Bayesian Neural Networkによる景気テキストの不確実性評価と景気指標の開発

書誌事項

タイトル別名
  • Bayesian Neural Network for Evaluate Uncertainty of Economic Text and Development of Economic Indicators

この論文をさがす

抄録

近年,ニュースやソーシャルメディア等のテキストを深層学習モデルで解析して,人々の景況感を予測しようという試みが行われている.既存の研究では景気が良い/悪いといった景気の方向性に着目しているが,景気とは本来,不確かなものであり,テキストの書き手の景況感も中心の周りに広がった分布で評価すべきと考えられる.実際,テキストには景気に良い面と悪い面が併記されたり,先行きの不確実性を強調したりする等,人間が読んでも書き手の景況感が単純に判断できないものも多い.このような書き手の景況感が不確実なテキストを除くことで,より確信的な景況感を持った意見だけを集めて景況感を評価することが可能となり,マクロな景況感の推定精度向上が期待される.本研究ではBayesian Neural Network(BNN)を用いることで景気テキストの不確実性を評価し,それを使って景況感評価の精度を高めることを提案する.実験ではまずBNNによって景気センチメント推定の精度が向上することを示した.次に,不確実性の高いサンプルを除いてセンチメントを集計することで,より精度の高い景気指標となることが確認できた.最後に,BNNで不確実性が高まるテキストの特徴やその時期について考察した.その結果,得られた不確実性と経済の不確実性指標として用いられているEconomic Policy Uncertainty(EPU)指数との統計的に有意な相関を確認できた.

In recent years, there have been attempts to use deep learning models to analyze text from news and social media to predict trends in the economy, stock prices, and so on. Existing research has focused on the direction of the economy, such as good/bad business conditions. However, there are many references to uncertainty such as “uncertain” and “uncertain” in texts about the economy. However, the business climate is inherently uncertain, and it is thought that the business condition of text writers should be evaluated in terms of a distribution. In fact, there are many texts that also human reader cannot judge writer's business sentiments because texts describe both positive and negative aspects of the economy or emphasize uncertainty about the future. By excluding such texts in which the writer's business sentiment is uncertain, it is possible to collect only opinions with a more confident business sentiment. This is expected to improve the accuracy of estimating macro business sentiment. In this study, we propose to use a Bayesian Neural Network (BNN) to evaluate the uncertainty of texts that mention the business condition and thereby improve the accuracy of business climate assessment. In experiments, we first show that the BNN improves the accuracy of business sentiment estimation. Second, we confirmed that excluding the highly uncertain samples from the aggregate yields a more accurate business cycle indicator. Finally, we discussed the characteristics of texts in which uncertainty increases in BNNs and the timing of such increases. We confirmed a statistically significant correlation between the obtained uncertainty and the Economic Policy Uncertainty (EPU) index, which is used as a measure of economic uncertainty.

収録刊行物

詳細情報 詳細情報について

問題の指摘

ページトップへ