LENS: Reliability-Aware Stock Prediction via LLM-Extracted Priors and News-Driven Semantic Graphs

Hao Zhou; Tiejun Ma; Luis Felipe Costa Sperb

cs.LGNov 3, 2026

LENS: Reliability-Aware Stock Prediction via LLM-Extracted Priors and News-Driven Semantic Graphs

Authors: Hao Zhou, Tiejun Ma, Luis Felipe Costa Sperb

Organizations: School of Informatics, The University of Edinburgh, Edinburgh, United Kingdom · Kingston Business School, Kingston University, London, United Kingdom

Abstract

Predicting stock returns requires modeling persistent company relations and time-varying market signals. Graph Neural Network (GNN) approaches combine static priors with dynamic relational graphs, but treat every extracted edge as a trustworthy binary fact, ignoring heterogeneous reliability across relational sources. We propose LENS (LLM-Extracted priors and News-driven Semantics), a static--dynamic graph framework for stock return prediction and ranking that models relational reliability as a first-class signal across the pipeline. On the static side, we extract a multi-relational company graph from a Large Language Model and aggregate it over multiple runs with confidence weighting to downweight low-consistency edges. On the dynamic side, we build a news-driven semantic graph with exponential-decay temporal aggregation to capture changing market context under uneven coverage. A context-dependent gated fusion module rebalances the two graphs at the per-stock, per-day level. On S&P 500, LENS improves AR@5 by 16.4% and SR@5 by 22.6% over the strongest baseline, and on NASDAQ 100 it achieves comparable improvements (15.5% and 20.9% respectively). Ablations isolate each component's contribution, and a human-validation study of sampled relations supports that the aggregation is grounded, providing evidence that reliability-aware LLM relations and news-based semantic graphs carry complementary information for graph-based financial prediction.

Figures & tables

Figure 1. Overview of LENS. The framework consists of (1) static relation prior construction via multi-run LLM extraction and reliability-aware aggregation, (2) dynamic semantic graph construction from temporally aggregated news embeddings, (3) a dual-component GNN encoder over the static and dynamic graphs, and (4) a gated fusion module followed by a return predictor. End-to-end pipeline of LENS: market features and news are encoded; a static multi-relational company graph is extracted by an LLM and aggregated across runs with reliability weights; a dynamic weighted semantic graph is built from time-windowed news embeddings; two graph encoders produce static and dynamic node representations; a gated fusion combines them to output next-day return/ranking scores.

	S&P 500							NASDAQ 100
	MRR $\uparrow$	AR@5 $\uparrow$	SR@5 $\uparrow$	AR@10 $\uparrow$	SR@10 $\uparrow$	AR@20 $\uparrow$	SR@20 $\uparrow$	MRR $\uparrow$	AR@5 $\uparrow$	SR@5 $\uparrow$	AR@10 $\uparrow$	SR@10 $\uparrow$	AR@20 $\uparrow$	SR@20 $\uparrow$
o3-mini	0.0232	0.2501	1.0407	0.1534	0.8023	0.1617	0.9481	0.0638	0.1872	0.7890	0.1549	0.7436	0.1206	0.7351
FinGPT	0.0362	0.3997	1.4424	0.3172	1.3423	0.2772	1.3100	0.0852	0.3127	1.0853	0.2104	0.9010	0.2098	1.0394
MLP	0.0359	0.3564	1.1679	0.3128	1.2111	0.2585	1.1595	0.0859	0.2734	0.8419	0.1953	0.7826	0.2167	0.9128
ALSTM	0.0275	0.2401	1.0153	0.2619	1.2322	0.2663	1.3107	0.0712	0.1684	0.7128	0.1745	0.8273	0.2089	1.0431
Transformer	0.0364	0.3287	1.2402	0.3235	1.4030	0.2628	1.2959	0.0915	0.2543	0.8957	0.2127	0.9381	0.2178	1.0250

Table 1. Main performance comparison on S&P 500 and NASDAQ 100. We report the mean over 10 seeds. For each metric within each market, the best result is in bold and the second-best is underlined .

\uparrow

indicates higher is better. AR and SR are reported with a 5 bps transaction cost. ** denotes

p<0.01

and * denotes

p<0.05

via paired t-test over 10 seeds.

Highrecall: subset breakdowns
Aggregation	# Triples	Relaxed	Strict
Single-run	30	83.3%	56.7%
Highrecall (LENS)	90	91.1%	54.4%
$f(e)\geq 0.6$ subset	28	92.9%	78.6%
non-competitor relations	52	96.2%	63.5%
Highrecall: by reliability tercile

Table 2. Human validation of the LENS static graph (Fleiss’

\kappa{=}0.58

three-way). Sample composition and definitions of Strict and Relaxed precision are given in 5.7 ; subset rows refer to the Highrecall graph.

Graph Source
Ablation	MRR $\uparrow$	AR@5 $\uparrow$	SR@5 $\uparrow$
w/o Static Graph	0.0361	0.3375	1.2143
GICS	0.0354	0.3724	1.3937
WikiData	0.0351	0.3649	1.4237
News Co-occurrence (static)	0.0370	0.3582	1.3670
Construction

Table 3. Ablation on static graph design (RQ2), S&P 500, mean over 10 seeds. No reliability weight uses the same Highrecall edge set as LENS but replaces

\rho(e)

with

1

	Strict	Highrecall	All
#Edges	3,162	5,424	10,627
Mean $\rho(e)$	0.67	0.49	0.32

Table 4. Static prior graph statistics across aggregation modes (S&P 500).

Figure 2. Heatmaps of (a) SR@5 and (b) MRR under different temporal aggregation settings ( $\tau\in\{1,5,10,20,40\}$ , $\alpha\in\{0.1,0.5,1,2\}$ and mean w/o decay). Two heatmaps compare SR at 5 and MRR across five temporal window lengths and four decay rates, plus mean aggregation without decay. Both metrics reach their highest values with a 10-day window and a decay rate of 0.5.

Dynamic graph	MRR $\uparrow$	AR@5 $\uparrow$	SR@5 $\uparrow$
Price	0.0342	0.3881	1.4627
News co-occurrence (dynamic)	0.0342	0.4238	1.5430
w/o dynamic graph	0.0267	0.2617	1.0854
LENS (LLM static)	0.0371	0.5232	1.8694

Table 5. Ablation on dynamic graph design (RQ3). We report the mean performance over 10 seeds.

Fusion strategy
Variant	MRR $\uparrow$	AR@5 $\uparrow$	SR@5 $\uparrow$
Equal (0.5/0.5)	0.0354	0.4361	1.6278
Concat (MLP)	0.0312	0.3608	1.3756
Learned	0.0371	0.4246	1.5899
Gated (LENS)	0.0371	0.5232	1.8694
Edge-feature injection (static GNN)

Table 6. RQ4: ablations on reliability-aware fusion. Upper panel varies the static/dynamic fusion module; lower panel varies the edge-feature injection in the static GNN (Section 4.5 ). All other components are held fixed. We report the mean performance over 10 seeds.

Figure 3. Static gate weights for representative stocks from different news coverage groups: NVR (low coverage, freq=0.094), BKNG (medium coverage, freq=0.458), and NVDA (high coverage, freq=0.692). Line chart of static gate weights over approximately 320 trading days for NVR, BKNG, and NVDA, representing low, medium, and high news coverage. NVR generally receives higher static weights, and all three stocks show increased weights around day 190.

Method	High	Medium	Low	$\Delta$ (H→L)
Dynamic Only	0.012690	0.012793	0.013374	5.39%
LENS (Equal)	0.012738	0.012768	0.013279	4.24%
LENS (Gated)	0.012752	0.012725	0.013158	3.18%

Table 7. MAE across different news coverage groups. We report the mean performance over 10 seeds.

Stock Price Forecasting

See all 22 papers

Relational Probing: LM-to-Graph Adaptation for Financial Prediction
Apr 11, 2026Yingjie Niu, Changhong Jin, Rian Dolphin +1LLMs for Graph Learning Financial Forecasting
University College Dublin, Belfield, Dublin, Ireland · Massive, Dublin, Ireland
Beyond Sentiment: Structured Information Extraction from Financial News
Jul 30, 2026Daohan Zhu, Sitong Ge, Ruofei Wang +4Financial Sentiment Analysis Document Information Extraction
School of ASEE, Beihang University, Beijing, China · School of BME, Beihang University, Beijing, China · CAIR and CECS, VinUniversity, Hanoi, Vietnam
RICE-Alpha: Reliability-Informed Correction with Event Graphs for LLM-Agent Stock Forecasting
Sep 27, 2026Tong Liu, Lanmiao Liu, Xiang HuStock Price Forecasting Financial Forecasting
Zircon Security · Utrecht University · The Max Planck Institute for Psycholinguistics +1

Apr 11, 2026J/K move · Enter open · S save

Relational Probing: LM-to-Graph Adaptation for Financial Prediction

Yingjie Niu, Changhong Jin, Rian Dolphin, Ruihai Dong

University College Dublin, Belfield, Dublin, Ireland · Massive, Dublin, Ireland

LLMs for Graph Learning Financial Forecasting Graph Structure Learning

Open paper AskPDF arXiv

Language models can be used to identify relationships between financial entities in text. However, while structured output mechanisms exist, prompting-based pipelines still incur autoregressive decoding costs and decouple graph construction from downstream optimization. We propose Relational Probing, which replaces the standard language-model head with a relation head that induces a relational graph directly from language-model hidden states and is trained jointly with the downstream task model for stock-trend prediction. This approach both learns semantic representations and preserves the strict structure of the induced relational graph. It enables language-model outputs to go beyond text, allowing them to be reshaped into task-specific formats for downstream models. To enhance reproducibility, we provide an operational definition of small language models (SLMs): models that can be fine-tuned end-to-end on a single 24GB GPU under specified batch-size and sequence-length settings. Experiments use Qwen3 backbones (0.6B/1.7B/4B) as upstream SLMs and compare against a co-occurrence baseline. Relational Probing yields consistent performance improvements at competitive inference cost.

Abstract

Stock Price Forecasting

Explore similar work

Relational Probing: LM-to-Graph Adaptation for Financial Prediction

Beyond Sentiment: Structured Information Extraction from Financial News

RICE-Alpha: Reliability-Informed Correction with Event Graphs for LLM-Agent Stock Forecasting