Research on time series information prediction using deep learning

2023 Fiscal Year Final Research Report

• Project/Area Number: 19K01591
• Research Category: Grant-in-Aid for Scientific Research (C)
• Allocation Type: Multi-year Fund
• Section: 一般
• Review Section: Basic Section 07030:Economic statistics-related
• Research Institution: University of Shizuoka
• Principal Investigator: Jun Rokui 静岡県立大学, 経営情報学部, 教授 (70362910)
• Project Period (FY): 2019-04-01 – 2024-03-31
• Keywords: LSTM / QRNN / GRU / RNN / LSTNet / GAN / ESN / グレンジャー因果性

Outline of Final Research Achievements

In this research, we proposed several new time series prediction methods which analyze the relationship of multiple time series and converge the prediction results using the regression model. By predicting the chronological relationship using the Granger causal test, which was awarded the Nobel Prize in 2003, we succeeded in extracting only the time series with the relationship from multiple time series selected at random. In the time series prediction, the prediction accuracy of the short-term prediction tends to be relatively good, but in the long-term prediction, the prediction error becomes large. For this problem, we succeeded in drastically reducing the error of the long-term prediction by extracting the noticeable interval with large variation.

Free Research Field

機械学習

Academic Significance and Societal Importance of the Research Achievements

本研究では、各起因子を取りまとめる機構と複数のLSTMを並列に動作させる機構を組み合わせることで、蜘蛛の糸のように複雑に収束する単一系列を可能な限り少ない誤差で予測する新たなRNNの枠組みを提案する．本研究はニューラルネットワークを基礎とする深層学習と経済学、社会学を融合させた領域横断的位置づけの研究である．工学的・統計的視点だけでなく、入出力データに対する知見も必要となるため、高い学術的独自性と創造性を有する．本研究の成果により、幅広い分野の時系列変化を実時間の範囲内で高精度に推定することができる．結果、株価や人口増減などあらゆる社会問題への解決方法として役立つものと確信する．