PaperlayerView exact revision on arXiv

Source-linked AI summary

Reasoning Externalization for Faithful Large Language Model Narratives of Stock Return Predictions

Sujung Kim, Seung Hwan Cho, Sangjin Park, Young-Min Kim

arXiv:2609.38869v2cs.AIcs.CE

TL;DR

Financial SHAP outputs can be difficult for non-experts to interpret, and temporal market context has not been systematically incorporated into LLM explanations. The paper combines temporal SHAP evidence with historical analogs and progressively externalizes reasoning; it reports improved faithfulness and temporal and relational accuracy, while analogs add human-rated usefulness.

  • Problem

    Numerical SHAP explanations can be difficult for non-experts, and a single-time-point explanation may not convey how a model interprets changing market conditions.

    Source: §1 · p. 1

  • Method

    The framework supplies temporal XAI evidence and similar historical regimes to LLMs, progressively externalizes numerical and relational reasoning, and links claims to evidence IDs.

    Source: §1 · p. 2, §3.5 · p. 6

  • Results

    Externalizing attribution changes and trend information improved faithfulness and temporal reasoning, while externalizing feature relations further improved relational reasoning.

    Source: §5 · p. 11

  • Takeaways & Limitations

    Structuring error-prone reasoning steps as verifiable evidence can improve financial XAI narrative reliability, while historical analogs provide complementary interpretive context.

    Source: §1 · p. 2, §5 · p. 11

    Structured Evidence Faithfulness checks claim-field correspondence with the evidence registry, not the semantic truthfulness of unrestricted narrative text.

    Source: §4.1.2 · p. 6

Abstract

from arXiv · show

In finance, interpreting machine learning predictions is essential, yet the numerical outputs of explainable AI can be difficult for non-experts to understand. While large language models (LLMs) can translate these outputs into natural language, they may produce errors when inferring numerical changes and feature relations. We propose an LLM narrative framework for cross-sectional stock return prediction that combines temporal Shapley additive explanations (SHAP) evidence with historical regime analogs. Temporal evidence tracks changes in the normalized global SHAP importance of an XGBoost model over six months. Historical analogs are past periods with similar changes in SHAP importance, their model performance and subsequent market returns are provided as comparative context. Using this framework, we conduct a controlled study of progressive reasoning externalization, sequentially providing raw SHAP sequences, deterministic temporal descriptors, and feature relations. Each generated claim is verified against provenance-linked evidence. Across Qwen3, externalizing numerical and relational reasoning improved evidence faithfulness as well as temporal and relational accuracy. Evidence faithfulness increased from 0.696 to 0.996 for Qwen3-32B-Instruct. While historical analogs did not improve structured automatic faithfulness, they received higher human-rated usefulness scores. These results suggest that externalizing verifiable reasoning enhances narrative faithfulness and that historical context adds interpretive value.

1. Introduction

Financial explanations must make model predictions understandable, but numerical SHAP outputs are difficult for non-experts and temporal market context is not systematically incorporated into LLM explanations. This study combines temporal XAI evidence and historical analogs with a controlled test of reasoning externalization and claim-level evidence verification.

1. Introduction

  • Numerical SHAP explanations can be difficult for non-experts, even though interpretability is important for financial decisions and regulatory compliance.LLMs can translate SHAP outputs into natural-language explanations, but the paper notes that a single time point does not convey how the model interprets changing market conditions.Source: §1 · p. 1
  • The framework supplies LLMs with temporal XAI evidence for cross-sectional stock-return narratives and retrieves historical periods with similar attribution trajectories as context.It uses an XGBoost model trained on preprocessed CRSP data and studies progressive externalization of numerical and relational reasoning.Source: §1 · p. 2
  • Each generated claim is checked against provenance-linked evidence, addressing the need to verify financial narratives for faithfulness.The study reports that externalizing numerical and relational reasoning reduces temporal and relational errors, while historical analogs add human-rated usefulness without improving structured automatic faithfulness.Source: §1 · p. 2

2. Related Work

Prior financial SHAP research examined feature contributions and their changes over time, while LLM explanation work translated model outputs into narratives and explored contextual inputs. The paper positions its approach at the intersection of temporal attribution patterns, historical regime retrieval, and evidence-faithful narratives.

2.1. SHAP Based Feature Analysis in Financial Prediction

  • Earlier financial SHAP studies mostly analyzed feature importance statically, while dynamic studies tracked attribution changes to monitor models rather than retrieve similar regimes for downstream analysis.Prior work reported contributions from momentum, trading, and sentiment features, and examined changes across market regimes.Source: §2.1 · p. 2

2.2. LLM-Based Explanation and Context-Augmented Financial Analysis

  • LLM explanation studies translated SHAP and other model outputs into narratives, but prior work raised concerns about fidelity and limited validation of practical decision usefulness.Reported fidelity failures include reversed SHAP signs, omitted dominant features, and features absent from the underlying evidence.Source: §2.2 · p. 3
  • Context-augmented financial LLM studies used point-in-time information or retrieved similar periods using macroeconomic variables, rather than temporal patterns in model feature attributions.The paper distinguishes its approach by using SHAP attribution dynamics to identify historical analogs.Source: §2.2 · p. 3

3. Methodology

The methodology builds monthly stock-return predictions and temporal SHAP evidence, retrieves historical analogs from similar attribution changes, and stores evidence with provenance identifiers. It varies which evidence and reasoning support are exposed to the LLM while keeping the underlying registry fixed.

3.1. Data Preparation and Stock Return Prediction

  • Monthly stock characteristics are used to predict each stock’s realized return in the following month, with temporal ordering preserved in the data split.The sample uses monthly CRSP data and separates training, validation, and test periods chronologically.Source: §3.1 · p. 3

3.2. Temporal XAI Evidence Construction

  • Monthly global feature attribution is computed from mean absolute SHAP values, normalized across features, and assembled into a six-month temporal trajectory.The resulting attribution vector represents each feature’s relative importance in the model’s global attribution for that month.Source: §3.2 · p. 4

3.3. Historical SHAP Based Analog Augmentation

  • Historical analogs are retrieved by comparing trajectories of changes in feature importance, with model performance and subsequent universe returns supplied as comparative context.Candidates must not overlap the query period, and the analog outcomes are not treated as direct forecasts or causal evidence.Source: §3.3 · pp. 4–5

3.4. Structured Evidence Registry

  • Predictions, attributions, temporal descriptors, feature relations, and analog evidence are stored in a unified registry with provenance identifiers.All conditions use the same underlying registry; they differ in which evidence units are visible to the LLM.Source: §3.4 · p. 5

3.5. Reasoning Externalization and Narrative Generation

  • Reasoning is progressively externalized by providing raw temporal sequences, deterministic temporal descriptors, and feature relations, with a separate condition adding historical context.Outputs pair narrative text with structured claims, require evidence IDs for factual claims, and permit abstention when evidence is unavailable.Source: §3.5 · pp. 5–6

4. Experiments and Results

The experiments evaluate prediction performance, structured evidence faithfulness, temporal reasoning, and human judgments of narrative utility. They also use a separate judge evaluation to assess temporal accuracy, evidence scope, and synthesis.

4.1.1. Prediction Performance Evaluation

  • Prediction performance is evaluated with RankIC, RankICIR, RMSE, and MAE, with RankIC computed as monthly cross-sectional Spearman correlation and averaged over the test period.RankICIR measures the stability of cross-sectional ranking performance over time.Source: §4.1.1 · p. 6

4.1.2. Deterministic Automatic Evaluation

  • Structured Evidence Faithfulness measures whether factual claims cite permitted evidence and match ground-truth subject–predicate–value tuples; evaluation also includes parse rate and abstention accuracy.The metric concerns structured fields rather than the semantic truthfulness of unrestricted narrative text.Source: §4.1.2 · pp. 6–7

4.1.3. LLM-as-a-Judge

  • The judge evaluation uses narratives to assess six-month trend identification, and narrative-plus-evidence pairs to rate evidence scope adherence and synthesis.A month-stratified sample of 400 paired cases is judged once by Claude Opus 5, with model and condition identifiers hidden.Source: §4.1.3 · p. 7

4.1.4. Human Evaluation

  • Ten non-expert evaluators compare paired C3-Relational and C3-Analog narratives using five-point ratings for understandability, usefulness, reliance calibration, and adoption intention.The evaluation separately examines whether historical analogs add value for interpretation and use.Source: §4.1.4 · p. 7

4.2. Experimental Setup and Prediction Model Selection

The study prioritizes narrative faithfulness over state-of-the-art prediction, selecting XGBoost on validation performance and fixing it for test-period XAI analyses.

4.2. Experimental Setup and Prediction Model Selection

  • The comparison included Linear Regression and Ridge Regression alongside Random Forest and XGBoost, with the study focused on narrative faithfulness rather than predictive performance competition.Tree-based models were included because they can capture nonlinear relationships and feature interactions while supporting TreeSHAP attribution.Source: §4.2 · p. 7
  • XGBoost achieved 1.4% higher RankIC and 9.8% higher RankICIR than Random Forest, while Random Forest had lower RMSE and MAE.The comparison shows XGBoost did not dominate across all predictive metrics.Source: Table 2 · p. 8, §4.2 · p. 8
  • XGBoost was selected using validation-period performance, fixed for all test-period XAI analyses, and interpreted with TreeSHAP.The test results were reserved for final out-of-sample evaluation.Source: §4.2 · p. 8

4.3. Historical Analog Characterization

The authors evaluate SHAP-trajectory analog retrieval as comparative context, testing its return-prediction error across retrieval settings and against alternative baselines.

4.3. Historical Analog Characterization

  • The sensitivity analysis varied temporal window length L across 3, 6, and 9 months and the number of retrieved analogs K across 1, 3, and 5.Comparisons included market-state retrieval, a recent-return baseline, and random retrieval, using MAE against realized next-month universe returns.Source: §4.3 · p. 8
  • The selected SHAP-Trajectory configuration, K=5 and L=6, minimized validation MAE among evaluated configurations at 0.03996.This configuration was selected before the test-period evaluation of C3-Analog.Source: Table 3 · p. 8, §4.3 · p. 8
  • SHAP-Trajectory retrieval was not consistently superior: the recent-return baseline had lower MAE in six of nine settings, and selected-setting MAE was 0.03996 versus 0.03997 for market-state retrieval.The retrieved analogs were therefore used as historical comparative context, not as evidence of consistent retrieval superiority.Source: §4.3 · p. 8

4.4. Reasoning Externalization and Narrative Faithfulness

The controlled comparisons show that externalizing temporal and relational reasoning improves structured claim performance, while historical analogs add no further automatic faithfulness gain.

4.4. Reasoning Externalization and Narrative Faithfulness

  • Across Qwen3 Instruct models, C1-to-C2 increased Structured Temporal Accuracy, while C2-to-C3 increased Relational Accuracy across the tested model sizes.Qwen3 Base models also improved in temporal accuracy from C1 to C2; Ministral and Llama were excluded from direct comparisons because of low parse rates.Source: Table 4 · p. 9, §4.4 · p. 9
  • For Qwen3-32B-Instruct, C3 raised Relational Accuracy from 0.5729 to 1.0000 and EF to 0.9964; C1-to-C2 raised EF from 0.6960 to 0.9032.C1-to-C2 also increased Structured Temporal Accuracy from 0.3396 to 0.9896, while C3 Structured Temporal Accuracy was 0.9313.Source: Table 5 · p. 9, §4.4 · p. 9
  • C3-Analog retained Relational Accuracy of 1.0000 but had lower EF and Structured Temporal Accuracy than C3-Relational; C0 abstention accuracy was 0.8386.The authors conclude that historical analog augmentation did not improve structured automatic faithfulness beyond C3-Relational.Source: §4.4 · p. 9

4.5. LLM-as-a-Judge Evaluation

The judge-based evaluation finds high temporal transfer and evidence-scope adherence, but low narrative synthesis scores.

4.5. LLM-as-a-Judge Evaluation

  • Textual Temporal Transfer Accuracy was 0.9975 (399/400) in both C3-Relational and C3-Analog, while Deterministic Temporal Accuracy was 0.8625 and 0.7425, respectively.The reported discrepancy arose from claim-type-specific aggregation, not an actual loss of temporal information in the narrative.Source: Table 6 · p. 10, §4.5 · p. 10
  • Evidence Scope Adherence was approximately 4.9 out of 5 in both conditions, while Narrative Synthesis scores were approximately 2 out of 5.The authors suggest evidence-constrained generation may limit flexibility in integrating multiple evidence items into a narrative.Source: Table 6 · p. 10, §4.5 · p. 10

4.6. Human Evaluation of Historical Analog Context

C3-Analog received higher human ratings than C3-Relational across all five evaluation criteria, despite slightly lower automatic faithfulness and structured temporal accuracy. The authors interpret historical analogs as complementary context for human-level usefulness rather than a way to improve automatic faithfulness.

4.6. Human Evaluation of Historical Analog Context

  • 3.6200 versus 3.1600 overall mean (+0.4600): C3-Analog rated higher than C3-Relational across all five human-evaluation criteria.The comparison covered understandability, decision usefulness, contextual usefulness, reliance calibration, and adoption intention.Source: §4.6 · p. 10
  • Adoption Intention had the largest between-condition difference (+0.5500), followed by Contextual Usefulness (+0.5000).Source: §4.6 · p. 10
  • C3-Analog's higher human ratings occurred despite slightly lower EF and Structured Temporal Accuracy than C3-Relational.Source: §4.6 · p. 10
  • Historical analogs are characterized as complementary context for human-level contextual usefulness, not as a mechanism for improving automatic faithfulness.Source: §4.6 · pp. 10–11

5. Conclusion

The conclusion finds that externalizing temporal and relational reasoning improves financial XAI narrative faithfulness and reasoning accuracy relative to raw temporal evidence. It calls for testing whether these findings generalize to other predictive models and financial decision-making tasks.

5. Conclusion

  • The study structured SHAP temporal evidence as numerical and relational information and linked factual claims to Evidence IDs for deterministic evaluation without LLM-as-a-Judge.Source: §5 · p. 11
  • Explicit attribution changes and trend information improved faithfulness and temporal reasoning versus raw temporal sequences; externalized feature relations further improved relational reasoning.Source: §5 · p. 11
  • Future work should test whether the findings generalize to other predictive models and financial decision-making tasks.Source: §5 · p. 11

A. Stock Level Characteristic Definitions

The appendix lists stock-level characteristic definitions, formulas, measurement windows, and CRSP sources. The supplied entries include adjusted-price-based characteristics, volume and return measures, and trailing-window indicators.

A. Stock Level Characteristic Definitions

Related papers

Loading 2609.38869v2…