Neural Networks for Time-Series prediction : Stock Exchange Forecasting

• Item type: National Convention(1)
• 公開日: 1997-03-12
• タイトル / 言語: en
• タイトル: Neural Networks for Time-Series prediction : Stock Exchange Forecasting
• 言語: eng
• 資源タイプ / 資源タイプ識別子: http://purl.org/coar/resource_type/c_5794
• 資源タイプ: conference paper
• 著者所属(英): en
• 著者所属(英): Lab. Prof. T.Okamoto, Artificial Intelligence (AI) Info-Systems Graduated School(IS) University of Electro-Communication
• 著者所属(英): en
• 著者所属(英): Lab. Prof. T.Okamoto, Artificial Intelligence (AI) Info-Systems Graduated School(IS) University of Electro-Communication
• 論文抄録(英) / 内容記述タイプ: Other
• 論文抄録(英) / 内容記述: This study attempts to simplify the complexity of the systems present in the economical world and to create a handy tool for economists, that usually deal with little mathematics (mostly linear functions, that can hardly be a correct approximation of real world behaviour) and a lot of "common sense". It is the authors belief that this "common sense" can be structured, organized and made reproducible by a computer. Most of the economical functions can be represented as Time Series (TS). Stock Exchange (SE) events are representable as TS (see [Ankenbrand et al., 1995]), which can be forecasted with certain accuracy with Neural Net-works(NN) (see [Gas et al., 1993]). In [Cristea et al., 1996] a forecasting theory based on NNs was developed, for general TS as well as for the particular case of Stock Exchange (SE), examining the components: Trend, Cyclus, Season and Irregular events. NNs were used for their parallel processing power. and ability to learn three of the four components mentioned above. For the forth, an economy based calculus was proposed. This theory was implemented here with help of a program with Motif interface and the results were studied. In [Komo et al., 1994] a similar forecasting tool was developed. Stock Market values are analized, divided into a training set and a test set, and presented to a NN tool. Differences exist though in the data-window size, both for inputs and outputs. While [Komo et al., 1994] works with a 1-prediction-step, the current study aims at a prediction for a variable number of future values. Also, input data is not only from the previous step, but considers the history of the TS. Therefore, data input has a three level hierarchy : stock market prices for training, a previous weights matrix for new data (testing) and current stock market prices, for active forecasting (prediction). The learning method in [Komo et al., 1994] is based on sequential data input ("epoch") while the current study proposes the more efficient bootstrapping method. Some partial results that are available look promising, although the study is not yet completed. The main difference from previous systems consists in the usage of both mathematical and economical rules for the system construction.
• 書誌レコードID / 収録物識別子タイプ: NCID
• 書誌レコードID / 収録物識別子: AN00349328
• 書誌情報: 全国大会講演論文集; 巻 第54回, 号 人工知能と認知科学, p. 309-310, 発行日 1997-03-12
• 出版者 / 言語: ja
• 出版者: 情報処理学会