跳转至

文章背景与核心概要

由于长程依赖、局部波动爆发、结构性转变以及非线性振荡行为的存在，非平稳时间序列的预测一直是一项艰巨的挑战。传统的基于 Transformer 的模型在面对突发的机制变化时往往表现不佳，因为它们的前馈神经网络层依赖于平滑且静态的激活函数。

为了克服这一局限性，作者引入了 QFCQT（Quantum-Fractal-inspired Chaotically Gated Quantformer，即受量子分形启发的混沌门控量化变压器）。QFCQT 专为在复杂波动动态下进行稳健预测而设计，它结合了数值编码器、可学习的李-振荡器（Lee-oscillator）激活模块以及平滑-混沌门控融合机制。基准测试表明，QFCQT 在性能上始终优于现有的各种主流预测模型。

QFCQT: A Chaotically Gated Quantformer Framework for Volatile Time-Series Forecasting

arXiv: 2608.07363 [cs.AI]
Submitted: August 7, 2026
Authors: Junkai Lin, Siqi Hou, Raymond Lee

📋 执行摘要

Forecasting non-stationary time series remains a formidable challenge due to long-range dependencies, local volatility bursts, structural shifts, and nonlinear oscillatory behaviors. Traditional Transformer-based models struggle with abrupt regime changes because their feed-forward blocks rely on smooth, static activations.

To overcome this limitation, the authors introduce QFCQT (Quantum-Fractal-inspired Chaotically Gated Quantformer). Designed for robust forecasting under complex volatile dynamics, QFCQT utilizes a combination of numerical encoding, a learnable Lee-oscillator activation module, and a smooth-chaotic gated fusion mechanism. Benchmarking tests demonstrate that QFCQT consistently outperforms established forecasting models.

🧠 核心架构与方法论

QFCQT 由三个主要结构组件组成：

  1. Quantformer 风格的数值编码器： 通过线性嵌入直接处理多变量时间序列输入，在时间维度上保持结构完整性。
  2. 可学习的李-振荡器激活模块： 将标量预激活映射到动态振荡响应中，并通过随时间的最大值池化（Max-over-Time pooling）对其进行高效汇总。该框架没有依赖单一的固定振荡器，而是采用八个参数化李-振荡器家族的软叠加（soft superposition），以捕获不同机制下的多样化非线性响应模式。
  3. 平滑-混沌门控融合机制： 自适应地平衡传统的平滑激活与对混沌敏感的响应，以处理市场波动中的突发转变。

注：“受量子分形启发”一词指的是基于软振荡器叠加和多尺度非线性响应的计算类比，而非形式化的量子力学或分形理论推导。

QFCQT consists of three main structural components:

  1. Quantformer-Style Numerical Encoder: Directly processes multivariate time-series inputs via linear embedding, maintaining structural integrity across temporal dimensions.
  2. Learnable Lee-Oscillator Activation Module: Maps scalar pre-activations to dynamic oscillatory responses, summarizing them efficiently through Max-over-Time pooling. Rather than relying on a single fixed oscillator, the framework uses a soft superposition of eight parameterized Lee oscillator families to capture diverse nonlinear response patterns across different regimes.
  3. Smooth-Chaotic Gated Fusion Mechanism: Adaptively balances traditional smooth activations and chaos-sensitive responses to handle sudden shifts in market volatility.

Note: The term "quantum-fractal-inspired" refers to a computational analogy based on soft oscillator superposition and multi-scale nonlinear responses, rather than a formal quantum-mechanical or fractal-theoretic derivation.

📊 实验结果

在多个标准的波动性基准数据集上进行了广泛的评估： * ETTh1 与 ETTh2（电力变压器温度数据集） * A股股指基准

性能表现： QFCQT 的表现始终优于强大的基线模型，包括： * Informer * LogTrans * LSTMa * HAT * COTN

Extensive evaluations were conducted across multiple standard volatile benchmarks: * ETTh1 & ETTh2 (Electricity Transformer Temperature datasets) * A-share Stock Index benchmarks

Performance: QFCQT consistently outperformed strong baseline models, including: * Informer * LogTrans * LSTMa * HAT * COTN

🔗 快速链接与资源