ASE: Attribution-Guided Semantic Evolution for Time Series Forecasting

Yuxuan Zheng; Zhaoyang Li; Purui Liu; Hengbo Xiao; Guannan He

cs.AIOct 16, 2026

ASE: Attribution-Guided Semantic Evolution for Time Series Forecasting

Authors: Yuxuan Zheng, Zhaoyang Li, Purui Liu, Hengbo Xiao, Guannan He

Organizations: Peking University

Abstract

External text can supply predictive information absent from numerical histories, but its value depends on how it is represented, combined and deployed. We present Attribution-guided Semantic Evolution (ASE), a modular framework for constructing textual factors, evaluating their conditional contribution, and deploying validation-selected numerical--textual predictors. ASE supports task-specific implementations with lexical measurements, semantic primitives or evolved event/state factors, coupled to suitable numerical experts. Its feedback-based implementation uses conditional refitting diagnostics to guide factor revision and combination search. We evaluate ASE on the four text-associated TESS datasets and dynamic CSI 300 stock ranking. The four forecasting implementations obtain test MSEs of 0.001041, 0.022439, 0.333115 and 1.8150, each below the corresponding published TESS score. On CSI 300, the frozen text-fusion implementation improves mean RankIC by 0.000642 over its numerical expert and by 0.000701 over matched shuffled text, with positive differences in all five seeds. Controlled text perturbations support functional reliance on text correspondence, while incumbent retention prevents regressions in the measured development criterion. Together, the experiments characterize ASE at two levels: forecasting performance of task-specific implementations and the conditional contributions of individual modules.

Figures & tables

Figure 1: The iterative feedback implementation of ASE. C1 proposes and compiles event/state definitions into an accumulated library. C2 searches numerical–textual subsets by refitting the same predictor, then returns conditional contribution and redundancy diagnostics to the next C1 generation. Libraries are fixed within each search round; the final selected predictor is frozen for assessment. Event and state names in the figure are illustrative.

Method	FNSPID	Electricity	Environment	Bitcoin
Persistence	0.001551	0.099490	0.518086	2.7678
Ridge	0.001612	0.029461	0.363842	4.0281
NLinear	0.001512	0.028769	0.405229	2.7600
DLinear	0.002096	0.028570	0.361886	3.7084
GRU	0.001615	0.029219	0.372453	4.0124
iTransformer	0.001725	0.025178	0.373132	3.7583

Table 2

Figure 2: MSE reductions of task-specific ASE implementations relative to published TESS scores, computed from the displayed main-table values. The complete twelve-method comparison is retained in the table. These are system-level point comparisons; baseline implementations and selection protocols differ as specified in the setup.

Comparison	ASE RankIC	Reference RankIC	ASE gain
ASE vs numerical expert	0.054989	0.054346	+0.000642
ASE vs matched shuffled text	0.054989	0.054287	+0.000701

Table 4

Figure 3: CSI 300 text-fusion evidence. Left: paired seed gains over the numerical expert and matched shuffled text. Right: mean gains and paired 20-day block-bootstrap 95% intervals. The numerical-reference interval includes zero; the matched-shuffle interval is positive under this historical protocol.

Time Series Forecasting

Time series forecasting predicts future values from ordered observations using statistical, neural, and probabilistic models to support decisions under uncertainty.

See all 861 papers

From Text to Forecasts: Bridging Modality Gap with Temporal Evolution Semantic Space
Mar 13, 2026Lehui Li, Yuyao Wang, Jisheng Yan +5Non-Stationary Time Series Forecasting LLM-Based Time Series Forecasting
School of Software, Shandong University, China · Boston University, Boston, MA, USA · State Key Laboratory of Complex & Critical Software Environment, Beihang University +1
ReasonCast: Agentic Demand Forecasting with Selective Semantic Reasoning
Aug 15, 2026Ziyue Yang, Chaolin Xu, Yijing Wang +5Time Series Forecasting Time Series Reasoning
Taobao and Tmall Group, Alibaba Group China
From Affect Prediction to Affect Forecasting: Evidence for Distinct Information Sources in Longitudinal Text
Jun 15, 2026Sadia Noor, Seemab Latif, Raja Khurram Shahzad +1Time Series Forecasting Valence-Arousal Modeling
School of Electrical Engineering and Computer Science (SEECS), National University of Sciences and Technology (NUST), Islamabad, Pakistan · Department of Communication, Quality Management and Information Systems, Mid Sweden University, Ostersund, Sweden

Mar 13, 2026J/K move · Enter open · S save

From Text to Forecasts: Bridging Modality Gap with Temporal Evolution Semantic Space

Lehui Li, Yuyao Wang, Jisheng Yan, Wei Zhang+4

School of Software, Shandong University, China · Boston University, Boston, MA, USA · State Key Laboratory of Complex & Critical Software Environment, Beihang University+1

Non-Stationary Time Series Forecasting LLM-Based Time Series Forecasting Multimodal Fusion Multivariate Time Series Forecasting

Open paper AskPDF arXiv

Incorporating textual information into time-series forecasting holds promise for addressing event-driven non-stationarity; however, a fundamental modality gap hinders effective fusion: textual descriptions express temporal impacts implicitly and qualitatively, whereas forecasting models rely on explicit and quantitative signals. Through controlled semi-synthetic experiments, we show that existing methods over-attend to redundant tokens and struggle to reliably translate textual semantics into usable numerical cues. To bridge this gap, we propose TESS, which introduces a Temporal Evolution Semantic Space as an intermediate bottleneck between modalities. This space consists of interpretable, numerically grounded temporal primitives (mean shift, volatility, shape, and lag) extracted from text by an LLM via structured prompting and filtered through confidence-aware gating. Experiments on four real-world datasets demonstrate up to a 29 percent reduction in forecasting error compared to state-of-the-art unimodal and multimodal baselines. The code will be released after acceptance.

Abstract

Time Series Forecasting

Explore similar work

From Text to Forecasts: Bridging Modality Gap with Temporal Evolution Semantic Space

ReasonCast: Agentic Demand Forecasting with Selective Semantic Reasoning

From Affect Prediction to Affect Forecasting: Evidence for Distinct Information Sources in Longitudinal Text