産業間の取引構造を用いた深層学習モデルによる生産指数の予測

COVID-19によって引き起こされた生産の混乱によってサプライチェーンの脆弱性が顕在化した．特に半導体のように，様々な産業に供給される製品の供給不足は産業間取引を通じて多くの産業に波及している．もし，足元の各産業の生産状況から，このような波及効果を考慮して，将来の各産業の生産を予測することができれば経済の混乱を緩和することが可能になると期待される．産業間の取引関係や波及効果のデータとして産業連関表がある．産業連関表を用いた従来の予測は静的な波及効果の予測が多く，時系列モデルと組み合わせたモデル化は限られたものであった．本稿では産業連関表をグラフと見做してGNNで解析し，さらにRNNを組み合わせた手法を用いることで産業間取引を特徴量に用いた生産指数の予測モデルを開発した．得られたモデルはCOVID-19後の生産が混乱した時期において統計的に有意な予測精度の改善が見られた．特に，自動車のように仕入先の多い産業ほど取引関係を考慮したモデルによる改善が大きく，これは定性的な理解と一致する．

The production disruption caused by COVID-19 has made apparent the vulnerability of the supply chain. In particular, shortages in the supply of products supplied to various industries, such as semiconductors, are propagated to many industries through inter-industry transactions. If it is possible to predict the future production of each industry from the current production status by taking into account such propagation, it would be possible to alleviate the economic turmoil. The Input-Output table provides data on trade relations and propagation effects among industries. Conventional forecasting based on the I-O table has been limited to static propagation effects, and modeling combined with time series models has been limited. In this paper, we developed a forecasting model of Indices of Industrial Production using inter-industry transactions from the input-output table. Specifically, we used GNN to analyze the input-output table as a graph, and then combined it with RNN. The obtained model showed statistically significant improvement in forecasting accuracy in the period of production disruption under COVID-19 occurred. Furthermore, the improvement in accuracy due to our model considering the input-output table is greater for industries with many suppliers, such as the automobile industry. This is consistent with qualitative understanding.