📈 Time Series

💬 ACL2026 · 5 paper notes

A Unified Framework for Modeling Heterogeneous Financial Data via Dual-Granularity Prompting

This paper proposes FinLangNet, a dual-module framework comprising DeepFM for static feature processing and a Transformer with a dual-granularity prompting mechanism for sequential behavior modeling, enabling multi-scale credit risk prediction. Upon deployment on the Didi Finance platform, the system achieves a 6.3 pp improvement in KS and a 9.9% reduction in bad debt rate.

Learning Uncertainty from Sequential Internal Dispersion in Large Language Models

This paper proposes the SIVR framework, which computes internal variance statistics (generalised variance, circular variance, and token entropy) across layers of LLM hidden states as token-level features, and aggregates full-sequence patterns via a lightweight Transformer encoder to estimate uncertainty and detect hallucinations, achieving significant improvements over baselines with stronger generalization.

STK-Adapter: Incorporating Evolving Graph and Event Chain for Temporal Knowledge Graph Extrapolation

This paper proposes STK-Adapter, which embeds three MoE modules at every layer of an LLM—ST-MoE for capturing spatiotemporal structure, EA-MoE for modeling event chain semantics, and CMA-MoE for deep cross-modal alignment—to address the spatiotemporal information loss and layer-wise dilution caused by shallow alignment between TKG embeddings and LLMs in existing methods, achieving significant improvements over SOTA on four benchmark datasets.

Temporal Leakage in Search-Engine Date-Filtered Web Retrieval: A Retrospective Forecasting Case Study

This paper presents a systematic audit of date filters in Google and DuckDuckGo, revealing that search engine date filtering critically fails in retrospective forecasting evaluation — 71% (Google) and 81% (DuckDuckGo) of questions have at least one page containing significant post-cutoff information leakage, causing the prediction Brier score to drop artificially from 0.24 to 0.10.

Time-RA: Towards Time Series Reasoning for Anomaly Diagnosis with LLM Feedback

This paper introduces Time-RA, a new task that upgrades time series anomaly detection from binary classification to generative reasoning diagnosis (detection + classification + causal explanation). It constructs RATs40K, the first multimodal benchmark comprising ~40K samples across 10 domains and 20 anomaly types, and validates the feasibility of this paradigm through an AI feedback annotation pipeline and LLM fine-tuning.