Paper Reading | Hermes: Integrating Multi-Scale Spatio-Temporal Hypergraphs with Industry-Leading-Lagging Relationships for Stock Time-Series Forecasting

Paper link: arxiv.org/pdf/2509.2366... Stock time-series forecasting provides critical decision support for investors, regulators, and analysts in the financial domain. Unlike multivariate time series in other domains, stock time series exhibit pronounced industry-level correlations; uncovering such correlations can significantly improve prediction accuracy. However, existing hypergraph-based methods only superficially capture industry associations and suffer from two key limitations: (1) insufficient modeling of leading-lagging interactions among industries, and (2) failure to model multi-scale information both within and across industries. To address these issues, this paper proposes the Hermes framework, which integrates a moving aggregation module and a multi-scale fusion module into a hypergraph neural network. Specifically, Hermes designs an edge-based moving aggregation module—combining sliding windows and dynamic temporal aggregation operations—to capture inter-industry leading-lagging dependencies; and a multi-scale fusion module—enabling edge-to-edge message passing across scales—to fuse multi-scale information while preserving consistency across scales. Experiments on multiple real-world stock datasets demonstrate that Hermes outperforms existing state-of-the-art (SOTA) methods in both efficiency and accuracy. 1. Background Stock time series typically comprise key indicators (e.g., opening price, closing price) of multiple stocks, and their accurate forecasting is vital for investment decisions. Yet, existing studies often treat stocks as independent entities, overlooking the intrinsic interdependencies embedded in market dynamics. In reality, rich industry-level associations exist among stocks: stocks within the same industry are jointly influenced by common macroeconomic conditions and policy factors (e.g., synchronized price movements of Apple and Microsoft in the technology sector); and cross-industry causal and temporal dependencies also exist (e.g., technological breakthroughs in the tech sector driving growth in the energy sector, forming leading-lagging relationships). Although conventional hypergraph methods can capture intra-industry associations via hyperedges (connecting multiple stocks within one industry), they face two major challenges: • Insufficient modeling of inter-industry leading-lagging interactions: Financial markets feature complex causal and temporal dependencies across industries (e.g., technology sector developments driving the energy sector), which current methods fail to adequately model dynamically. • Lack of multi-scale information modeling: Stock time series exhibit multi-scale characteristics (e.g., coarse-grained overall trends versus fine-grained cyclical patterns); single-scale analysis cannot capture scale-dependent association differences (e.g., two industries may be positively correlated at a coarse scale but negatively correlated at a fine scale). 2. Problem Definition This paper aims to address two core problems in stock time-series forecasting: 1. Deep modeling of inter-industry leading-lagging relationships: How to capture dynamic temporal dependencies (e.g., changes in one industry preceding or lagging behind another) within a hypergraph framework. 2. Effective fusion of multi-scale information: How to decompose and integrate industry-level associations across different time scales (coarse- vs. fine-grained), while maintaining consistency across scales. 3. Methodology The Hermes framework comprises five core components: multi-scale feature extraction, adaptive spatio-temporal hypergraph construction, hyperedge moving aggregation module, hyperedge multi-scale fusion module, predictor, and optimization objective. 3.1 Multi-Scale Feature Extraction • Multi-scale decomposition: Apply 1D convolutional layers of varying kernel sizes to downsample the raw time series X, generating S distinct-scale time series. • Causal mixing: Apply MLP layers along the temporal dimension to ensure each time step depends only on historical information (causality), yielding enhanced representations.

Moore Threads Buy-Side Investment Analysis Report

Witness the journey of a middle-aged man turning his Range Rover into a Volkswagen, his black hair into white, and his seafood dinners into pickled vegetables—all courtesy of the stock market! (There’s a bonus at the end—you know what we mean.) ************************************************** Moore Threads (688795.SH) Buy-Side Investment Analysis Report Report Date: September 7, 2026 | Market Data As Of: Close on September 7, 2026 | Data Sources: IPO Prospectus / 2026 Semi-Annual Report / Brokerage Research Reports / IDC Module I: Executive Summary (Conclusions First) 【Investment Rating】Avoid (Do Not Buy, Do Not Participate) 【Core Conclusion】This is a textbook case of a "good company, bad price": its technical roadmap is scarce, its founder’s capabilities are top-tier, and its financial inflection point is genuine—but its current market capitalization of ~¥195.3 billion implies a 2025 revenue of only ~¥1.5 billion (P/S ratio ~130x), leaving no margin of safety. Compounded by an impending flood of share unlocks and dilution from a potential Hong Kong IPO, buying now amounts to carrying the IPO pre-investment shareholders’ burden. 【Fair Valuation Range】Based on brokerage consensus 2026 revenue forecast of ¥3.46 billion and a P/S multiple of 40–50x (midpoint between Hygon’s 36x and Cambricon’s 88x), the fair market cap is ~¥138–173 billion, translating to a fair share price of ~¥295–368. At the current price of ¥415.48, shares trade ~13% above the upper bound of this fair range. 【Core Investment Thesis】In one sentence: Domestic GPU substitution is among the most certain, decade-long structural trends—and Moore Threads is China’s sole full-featured GPU supplier. But being right about the sector does not guarantee being right about the stock. It has yet to build a "real moat" (CUDA’s ecosystem remains insurmountable), remains unprofitable (cumulative losses ~¥5.9 billion), exhibits extreme customer concentration (91%), and faces relentless selling pressure from IPO-pre-investment shareholders over the next 12 months. 【Key Catalysts】① Accelerated domestic substitution and ramp-up of large AI-computing orders; ② Quarterly profitability achieved in 2027; ③ Deployment of a 100,000-GPU cluster. 【Major Risks】① Share unlock-related selling (today’s 5.48% unlock could trigger a 20% daily limit-down); ② Dilution from potential Hong Kong IPO; ③ Competition (Cambricon is already profitable; Huawei Ascend dominates the training segment); ④ Extreme customer concentration; ⑤ Persistently negative operating cash flow. 【Expected Return & Time Horizon】Even under the most optimistic 2027 brokerage forecast—¥501 million net profit, valued at 100x P/E—the implied market cap would be only ~¥50 billion, starkly divergent from the current ¥195.3 billion valuation. Medium-term downside risk significantly outweighs upside potential. Module II: Company Overview (Understanding the Investment Target) I. Core Business & Positioning Moore Threads Intelligent Technology (Beijing) Co., Ltd., founded in October 2020, is a fabless GPU (graphics processing unit) and related product R&D, design, and sales company. Centered on its self-developed full-featured GPU, the company launched MUSA (Meta-Unified System Architecture)—a unified platform integrating scientific computing, 3D graphics rendering, AI training & inference, high-performance computing, and physics simulation. It is among the very few Chinese enterprises capable of developing full-featured GPUs, thereby filling a critical technological gap in China’s high-end GPU landscape.

Methanex Corporation (MEOH) Stock Forecasts

Jul 29, 2026

Argus Research lowered its price target for Methanex Corporation (MEOH) to $56.00, noting the company produces methanol feedstock for adhesives and solvents. The report also covers adjusted targets for peer chemical firms like Eastman Chemical and CF Industries.

Yahoo! Finance Canada

2026-07-01

RBC Cuts Price Target on Eastman Chemical to $69 From $82, Keeps Sector Perform Rating

2026-07-28

EMN: Recycling And Cost Discipline Will Support Resilient Earnings Through Commodity Headwinds

2026-07-29

Methanex Corporation (MEOH) Stock Forecasts

2026-07-29

Materion Corporation (MTRN) Stock Forecasts

2026-08-08

Royal Bank of Canada Increases Position in Eastman Chemical Company $EMN