Currency Exchange Rate Prediction with Long Short Term Memory Networks based on Attention and News Sentiment Analysis

Please use this identifier to cite or link to this item: http://ir.lib.ncu.edu.tw/handle/987654321/81517

Title:	Currency Exchange Rate Prediction with Long Short Term Memory Networks based on Attention and News Sentiment Analysis
Authors:	李謦伊;李
Contributors:	數學系
Keywords:	預測匯率;Currency Exchange Rate Prediction;Long Short-Term Memory based on Attention;News Sentiment Analysis
Date:	2019-08-16
Issue Date:	2019-09-03 15:58:41 (UTC+8)
Publisher:	國立中央大學
Abstract:	匯率預測是典型的時間序列預測問題，其通過使用歷史數據由自回歸整合移動平均（ARIMA）、季節性ARIMA以及人工神經網絡（ANN）與長短期記憶（LSTM）進行建模。在本研究中，我們將討論LSTM變體，LSTM attention的匯率預測表現。我們將實驗分為三個部分：預測貨幣匯率、預測貨幣匯率的趨勢、嘗試預測更長的一天的匯率。為了更好地預測澳幣兌美元的匯率，我們進一步增加了基於SnowNLP庫的新聞文章的情緒分析以及提及澳幣上漲的新聞文章的簡單關鍵詞匹配。我們還比較了使用不同維度的輸入特徵的預測表現，例如7天、30天和60天以及不同的輸入特徵的組合，例如歷史數據、差異和比率。此外，我們使用兩種不同的策略來預測更遠的未來並比較效能。我們的實驗結果表明，將新聞文章的情緒分析添加為特徵可以降低至少15％的預測誤差。與僅使用歷史數據的效能相比，使用匯率及匯率比率的預測效能降低了12％的預測誤差，但是使用匯率及匯率差分並沒有對預測匯率有所貢獻。以及使用7天輸入的預測效能優於其他天的輸入的預測效能。最後，我們比較了不同方法的匯率預測效能，LSTM attention加入新聞的情感分析的效能優於其他方法。;Exchange rate prediction is a typical time series prediction problem which has been modeled by Autoregressive integrated moving average (ARIMA), Seasonal ARIMA as well as Artificial neural networks (ANN) such as Long Short-Term Memory (LSTM) using historical data. In this study, we will discuss the exchange rate prediction of the variant of LSTM, LSTM based on attention. We divide our experiments into three parts, predict currency exchange rates, predict the trends of currency exchange rate, and try to predict exchange rates for a longer day. To better predict the exchange rate of the Australian dollar against the US dollar, we have further added the sentiment analysis of the news articles based on SnowNLP library as well as simple keyword matching on news articles that mention the rise in the Australian dollar. We also compare the performance of using different sizes of input features, such as 7 days, 30 days and 60 days, as well as the different combinations of features, such as historical data, differences and ratios. In addition, we use two different strategies to predict a farther future and compare the performance. Our experimental results show that adding sentiment analysis of the news articles as features can reduce prediction error by at least 15\%. The exchange rate prediction performance of using rate ratio is reduced by the test error of 12\%, compared to the performance using only historical data, but the use of the difference as our feature does not contribute to the prediction. The performance of using 7-day input is superior to the other inputs. Finally, we compared the exchange rate prediction performance of different methods, LSTM based on attention with news sentiment analysis outperforms other methods.
Appears in Collections:	[Graduate Institute of Mathematics] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	250	View/Open

社群 sharing

Loading...