PaperlayerView exact revision on arXiv

Source-linked AI summary

When News Gets Old: Portfolio Allocation with Reinforcement Learning and Large Language Models

Bolin Song, Surui Xiang

arXiv:2610.30564v1cs.LGq-fin.PM

TL;DR

The paper examines whether aging financial-news signals should be retained with timing information or filtered before portfolio selection. It compares RL and direct-LLM portfolio policies using SEC-disclosure signals, then finds that filtering old signals yields the LLM’s highest average net return, while timing information does not consistently improve RL returns.

  • Problem

    As reports age, their usefulness for new trades becomes less certain, raising whether policies should receive timing information or have old signals removed.

    Source: §I · p. 1, A. Observations and News Timing · p. 1

  • Method

    The paper compares RL and direct-LLM portfolio policies using fixed LLM-extracted SEC-disclosure signals, with old signals either retained, accompanied by timing inputs, or filtered.

    Source: A. Observations and News Timing · p. 1, §V · p. 4

  • Results

    Filtering old signals gives the LLM its highest average net return, reaching 18.43% per year after costs at a four-week delivery delay and exceeding all three news-using RL policies.

    Source: §V · p. 4

  • Takeaways & Limitations

    Signal filtering is a practical part of LLM-based portfolio design, while RL sensitivity to news timing and investment performance are distinct outcomes to evaluate together.

    Source: §V · p. 4

Abstract

from arXiv · show

Financial news can influence investment decisions long after publication. Timely news may reach an investor before prices fully adjust, whereas older news may already be reflected in prices. We examine two ways to handle aging signals: retain them and give a portfolio policy timing information, or filter them out before portfolio selection. A fixed large language model (LLM) extracts signals from public SEC disclosures. We compare these treatments in reinforcement learning (RL) policies and LLMs that choose portfolios directly. We evaluate returns after trading costs in historical tests over 2023-2025. RL choices respond to timing information in a controlled test, but historical returns show no consistent benefit from adding it. For the LLM, filtering old news produced the highest average net return, reaching 18.43% per year when news delivery was delayed by four weeks. This exceeded the returns of all RL strategies that used news.

I. INTRODUCTION

The paper asks whether aging SEC-disclosure signals should be retained with timing information or filtered before portfolio selection. It compares these treatments in RL and direct-LLM portfolio policies, with timing inputs representing report age and time since arrival.

I. INTRODUCTION

  • The paper compares retaining signals with timing information against removing signals once their reports are more than eight weeks old.It also tests signals without timing information and measures returns after trading costs under zero-, one-, and four-week delivery delays.Source: §I · p. 1
  • For the LLM policy, filtering old signals produces the highest average net return; at a four-week delivery delay it earns 18.43% per year after costs, exceeding all three RL policies that use news.The paper contrasts this with RL, where timing information changes choices but does not consistently improve historical net returns.Source: §I · p. 1

A. Observations and News Timing

  • A fixed LLM extracts positive, neutral, or negative signals from SEC disclosures, which RL or an LLM combines with market data and holdings to select a portfolio.Signals persist until newer reports arrive unless filtered.Source: A. Observations and News Timing · p. 1, §I · p. 1
  • Timing information comprises weeks since publication and weeks since arrival, so a report published four weeks earlier but arriving today has age four and time since arrival zero.The RL policy and direct LLM baseline choose portfolios independently, with separate holdings under common backtesting rules.Source: A. Observations and News Timing · p. 1, Figure 1 · p. 2

B. Actions, Reward, and Learning

Policies choose among predefined target portfolios or keep their holdings, with trades incurring costs. RL and LLM policies receive corresponding information inputs, while each RL setting uses a separately trained Double DQN.

B. Actions, Reward, and Learning

  • Each week, a policy selects a predefined target portfolio or keeps its holdings; rebalancing incurs trading costs, while HOLD preserves existing shares.Strategies are fully invested and share an execution schedule.Source: B. Actions, Reward, and Learning · p. 2
  • The reward and objective use wealth after trading costs, with the horizon, discount factor, and reward-scale parameter specified in the accompanying definitions.The supplied passage identifies T as the horizon, γ as discounting later rewards, and κ as setting reward scale.Source: B. Actions, Reward, and Learning · p. 2
  • Each RL information setting has a separately trained Double DQN; during evaluation it selects the highest-value action without exploration or further training.LLM policies choose from the same portfolio list, and invalid outputs map to HOLD.Source: B. Actions, Reward, and Learning · p. 2

III. EXPERIMENTAL DESIGN

The historical comparison tests weekly allocation for ten US stocks over 2023–2025 using historical prices and signals from periodic SEC disclosures. Methods share the data and trading rules, while portfolio decision models and report-delivery delays vary.

III. EXPERIMENTAL DESIGN

  • The evaluation covers ten US stocks over 2023–2025, with each test year beginning from an equal-weight portfolio.RL training and validation use data from years before each annual test.Source: A. Historical Comparison · p. 2
  • Reports arrive with zero-, one-, or four-week delays, and trades execute at the following week’s first open with a fee of 10 basis points per dollar traded.Report text and publication dates remain unchanged across delay settings.Source: A. Historical Comparison · p. 2
  • GPT-4.1 mini extracts signals and selects portfolios for the LLM results; at a four-week delay, Llama 3.3 and GPT-6 Astra are also tested as decision makers using the same signals.Model settings are given in Appendix C.Source: A. Historical Comparison · p. 2

B. Decision Changes and Net Returns

The design separates historical returns under delivery delays from a controlled test of whether older reports change RL portfolio choices. It also specifies how portfolio outcomes and uncertainty are evaluated.

B. Decision Changes and Net Returns

  • The controlled test increases report age and time since arrival equally while holding reports, market data, holdings, and original signals fixed; filtered policies reapply the eight-week cutoff.This test measures changes in portfolio choice, not returns from those changes.Source: B. Decision Changes and Net Returns · pp. 2–3
  • Investment outcomes include annualized return, Sharpe ratio, maximum drawdown, trading volume, and fees, with News serving as the return baseline for News + Age and Filtered News.Metrics are averaged equally across annual tests.Source: B. Decision Changes and Net Returns · p. 3
  • Paired time-block intervals resample the same weeks across methods and trained policies to distinguish uncertainty in return differences from RL training variation.Supporting checks vary fees and examine signal persistence in a synthetic market.Source: B. Decision Changes and Net Returns · p. 3

IV. RESULTS

Filtering old signals gave the LLM policy its highest average net return. In the controlled test, RL choices also shifted as reports aged.

A. LLM Portfolio Returns

  • At a four-week delivery delay, the filtered LLM earned 18.43% average annualized net return, beating all three news-using RL policies; RL Market Data returned 19.07%.Source: A. LLM Portfolio Returns · p. 3
  • The effect of adding timing information to LLM policies varies by model and year.Source: A. LLM Portfolio Returns · p. 4

B. RL Responses to News Age

C. RL Portfolio Returns

Historical returns did not consistently improve when RL received timing information, and higher trading fees reduced its net returns. The conclusion distinguishes RL’s responsiveness to news timing from investment performance.

C. RL Portfolio Returns

  • At four weeks, News + Age returned 18.02% versus News at 17.81% and Filtered News at 18.03%; confidence intervals for both gains over News included zero.Source: C. RL Portfolio Returns · p. 4, Figure 3 · p. 4
  • At a four-week delay, filtering removed roughly 70% of nonzero news signals, with no clear return advantage from retaining them and supplying timing information.Source: C. RL Portfolio Returns · p. 4

D. Trading Costs

  • Raising fees from 10 to 30 basis points per dollar traded lowered RL with News + Age’s average annualized net return from 18.02% to 12.32%.Source: D. Trading Costs · p. 4

V. CONCLUSION

  • With a four-week delivery delay, LLM filtering earned 18.43% average annualized net return, exceeding all three RL policies that used news.Source: §V · p. 4
  • For RL, responding to news timing and improving investment performance are distinct outcomes that should be evaluated together.Source: §V · p. 4

APPENDIX A ANNUAL RESULTS AND ADDITIONAL DIAGNOSTICS

Historical timing and filtering comparisons are uncertain across annual periods and models, while trained policies show varied allocation responses as reports age. The execution rule accounts for costs using actual execution holdings, distinct from decision-time policy inputs.

A. Uncertainty and Annual Variation

  • None of the 36 primary annual timing/filter comparisons has an eight-week-block 95% interval excluding zero; timing effects also vary across training runs.With 13-week blocks, only mini filtering at zero delay in 2025 excludes zero; these intervals are unadjusted and exploratory.Source: A. Uncertainty and Annual Variation · p. 5

B. LLM Portfolio Policies

  • All nine four-week-delay paired intervals for LLM News + Age versus News contain zero, and timing effects vary by model and annual period.The comparison holds mini-extracted signals fixed and changes only the portfolio-selection model.Source: B. LLM Portfolio Policies · p. 5

C. Variation in the Behavioral Response

  • At twelve weeks of added age, action-change rates range from 25.6% to 74.4% for News + Age and 7.7% to 41.0% for Filtered News.Signal-aligned allocation responses do not decline steadily with age: some trained policies increase them, while others decrease them.Source: C. Variation in the Behavioral Response · p. 5

A. Self-Financing Execution

  • Fees use actual execution holdings, while policy inputs use only decision-time information; HOLD leaves holdings unchanged and incurs no trade.The fraction of pretrade wealth left after fees accounts for the fee on dollars bought or sold, including cash needed to pay the fee.Source: A. Self-Financing Execution · p. 5

B. Annual Folds and Training

The annual-fold design trains RL policies anew across historical folds and evaluates fixed-budget checkpoints. The historical input and action spaces retain feature slots, offer defined portfolio choices, and include HOLD.

B. Annual Folds and Training

  • Table V is scoped to annual return differences at four-week delay with paired 95% intervals.The caption identifies the table’s scope but supplies no cell values.Source: Table V · p. 6
  • Each RL method is retrained in every fold with ten seeds, yielding 120 trained policies; final fixed-budget checkpoints are evaluated greedily.Validation blocks do not select hyperparameters, seeds, filtering thresholds, or checkpoints; overlapping training windows reuse observations from the same price history.Source: B. Annual Folds and Training · p. 6
  • Training uses 52-week windows with 26 weeks of prior feature history and mixes zero-, one-, and four-week delays.The folds contain 52, 104, and 156 eligible window starts, respectively.Source: B. Annual Folds and Training · p. 6
  • The Double DQN target is trained with a Huber (smooth L1) loss against the network’s predicted action value.The target uses the online network and a periodically copied target network; the passage defines termination and the parameter vectors.Source: B. Annual Folds and Training · p. 6

C. Inputs and Portfolio Choices

  • Historical RL networks retain 80 feature slots, with unavailable information zero-masked; market features summarize returns and volatility over stated weekly windows.Timing inputs are capped at 104 weeks, and missing reports have signal zero.Source: C. Inputs and Portfolio Choices · p. 6
  • The historical action menu has 57 choices, including equal allocation, single- and two-stock tilts, and HOLD; target portfolios are long-only and fully invested.HOLD retains existing holdings; the supporting four-asset simulator uses a separate 12-action menu.Source: C. Inputs and Portfolio Choices · p. 6

D. Supporting Simulator and Public Data

The supporting simulator varies hidden-state persistence and tests signal and timing effects, while the public-data evaluation uses a defined SEC corpus and reports extraction compliance separately from semantic accuracy. Behavior and cost checks use fixed observations or actions, and annual metrics are computed separately before equal-year averaging.

D. Supporting Simulator and Public Data

  • The simulator models hidden states that persist with probability .85 or .97 and uses report labels that match the hidden state with probability 0.80 when non-neutral.Policies observe stored signals, not the hidden state or report label; the test uses 50 validation and 100 test paths per persistence value.Source: D. Supporting Simulator and Public Data · p. 6
  • At four-week delay, adding timing changes simulated annualized returns by +0.35% in the faster-changing market and −0.32% in the slower one; both path-paired 95% intervals include zero.Replacing stored signals with true semantic labels changed the fixed RL policy’s mean return by −0.02% ([−0.16, 0.13]%).Source: D. Supporting Simulator and Public Data · p. 6
  • The historical corpus contains 277 SEC excerpts published from 2019 through 2025, including 119 reports in the 2023–2025 cohorts.The ten-stock universe is selected by a stated procedure; later returns or coverage do not select or replace stocks.Source: D. Supporting Simulator and Public Data · p. 7
  • The extractor correctly classifies 191 of 200 held-out synthetic reports, while 228 of 277 real-text outputs meet format and verbatim-evidence requirements.The real-text rate measures compliance, not semantic accuracy; semantic accuracy cannot be assessed without human labels for these excerpts.Source: D. Supporting Simulator and Public Data · p. 7

E. Behavior and Cost Checks

  • Behavior probes shift already-visible report times while preserving missing-report sentinels; annual cost checks replay saved actions at 0, 10, and 30 basis points.At 10 basis points, replay reproduces the recorded weekly paths and metrics.Source: E. Behavior and Cost Checks · p. 7

F. Prices and Statistical Summaries

  • Annual tests compound returns and calculate metrics separately, then average each metric with equal year weights; holdings reset at each annual boundary.The reported metrics include net annualized return, Sharpe ratio, maximum drawdown, trading volume, and fees.Source: F. Prices and Statistical Summaries · p. 7

APPENDIX C PROMPTS AND A RECORDED PORTFOLIO DECISION

The appendix presents prompt materials and model settings, then specifies how input contracts vary across portfolio prompts and how decisions are recorded.

APPENDIX C PROMPTS AND A RECORDED PORTFOLIO DECISION

  • The appendix’s prompt boxes show SEC extraction instructions and a portfolio-decision prompt, with minor wording edits and typography changes for readability.Exact executed prompts, full requests, and the supporting synthetic-report prompt remain in the reproduction materials.Source: Appendix C · p. 7

A. Model Settings and Input Contracts

  • The extractor and main direct baseline use gpt-4.1-mini-2025-04-14.The supporting direct models request Llama-3.3-70B-Instruct-Turbo and gpt-6-astra; Astra is an API alias, not a dated snapshot.Source: A. Model Settings and Input Contracts · p. 7
  • The mini extractor uses temperature zero and a 192-token output limit.The appendix separately specifies generation settings for direct models and notes that Astra sends no temperature field.Source: A. Model Settings and Input Contracts · p. 7
  • Within each model, prompts retain the objective, catalog, and output contract while market, News, and News + Age inputs contain five, six, and eight fields per asset.Smaller prompts remove absent-field definitions and adjust the stated field count.Source: A. Model Settings and Input Contracts · p. 7
  • Filtered News uses the News prompt with locally zeroed signals and does not send the age threshold or numerical timing inputs.Market-only trajectories are reused across delay views.Source: A. Model Settings and Input Contracts · p. 7
  • Of the test requests, 2,825 return valid actions; one service failure uses HOLD without retry, and no API call occurs inside RL training or evaluation.Decision requests and raw responses remain in the reproduction package.Source: A. Model Settings and Input Contracts · p. 7

B. From a Report to a Portfolio

The pipeline converts public SEC excerpts into operating-performance signals and supplies those signals, market features, holdings, and report-timing information to a portfolio policy. The policy selects from a catalog of actions under stated trading costs.

B. From a Report to a Portfolio

  • In the Boeing example, the extractor assigns a positive signal to a filing excerpt reporting a decrease in operating loss.The reported decrease is $2,118 million, and the excerpt was published 9.3100 weeks before the decision cutoff.Source: B. From a Report to a Portfolio · p. 7
  • The decision input combines anonymous market features, holdings, and stored signals with publication age and time since the report became available.The timing features are capped at 104 weeks; missing text has signal 0 and both times 104.Source: B. From a Report to a Portfolio · pp. 7–8
  • Signals classify excerpt-supported operating performance or outlook; lower costs or losses may be favorable, but the signal is not verified economic or market ground truth.The extraction rules also exclude trading recommendations and return predictions.Source: B. From a Report to a Portfolio · p. 8
  • Each asset’s eight features include past returns, volatility, current portfolio weight, a cached signal, and two measures of report timing.The prompt defines the return and volatility fields and describes signals as negative, neutral, positive, or unavailable.Source: B. From a Report to a Portfolio · p. 8
  • The policy chooses an action from a supplied catalog; HOLD preserves holdings without a fee, while other actions rebalance at the next execution price.Trading costs are 10 basis points for each dollar bought and each dollar sold.Source: B. From a Report to a Portfolio · p. 8

Related papers

Loading 2610.30564v1…