Its pipeline combines evidence-based universe selection, portfolio-grounded student training, screening, and regime-aware weighting.

3 Method

  • AlphaLedger separates stock eligibility, factor generation, and factor weighting, with each decision using only information available at decision time.The account ranks stocks within the selected universe using a weighted factor library.Source: §3 · p. 3, §3.1 · p. 3

3.1 Problem Formulation

  • A student generator writes each factor as a sequence of grammar actions, with training labels matured by the cutoff and deployment limited to information available by t−1.The prefix contains the preceding actions, and the legal next-action set constrains generation.Source: §3.1 · p. 3

3.2 EviPool: Evidence-Driven Universe

  • EviPool uses lagged evidence to select the stock universe, while factor scores rank stocks within it; its retention band limits membership churn.At scheduled refreshes, the pool retains eligible members within a wider rank threshold and fills vacancies with high-scoring eligible stocks.Source: §3.2 · pp. 3–4

3.3 EC-OPD: Economic-Context On-Policy Distillation

  • EC-OPD conditions a frozen teacher on each completed student factor and its economic ledger, then distills legal-action distributions on the student’s own prefixes.The student minimizes reverse KL divergence; the outcome ledger is teacher-only, and training records must mature before targets are computed.Source: §3.3 · pp. 4–5, Table 1 · p. 5

3.4 Library Regeneration and Screening

  • The student regenerates programs without outcome feedback, then screening adjusts their return series for six style streams and selects passing candidates by mean-to-standard-deviation ratio.Both signs are tested, and candidates must have positive means in both chronological halves; the selected sign orients the deployed factor.Source: §3.4 · p. 5

3.5 RRG: Regime-Residual Gate

  • RRG weights factors using matured returns from matching market states, shrinks sparse estimates toward overall means, and bounds deviations from equal weighting.Insufficient records or zero dispersion produce equal weights; each factor retains a minimum allocation.Source: §3.5 · p. 5

4 Experiments

Experiments test AlphaLedger on Chinese A-shares using a fixed backtest protocol and compare its results across shared-universe baselines, robustness checks, ablations, and deployment case studies. The full framework reports a 55.4% annualized net return and 2.06 Sharpe ratio; in the common CSI500 universe, it outperforms GPT-5 by annualized return and Sharpe ratio.

4.1 Dataset and Metrics

  • The test uses Chinese A-share price, volume, money-flow, and issuer-linked news data, with development labels ending by December 31, 2024 and backtesting from January 2025 to July 2026.Strategies use the same long-only account, rebalance every five trading days toward 40 equal-weighted stocks, and pay 5 bps to buy and 15 bps to sell.Source: §4.1 · p. 6, §4.2 · p. 6

4.4 Robustness

  • Five libraries sampled from the same student checkpoint have mean terminal NAV 1.906, and even the lowest-performing seed reaches 1.81.Every seed exceeds ChiNext and STAR50 in Sharpe ratio, while maximum drawdown varies from 10.34% to 17.30%.Source: §4.4 · p. 6

4.5 Ablations

  • EviPool reaches NAV 1.951 and Sharpe ratio 2.06, versus NAV 1.251–1.740 and Sharpe ratio 1.03–1.56 across listed index universes.Relative to CSI500, cumulative net return is 43.5 percentage points higher and maximum drawdown falls from 12.3% to 10.8%.Source: §4.5 · p. 6, Table 3 · p. 7
  • Removing any component lowers both return and risk-adjusted performance; excluding EviPool produces the largest ARR decline, while excluding EC-OPD produces the largest Sharpe-ratio decline.The declines are 55.4% to 31.6% in ARR without EviPool and 2.06 to 1.36 in Sharpe ratio without EC-OPD.Source: §4.5 · p. 6, Table 4 · p. 9

4.3 Main Results

  • 31.6% annualized return and 1.43 Sharpe ratio versus GPT-5’s 22.0% and 1.26 in the common CSI500 universe; AlphaLedger’s EviPool configuration reports 55.4% and 2.06.AlphaLedger also has the highest RIC (5.29%) and RICIR (0.43) among CSI500 methods; Table 2 covers the test period and distinguishes the EviPool full model from shared-CSI500 strategies.Source: §4.3 · p. 6, Table 2 · p. 7

4.6 Case Studies

  • The June 2026 sell-off sees the large-order-flow tilt rise to 8.6 points above equal weighting before new-state evidence accumulates, then recede to 0.6 points as matured records favor other families.Figure 4 summarizes mean intent-family weight shifts by market state and tracks the four largest tilts over time.Source: §4.6 · pp. 7–8, Figure 4 · p. 8
  • The style-controlled portfolio reaches NAV 1.388 and Sharpe ratio 1.30 versus NAV 1.865 for the coverage-matched reference, showing sensitivity to exposure constraints.This is a sensitivity analysis, not a causal decomposition; the projection changes stock selection and weights, and the descriptors omit industry and fundamental risks.Source: §4.6 · p. 8

5 Conclusion

AlphaLedger combines portfolio-grounded training for a local factor generator with evidence-driven deployment. Experiments report competitive performance, component ablations, and repeated-generation evaluation.

5 Conclusion

  • AlphaLedger trains a local factor generator with portfolio-grounded distillation and deploys it using an evidence-driven universe and regime-aware weights.EC-OPD guides student-generated trajectories; EviPool and RRG adapt eligibility and allocation using lagged evidence.Source: §5 · p. 9
  • Experiments report competitive performance in EviPool and CSI500, alongside component ablations and repeated-generation evaluation.Source: §5 · p. 9
  • Future work will explore generator updates as evidence matures, cross-market transfer, and additional multimodal sources.Source: §5 · p. 9

Limitations

The evaluation has fixed components and limited replication, unequal candidate budgets across methods, and a narrow market and trading scope.

Limitations

  • The student is trained once with a frozen teacher, and the screened factor library stays fixed during testing; only the universe and weights adapt.Source: Limitations · p. 9
  • Five generation seeds measure sampling variability from one checkpoint, not training-run uncertainty or statistical significance across market periods.Source: Limitations · p. 9
  • Hosted-LLM baselines use 60 candidates versus 440 student generation slots, so results compare reported configurations rather than equal-compute performance.Pretraining overlap with the historical test period cannot be ruled out for models released after it began.Source: Limitations · p. 9
  • Experiments cover Chinese A-shares under a long-only protocol; transfer to other markets and long–short trading remain untested.The style analysis uses five reconstructed technical descriptors, so unexplained return is not a complete risk-adjusted alpha estimate.Source: Limitations · p. 9

Ethics Statement

The system is presented as a research-signal tool, not investment advice; its backtests do not guarantee future returns.

Ethics Statement

A Evaluation Metrics

The evaluation defines account- and factor-level metrics and describes baseline construction. It reports AlphaLedger leading hosted-LLM baselines across four universes.

A Evaluation Metrics

  • The account metrics include terminal NAV, stress NAV at doubled trading costs, annualized return, Sharpe ratio, and maximum drawdown, for which lower is better.Source: A Evaluation Metrics · p. 11
  • Factor evaluation uses clipped cross-sectional Pearson IC and Spearman rank IC, summarized by absolute mean and ICIR across rebalance dates.Style exposures are holdings’ cross-sectional z-scores relative to CSI1000 constituents on the same date.Source: A Evaluation Metrics · p. 11

B Baseline Details

  • Baseline libraries are screened by 20-day RIC and generally combine up to 40 factors with equal weights; AlphaGen retains its learned weights.Machine-learning and deep time-series baselines are trained on CSI500 data through December 2024; all are frozen before testing.Source: B Baseline Details · p. 11
  • AlphaLedger leads hosted-LLM baselines in all four universes, by 9.6 points of ARR in CSI500.In CSI500, AlphaLedger earns 31.6% annually with a Sharpe ratio of 1.43, compared with 22.0% and 1.26 for GPT-5, the strongest baseline.Source: B Baseline Details · p. 11
  • Hosted API-Direct models receive grammar, fields, and abstract intent, returning 60 candidates versus 440 student slots without market data, outcomes, or economic feedback.Source: B Baseline Details · p. 11

C Implementation Details

AlphaLedger fixes its implementation settings before evaluation and defines factor programs through a typed grammar over five data families. Eleven intents guide elicitation without prescribing formulas or portfolio weights; selected factors can share signal behavior across intents.

C Implementation Details

  • All reported settings were fixed before evaluation; final experiments used two NVIDIA H800 GPUs with 80 GB of memory each.Table 7 lists AlphaLedger and machine-learning-baseline hyperparameters and settings.Source: C Implementation Details · p. 11

D Factor Grammar, Intents, and Example Factors

  • Factor programs combine fields from five data families with time-series and cross-sectional operators, arithmetic, and a conditional operator.Programs use windows of 3, 5, 10, 20, or 60 days and are evaluated on data through the previous close.Source: D Factor Grammar, Intents, and Example… · p. 11
  • Eleven generation intents describe desired signals and may constrain data families or horizons, but do not prescribe formulas, signs, or portfolio weights.Slots are interleaved so each intent receives the same number of proposals; Figure 4 treats intents as elicitation contexts, not validated mechanisms.Source: D Factor Grammar, Intents, and Example… · p. 11
  • Tables 5 and 6 provide additional baseline comparisons, with Table 6 covering four universes under shared trading rules and costs.Table 5 includes machine-learning, deep time-series, and API-Direct baselines; Table 6 transfers API libraries screened on pre-2025 CSI500 data unchanged.Source: Table 6 · p. 12, Table 5 · p. 12
  • One selected factor per intent illustrates varied constructions, while recurring fields and operators mean different intent labels need not produce different signal behavior.Each selected factor reads at least two data families; development-set orientation is applied before deployment.Source: D Factor Grammar, Intents, and Example… · p. 12

E Sampling Robustness

Five factor libraries generated from one checkpoint show comparatively stable terminal NAV, with greater relative variation in maximum drawdown. These repetitions measure sampling variability, not independent training-run uncertainty or performance across additional market periods.

E Sampling Robustness

  • Across five generations from one trained checkpoint, mean terminal NAV was 1.906 ± 0.084, SR 1.868 ± 0.253, and MDD (12.76 ± 2.94)% (sample SD); NAV CV was 4.4%.Source: E Sampling Robustness · p. 12
  • All five generations have NAV above 1.80 and MDD below 18%, while the repetitions do not assess independent training runs or additional market periods.The experiment measures sampling variability from one checkpoint.Source: E Sampling Robustness · p. 13

F Case Studies

Case studies characterize AlphaLedger’s style exposures, factor similarities, index-relative performance, return attribution, and unconstrained sector rotation. Style-controlled portfolios retain positive returns against the market reference, though their NAVs fall below the coverage-matched reference.

Case-study scope

  • These diagnostics measure sensitivity to five technical styles but do not identify learned factors’ causal contribution or control all risk exposures.The style exposures use previous-close descriptors on actual beginning-of-day holdings; Figure 8 shows five CNE5 styles and news attention separately.Source: F Case Studies · p. 13, Figure 8 · p. 17

Style exposures

  • Average style exposure is −1.58 for size, +0.35 for beta, −0.26 for momentum, +0.06 for residual volatility, and +0.74 for liquidity.Size is the dominant average tilt; news attention averages −0.39 despite a temporary rise during the June–July 2026 rotation, as EviPool eligibility is followed by ranking from factor scores and RRG weights.Source: F Case Studies · pp. 13–14, Figure 8 · p. 17

Margin over indices

  • AlphaLedger finishes ahead of all five indices and remains above CSI300, CSI500, and CSI1000 throughout the displayed period.It leads ChiNext on 64% of trading days and STAR50 on 79%; Figure 6 plots NAV relative to each index, with green above and red below.Source: F Case Studies · p. 14, Figure 6 · p. 15

Factor similarity

  • More than half of selected factors have every pairwise score correlation below |ρ| = 0.3, while other factors form correlated groups that cross intent labels.The library includes near-duplicate pairs and factors moving against the dominant group, so the eleven intents are treated as complementary elicitation strategies rather than disjoint behaviors.Source: F Case Studies · p. 15, Figure 7 · p. 16

Return attribution

  • The fitted common intercept contributes +22.9 pp, while the model-unexplained component contributes +70.5 pp and accumulates strongly in 2026.The four non-beta styles jointly contribute −1.1 pp; the attribution uses an intercept and five reconstructed CNE5 styles and omits trading costs.Source: F Case Studies · pp. 15–16, Figure 9 · p. 17

Style-controlled portfolios

  • Joint control of all five measured styles retains 38.8% cumulative net return, while controlled portfolios finish below the coverage-matched reference.Single-style controls finish at NAVs from 1.494 for size to 1.730 for residual volatility; all controls finish above the broad-market reference, and the joint control reaches NAV 1.388.Source: F Case Studies · p. 16, Figure 10 · p. 18

Sector rotation

  • Materials decline from 20% of holdings in early 2025 to about 10% by mid-2026, while technology and electronics rise from 16% to 26%.The portfolio has no sector constraint; healthcare stays between 12% and 19%, and the rotation follows market leadership while emerging from factor scores and the moving universe.Source: F Case Studies · p. 16, Figure 11 · p. 18

Related papers

Loading 2610.23744v1…