特徴品質適応注意Transformer：EU ETS炭素価格の多期間予測のためのTransformerベースモデル
Feature Quality Adaptive Attention Transformer: A Transformer-Based Model for Multi-Horizon EU ETS Carbon-Price Forecasting (原題)
Guo Rui
🤖 gxceed AI 要約
日本語
EU ETS炭素価格を5・10・21営業日先まで予測するTransformerモデルFQA-ATTを提案。相互情報量とランダムフォレスト重要度を組み合わせた特徴品質ゲートと品質適応注意機構で、情報漏洩を制御した検証手順を構築。10モデル中で検証MAEが最小、短期ではXGBoost等より有意に高精度だが、持続性ベンチマークとの有意差は限定的。
English
Proposes FQA-ATT, a Transformer with a feature-quality gate (mutual information + random forest importance) and quality-adaptive attention for 5/10/21-day EU ETS carbon price forecasting. A leakage-controlled protocol yields the lowest validation MAE among ten models; on 2024-2025 test data it ranks 2nd-3rd. Significantly beats XGBoost/LightGBM/PatchTST at short horizons but rarely differs from the persistence benchmark.
Unofficial AI-generated summary based on the public title and abstract. Not an official translation.
📝 gxceed 編集解説 — Why this matters
日本のGX文脈において
日本ではGXリーグやカーボンクレジット市場、SSBJ開示と連動する炭素価格リスク評価に応用可能。EU ETSの予測手法は日本の炭素価格制度設計や企業の内部炭素価格設定の高度化に示唆を与える。
In the global GX context
Adds to the growing ML-for-carbon-markets literature relevant to TCFD/ISSB climate risk assessment and transition finance. The leakage-controlled protocol offers a methodological benchmark for carbon price forecasting that can inform internal carbon pricing and policy evaluation globally.
👥 読者別の含意
🔬研究者:炭素価格予測におけるTransformerの有効性と持続性ベンチマークの重要性を示す。
🏢実務担当者:内部炭素価格設定やEU ETSエクスポージャー管理の予測精度向上に活用可能。
🏛政策担当者:炭素市場の価格予測手法は政策評価や市場安定性モニタリングに参考となる。
📄 Abstract（原文）
Forecasting the carbon price of the European Union Emissions Trading System (EU ETS) supports energy planning, risk assessment, and climate-policy evaluation. The price is non-stationary and responds to several energy markets, so medium-horizon forecasts are difficult to evaluate without information leakage. This protocol describes a leakage-controlled workflow and the Feature Quality Adaptive Attention Transformer (FQA-ATT) for forecasting at 5, 10, and 21 trading days. A Feature Quality Gate scales each input by a prior that combines mutual information (MI) and random forest (RF) importance computed on training data. A Quality-Adaptive Attention module adds a learnable per-head temporal decay and a quality-energy bias to scaled dot-product attention. Depth-wise convolutions with 1-, 3-, and 5-day kernels extract local patterns. The trading-day panel covers November 2017 to 29 December 2025 and contains 41 lagged features. Each input window ends at the forecast origin, and every target lies after it. Over five random seeds, FQA-ATT gave the lowest validation mean absolute error (MAE) of ten models at all three horizons. On the 2024–2025 test period, its MAE was 2.37, 3.19, and 4.79 EUR, which ranked second, third, and second. FQA-ATT was significantly more accurate than XGBoost, LightGBM, and PatchTST at short horizons. FQA-ATT and a linear FITS-style model were the only learned models whose accuracy never differed significantly from the persistence benchmark. These results show that the protocol gives realistic accuracy estimates and that the feature-quality prior helps attention-based models remain competitive with persistence on carbon returns.
🔗 Provenance — このレコードを発見したソース
- openalex https://doi.org/10.5281/zenodo.23231191first seen 2026-10-10 05:17:12
🔔 こうした論文の新着を逃したくない方は キーワードアラート に登録（無料・3キーワードまで）。
gxceed は公開メタデータに基づく研究支援データセットです。要約・翻訳・解説は AI 支援で生成されています。 最終的な解釈・検証は利用者が原典資料に基づいて行うことを前提とします。