Bitcoin is a crypto currency that is a peer-to-peer(P2P) network systems based on distributed ledger technology and is being used as an alternative payment system. Reliability and safety are very important aspects of payment system. However, in recent years, with an increase in value, crypto currency becomes a target of a malicious users and the attacks that strike the vulnerability of the system are regarded as a problem. Such a problem significantly reduces reliability and safety as a payment system for crypto currency. Therefore, it is necessary to pay attention to cyber security risks inherent in the system, as well as price uctuations usually focused on financial asset prices. In this research, we propose to use information observed in darknet as alternative data. It is useful in evaluating the risk in the crypto currency market. The darknet is a name of an IP address space unallocated by terminals or the like among spaces that can be assigned IP addresses. The darknet is mainly used for observing signs of security incidents. This is useful for investors to grasp potential risks of crypto currency markets and is important for service providers to explain security risks and measures. In addition, the darknet observation information has a prospect of utilizing not only the crypto currency but also the monitoring of the security risk of companys.

In this research, we use English-language news related to the Bitcoin posted on the Internet as a data source, and use LDA(Latent Dirichlet Allocation) which is one of probabilistic topic models, for each article. We judged what kind of topics (keyword group) the sentence is composed of, and quantitatively expressed the excitement of topics in the period using different topic distribution for each article obtained. Furthermore, by analyzing the relationship with the bitcoin's market price on the Internet, we try to evaluate the influence of the news. We show to the relation between the quantity concerning the bitcoin's price (BTC/JPY) in the target period and the excitement of the topic in the news article related to bitcoin.

IPO 及び New Stage（市場変更）時に放映される StockVoice TV 出演社の社長や他出演者の表情・声の特徴量・テキストを基にした感情特徴量が株価に与える影響について、機械学習を通じた検証を行った。価格データについては、株式会社 Quick 様から提供して頂いたデータを利用し、検証対象として番組全体の感情特徴量と翌日の終値、及び番組中の一分ごとの感情特徴量と次点の一分足の終値の学習を行った。本研究では、目・耳・言葉の三系統のマルチモーダルな特徴抽出を行う事で予測の精度を上げることが出来るのではないか、という仮説のもと検証を行った。日足に関しては、SVM・ロジスティック回帰に特徴量を入力し二値検証を行い、１分足は５秒ごとの特徴量を入力とし、XGBoost 及び DNN を用いた３クラス分類による検証を行った。

The European Central Bank (ECB) decides the euro area's monetary policy at the monetary policy meeting of the Government Council. After the ECB's monetary policy meeting, the ECB president Mario Draghi and the vice-president Vitor Constancio hold a press conference to explain the monetary policy management. In this research, using facial expression recognition algorithms based on deep learning, we analyzed the presidents' facial expression in the press conference and estimated the emotional indexes such as "Happiness", "Anger", "Sadness" and "Surprise". As a result, we found that the president Draghi's index of "Happiness" decreased and the index of "Sadness" increased just before making major policy changes in the phase of ECB policy normalization. In addition, the vice-president Vitor Constancio's index of "Happiness" tends to change in the opposite direction to Draghi's index of "Happiness" regardless of the policy change. We believe that the inverse correlation may have the adjustment effect that the impression of the whole press conference will be neutral.

We analyze the impact of advanced machine learning methods on the performance and risk characteristics of popular Japan equity strategies over the last 10 years. We then propose a possible approach to enhancing each strategy through advanced risk control and we analyze the results.

市場の「流動性」に関心が高まっている．流動性は金融市場の盛況を表す目安とされ，「取引のしやすさ」ともいうことができる．実証研究では，それぞれの研究目的に沿うような流動性指標を用いて，その有用性を議論していた．しかし，それらの指標が市場内外のどの要因の影響を受けて変化するのかは明らかにされていない．そこで本研究では，市場内のどの要素が，流動性指標に影響を与えるのかを人工市場を用いて調査した．その結果，4 つの流動性指標（Volume，Tightness，Resiliency，Depth）は，人工市場のパラメータのうち，1）ティックサイズ，2）投資家の注文戦略を決める成分（ファンダメンタル成分，テクニカル成分，ノイズ成分）から影響を受ける可能性があることが分かった．

有価証券報告書には，業績の他，リスク対策や企業の施策等，決算短信にはない情報の記載もある．また，先行研究では，多くの情報から重要な文を効率よく抽出する方法として，因果関係文を重要文とした手法が提案されている．しかし，抽出対象を決算短信等とした報告はあるが有価証券報告書とした報告はない．そこで本稿では，この手法を応用し，有価証券報告書専用の因果関係文を抽出する判別モデルを提案した．そして，判別モデルの評価等を行い，高い性能であることなどを示した．この判別モデルにより有価証券報告書独自の投資判断に有益な情報の効率的な抽出が期待できる．

本研究では，経済新聞記事などの経済テキストから、日経平均株価などの市況について 言及している文書のみを抽出し，それらの内容を自動的に要約することにより，ファンドの運用報告 書における市況分析コメントを自動生成する手法の開発を行う．本手法では，まず経済新聞記事から 深層学習により日経平均株価の市況について言及している記事を抽出する．次に抽出された記事の 中から例えば「ギリシャへの金融支援協議が難航していることや、中国・上海株の値動きへの警戒感 から、投資家のリスクオフの動きが強まった。」のような日経平均が大幅に変動した理由について言 及している文を抽出する．そして，抽出された文を時系列順に並べることで市況分析コメントを自動 生成する．

複数のテキストデータから経済・金融事象を背景知識まで含めて可視化することは, 経済・金融事象の理解の助けになり有用である. しかしながら, 経済・金融事象の連鎖を手動で抽出することは非常に時間とコストがかかる. そこで, 本研究では経済・金融事象の連鎖を因果関係として扱い, 各事象を表したベクトル間の類似度を用いて因果関係の連鎖を構築する手法を提案する. また,提案手法における問題点から今後の提案をまとめる.

アナリストレポートは，証券会社のアナリストが企業の経営状態や収益力などを調査してまとめたものである．以前の研究では，学習データを自動生成し，深層学習を用いてアナリストレポートに対して極性を付与する手法を提案した．そこで本研究では，深層学習を用いて極性付与されたアナリストレポートと株式リターンとの関連性について実証分析を行った．その結果，アナリストによる定量情報に対する予想変更が無いレポートについて，市場は該当レポートに付与された極性に沿って短期的に反応することや，レポートにて言及された銘柄は，極性の違いによってその後の長期にわたり株式リターンに違いが表れることが明らかになった．

経済文書のような専門的な文書は非専門家にとって読みにくい場合が多い．そのため，非専門家を対象に経済文書上のセンチメントを単語単位で可視化するようなサポートシステムを構築することには一定の需要があると思われる．経済文書上のセンチメントを可視化する手段の一つとして近年提案された「ニューラルネットワークモデルの解釈」に関する手法，LRP (Layer-wiseRepresentation Propagation) を用いるという手段がある．しかし現状 LRP が日本語の経済文書の可視化に有用かどうかは調査されておらず，その性質についての詳細な分析もあまりされていない．また，LRP の Attention RNN への適用方法は未だ提案されていない．本報告では LRP の AttentionRNN への適用方法を提案し，また，LRP が日本語金融テキストの可視化に有用かどうかを検証する．さらに，実データを用いた検証の中で LRP を用いた日本語文書可視化の性質について分析する．

金融市場におけるテキストデータは、投資家にとって分析対象の一つであり、その重要性は日に日に増している。決算短信をはじめとする企業の業績について書かれたドキュメント、証券会社のアナリストが書いた個別企業についてのアナリストレポート、更にはツイッター等のSNSで発信している個人投資家のつぶやき等、多種多様なテキストデータが存在する。定性的な投資判断の材料になるとともに、これらのテキストを用いて、モデルを構築し、定量的な分析を行うこともある。本研究では、中央銀行が発行する議事録やステートメントなどの公的な文章を対象に、定量的な数値化を目的としている。具体的には、日本銀行の発行する金融政策決定会合の議事要旨から物価、生産、雇用等の各トピックに対するセンチメント指数の構築を目指す。

In recent years, textual information, which is unstructured data attracts attention as new analytical data in the financial and economic fields and it is expected to structure knowledge on this domain. One such knowledge is a sentiment polarity dictionary in which each word is representing positive or negative. In building the dictionary, it is costly to add the polarity value to a vast number of words manually. Therefore, in this research, we propose a the dictionary construction model especially considering the synonymity and symmetry of words. As a result of the experiment, the proposed method is a more accurate than the model of the previous research. In addition, we extended the conventional dictionary using the proposed method, and we showed that the extended dictionary has higher accuracy than the dictionary which is not extended.

Recently, with the increase of individual investors, the necessity of investment support technologies is increasing. Although analyst reports on which professional securities analysts forecast business performances or stock prices of companies are regarded as important investment decision materials, writing an analyst report is heavy burden. In this research, we summarize newspaper articles and support the generation of analyst reports by using knowledge of information features which are referred to as reasons for analysts' forecasts of business performances or stock prices in analyst reports.

This paper aims to predict a company's financial index by analyzing articles about the company. The authors propose MultiMedLDA, which is one of supervised topic models. MultiMedLDA assumes that each document has two types of labels, discrete value label and continuous one. It models relation between each document and these labels, and predicts an unknown label based on known labels and the documents. Making use of not only documents but also the known labels, it improves prediction accuracy. We evaluated our model with data from the "Japan Company Handbook". Using comments for each company as a document, the type of industry as a discrete value label and the company's ROE (Return On Equity) as a continuous value label, we predicted the ROE in the evaluation.

In the economic and financial fields, there is a growing interest in obtaining new knowledge from large quantities of data, such as corporate financial data and international exchange transactions. On the other hand, as a technical trend of recent data analysis, deep learning-based models have been successfully applied to various data, such as images, text, and audio. Especially, Recurrent Neural Network (RNN) and its extension of Long Short-Term Memory Network (LSTM) have been developed as deep learning for sequential data or time series. However, regardless of its importance, LSTM has not applied to corporate financial time series, such as in the Financial Statements Statistics of Corporations, to the best of my knowledge. In this research, considering the above-mentioned trends, we conduct regression analysis using LSTM for corporate financial time series. For experiments, we obtain the capital investment rate and other financial indicators, such as the cash flow ratio, for each target company from the Financial Statements Statistics of Corporations, and then use them as the objective and explanatory variables, respectively. By changing the number and types of explanatory variables used in the experiments, we evaluate the contribution of each explanatory variable to regression power to the objective variable at several time steps ahead. Furthermore, as baseline methods for the regression tasks, we evaluate the regression power of classical methods: Autoregressive Integrated Moving Averaging (ARIMA), and discuss the comparative evaluation with the LSTM approach.

Predicting the price movements of stocks based on deep learning and high frequency data has been studied intensively in recent years. Especialy, limit order book which describes supply-demand balance of the market is used as feature of a neural network, however, these methods do not utilize the properties of market orders. On the other hand, order encoding method of our prior work can take advantage of these properties. In this paper, we apply some types of convolutional neural network(CNN) architectures to order-based features to predict the direction of mid-price movements. The results show that smoothing filters which we propose to employ over embedding features of orders improve accuracy. Furthermore, inspection of embedding layer and investment simulation are conducted to demonstrate the practicality and effectiveness of our model.

Open-high-low-close price (also OHLC) series have been widely used for the pricemovement analysis of financial time series, including to draw candlestick charts. Modeling these data is complicated by the fact that such data are often unlikely to be samples of stationary stochastic processes, as can be seen in the well-known phenomenon of volatility clustering. In this research, first we try to remedy this matter by using the sequences of differences between high and low prices, which are pointed out to often have higher autocorrelations than the absolute returns of close-price series, and normalize the scales of OHLC by their exponential moving averages. Under our experimental conditions, the Earth Mover's Distance (EMD) between normalized S&P500 training and test data is about one-seventh of the EMD between the unnormalized data. Second, we try to model the normalized data by introducing 6 generative models for them. The EMDs between data generated by our learned models and the normalized test data are about one-sixth of the EMD between the normalized test data and the delta distribution located at the barycenter of the normalized training data. However, they are about 5 times larger than the EMD between the normalized test and training data.