値上がりが期待できる株式を選別して投資するアクティブ投資は，投資先企業に本源的に存在する価値に基づいて投資するため，企業価値に即した適正な価格を発見し，市場価格をその価格に近づける（市場を効率的にする）という主張がある．しかしながら，実際のアクティブ投資が市場を効率的にしているかどうかは分かっていない．近年の実証分析は，比較対象となるインデックス（日経平均株価などの指数）から大きく乖離した割合で銘柄を保有しインデックスに対して大きく異なる利益となるリスクをとったうえで売買量が少ない，いわいる"忍耐強い(Patient)アクティブ投資"が利益を得ていることを明らかにした．今後このような投資が増えることが予想されるが，売買量が少ないにも関わらず市場価格に影響を与え市場を効率的にするのかどうかは重要な論点である．そこで本研究では，忍耐強いアクティブ投資の特徴を反映した投資家を導入した人工市場モデルを構築し，このような売買量が少ない投資家が市場価格に影響を与え市場を効率的にするのかを議論した．その結果，忍耐強いアクティブ投資はまれに起こる，市場価格が企業価値に即した適正な価格から大きく乖離して市場が不安定になり，市場がさらに非効率になりそうなときのみに多く売買を行い，市場を効率化することに寄与していることが示された．市場価格の変動が大きくなると，投機的な注文はその変動をさらに大きくすることがあるが，忍耐強いアクティブ投資の注文は，このような増幅を防いでいる可能性も示された．本研究ではさらに，"忍耐強くない(Impatient)アクティブ投資"もモデル化し，アクティブ投資家の数を一定に保ちその構成を変化させた場合も分析した．その結果，忍耐強くない投資家が減り，忍耐強い投資家が増えると，アクティブ投資の総利益のみならず，投資家1人あたりの利益も増えることが分かった．つまり，忍耐強い投資家が増えることにより，競合が増えて利益を奪い合うのではなく，仲間が増えることにより売却したい価格に到達しやすくなり投資機会が増え利益を獲得しやすくなることが分かった．このことは，忍耐強い投資は市場が効率的になると利益が減るというよりは，市場が非効率すぎて利益が減るという側面もあることを示している．

A financial market is a diversified dynamical system with many constraints, and price movements are modeled in terms of the micro properties of each transactions and the macro properties of dynamical systems. These two properties must be bridged by the multiplicity. The model forcuses on the size of tick and the number of transactions with the price movement compared with the previous transctions and explains stochastic nature of short-term volatilities and persistently stable long-term volatilities as a results of unique behaviour of heterogeneous market participants.

多くの人工市場モデルにおいて、エージェントが注文を出す順序や頻度に対して十分な考察が行われておらず、エージェントはランダムもしくは固定の順序に応じて注文を出し、各エージェントの注文頻度はほぼ同一であることが多い。ところが、このようなモデルにおいては市場の状態に応じてエージェントが自分の注文を能動的にキャンセルすることが出来ない。株式市場においてキャンセルは注文総数の40%以上を占め、市場に大きな影響を及ぼしており、これを適切にモデル化することが求められている。そこで、本研究はその第一歩として、エージェントがキャンセル注文を出す確率が板に指値注文を出してからの経過時間に依存するモデルを構築し、キャンセルされた注文の寿命(キャンセルされるまでの経過時間)の分布の冪乗則を再現した。

A credit scoring model is a useful tool for Small and Medium-sized Enterprises (SMEs) lending. In this study, we investigated methods to improve the accuracy of the scoring model using machine learning. As a result, we have shown that Gradient Boosting Decision Tree (GBDT) can obtain the highest accuracy. We found out that GBDT shows better performance than other methods especially when we use more than 10000 learning data. In addition, we demonstrated that ensemble learning can further improve accuracy. According to our simplified estimation, it was suggested that the ensemble learning can reduce the default rate by 16% compared with the conventional method.

We analyze Japan equity portfolios for their risk exposures to North Korea geopolitical destabilization effects over the past few months, by using an implementation of the EM algorithm integrated with a GARCH process. The model identifies which latent factors in the Japan equity market relate to North Korea and predicts their risk impact for the near future.

In This paper, we propose a construction method of an artificial intelligence for TOPIX trading. We apply a Multi-layer Neural Networks as a prediction method, and optimize to maximize the Information Ratio using GA. We also conduct a simulation using TOPIX market data from January 2014 to December 2016, showed the effectiveness of the proposed construction method from the information ratio level.

In recent years, many researchers have taken keen interest in analyzing various kinds of relational data, such as social networks and financial networks. These data can be expressed as a graph or network where each vertex or node is an entity and each edge or link is a relation between a pair of entities. Moreover, each link is often associated with continuous and/or discrete relational attributes, such as in financial networks, the interest rate for a transaction and whether the transaction is international or intranational. In this paper we focus on max-margin latent feature relational models (called Med-LFRM) that are based on Indian buffet process (IBP) and maximum entropy discrimination (MED). For the estimation of model parameters, the Bayesian estimation is deemed equivalent to minimizing an objective function, which involves misclassification errors. We focus on link prediction problem for the networks with continuous and discrete relational attributes. We also focused on the time dependent analysis for the networks, and therefore, we estimated the model parameters considering the observations in the previous time interval. We demonstrate, through experiments with inter-bank financial networks, the effectiveness of the above model in terms of the link prediction performance.

The forecasting the stock price of a particular has been a difficult task for many of analysts and researchers. In fact, investors are highly interested in the research area of stock price prediction. However, to improve the accuracy of forecasting a single stock price is a really challenging task, therefore in this paper, I propose a sequential learning model for prediction of a single stock price with corporate action event information and Macro-Economic indices using LTSM-RNN method. The results show the proposed model is expected to be a promising method in the stock price prediction of a single stock with variables like corporate action and corporate publishings.

We study the limits of prediction accuracy of Bitcoin price data in CNY currency using tick data from the OKCoin Bitcoin exchange (source: Kaiko data). The tick data contain the price, volume, and trade direction, and are transformed to the OHLCV format using standard methods. In this report, we deploy the Support Vector Machine algorithm by Vapnik to estimate the sign of the hour-to-hour transaction return using a sampling moving window of varying size on the past data. Several kernel functions are validated. Our first results for all months of the year 2015 show that the hit ratio accuracy level (the fraction of correctly predicted upward or downward events) does not exceed 60%. It remains to be established whether this low result corresponds to the causal extraction limit inherent in the data, or whether it can be improved by deploying other methods, such as LSTM networks in deep learning.

機械学習を用いた株価予測については近年多く研究されている．しかしながら，その多くが株価ないし株価指数そのものをそのまま予測対象としている．一般に株価は非常に複雑な振る舞いを示し，予測が難しい．一方で，同業種同規模などの似通った銘柄間では価格差 (スプレッド) が平均回帰すること (所謂共和分性) が知られている．そこで本研究では，株価そのものではなく，共和分性を満たす株価ペアのスプレッドを機械学習で予測する手法を提案する．具体的には，LSTM を用いて，定常性を満足するモデルで最も代表的な AR(1) 過程に従う人工的な時系列データを事前学習する．学習した LSTM を用いてペア・トレード戦略に適用した結果，単純な AR(1) 過程や実データを直接学習させた LSTM よりも良好な結果が得られた．

This research has two objectives: (1) to model and analyze the momentum effect, (2) to propose a portfolio reconstruction algorithm that can use the momentum effect to obtain excess profit. The momentum effect tends to be present in the stock market, and describes the phenomenon whereby rising (declining) stocks tend to continue to rise (decline). However, because existing research does not separate momentum effects from stock price fluctuations it is not always possible to obtain excess return when working with an unknown data set that contains a momentum effect. In this research, we define a new External Force Momentum Effect (EFME) model based on bias in stock price rises (declines). We prepared an artificial data set that contained this momentum effect and constructed a portfolio with the proposed algorithm. The relationship between the EFME model and excess return is then analyzed to verify that excess profit can be obtained. Additionally, we confirmed that the proposed method can obtain higher excess return than the existing method when applied to artificial and real stock data sets.

In this research, we propose a method to automatically classify sentences including causal information concerning business performance (e.g. "Orders of semiconductor manufacturing equipment were good.") extracted from summary of financial statements of companies based on business segments of the companies. Moreover, we propose a method to extract performance sentences from summary of financial statements. For example, the sentences including causal information extracted form summaries of financial statements of SUBARU Co., Ltd. are classified to either "automobile" segment or "aerospace" segment. In addition, our method extracts performance sentences, e.g. "Sales were \3,262.0 billion, an increase of \93.7 billion (2.9%) compared with the previous fiscal year.", by deep learning and automatically generates training data.

In this paper, we proposed a method for supporting investors. Our method extracts basis information on analyst's forecasts from analyst reports and assigns polarity to the analyst reports. Analyst reports which are written about a company's performance or profitability by securities analysts are useful for investment but investors can only read it a little because many reports are published. Therefore, a system which judges investing by an artificial intelligence technique is required. By giving polarity to the analyst reports, the proposed method catches a slight change in performance. This ability of method is useful to judge whether investors need to read analyst reports carefully.

These days, a growing number of individual investors is attracting public attentions even in Japan, and securities companies are actively providing them with investment informations. Especially, analyst reports written by professional security analysts are important investment judgment materials, but their timing of publication varies by brands. In this paper, using the structures of causal relationships of sentences in analyst reports, the ways of security analysts paying attention to cause informations concerning business performances were learned, and newspaper articles including similar causal relationships were extracted. We aim to realize a real-time investment supporting system.

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.

In recent years, many people attempt to apply machine learning to a broad range of fields. At the same time, thanks to improvements of the information technology, we can use the enormous number of the past financial data as a learning data . In this paper, we describe prediction results of the foreign exchange price. The prediction model was made by using Convolutional Neural Network(CNN). Then, we compare the results of the model which learned numeric data with the model which learned image data.

Recent technological breakthrough has enabled us to study the microstructure of financial markets using the high-frequency trading data. In this presentation, we review our recent preprint (arXiv: 1703.06739), in which individual traders's strategies are analyzed on the basis of informative order book data with anonymized trader identifications. We empirically study the trend-following behavior of individual traders on the basis of conditional statistical analysis. We then propose a microscopic model of financial markets on the basis of the empirical finding of trendfollowing of individual traders. We further develop a systematic theory to our microscopic model paralleling to the mathematical formulation of kinetic theory. Finally, the agreement between empirical results and our theoretical predictions are shown in terms of the order book profile and the price movement distribution.

Fundamental factor models are one of the important methods for the quantitative active investors (Quants), so many investors and researchers use fundamental factor models in their work. But often we come up against the problem that highly effective factors do not aid in our portfolio performance. We think one of the reasons why is that the traditional method is based on multiple linear regression. Therefore in this paper, we tried to apply our machine learning methods to fundamental factor models as the return model. The results show that applying machine learning methods yield good portfolio performance and effectiveness more than the traditional methods.

In this paper, we propose order-based approach to predict future movements of a stock price. Our models employ a convolutional neural network(CNN) over embedded orders that have quantitative and qualitative variables. For each dataset of stock codes, the models outperform traditional feature-based approaches. Furthermore, we show that training under less influence of noise can be performed by applying an averaging filter to embedded feature space. Analysis of the embedding layer reveals that the models put emphasis on the features of market orders that are correlated with price return.

人間の経済行動には非合理的な行動傾向があることが、行動経済学におけるプロスペクト理論において知られている。この理論では、人は利得している状態の時は損失回避行動をとるのに対して、損失を出している状態の時はリスク選好行動を取ることを示している。今回、外国為替証拠金取引のデータとして、28 ヶ月にわたる 8 万 1300 人による 2800 万回の取引データを解析した。そして、この取引の参加者達にもプロスペクト理論から示唆される不合理な投資行動パターンが存在することを示した。さらに、この取引の参加者達を、利益を得ているグループと損失を出しているグループに分けた時、損失を出しているグループの方がより強くプロスペクト理論から示唆される非合理な取引行動をとっていることがわかった。

After the Bank of Japan (BOJ)'s monetary policy meeting, the BOJ governor makes a press conference to explain the monetary policy. In this research, using facial expression recognition algorithms based on deep learning, we analyzed the governor's facial expressions in the press conference and estimated the emotional indexes such as "Happiness", "Anger", "Sadness", "Surprise". As a result, we found that the indexes of "Anger" and "Disgust" increased much just before making major policy changes. On the other hand, the index of "Sadness" tended to decline after the monetary policy changes. This suggests that the information based on the facial expression analysis can be a useful material for forecasting the future monetary policy.

本研究では、Agent Based Model （ABM）の未知パラメターを、ニューラルネットワークの Adversarial Training を応用したベイズ的アプローチによって推定する方法を提案する。ABM のベイズ推定では、（1）パラメターの探索値毎のシミュレーション、（2）シミュレーション生成データと実データを突き合わせた尤度の計算、（3）事後分布のサンプリング、という計算負荷の大きいプロセスをくり返す。そのため、特に、推定するパラメター数が複数になる場合、（2）と（3）における効率化が必須となってくる。（2）の尤度計算について、パラメトリックな分布を仮定して省力化する手法もあるが、特定化の誤りによる大きな精度損失が懸念されることから、より柔軟な近似手法への需要がある。本研究では、こうした問題意識から、Adversarial Training によって、ABM の推定において尤度を近似するニューラルネット（判別器）を、現実データから適合的に獲得するアプローチを検討した。

In this paper, we propose a methodology to effectively capture credit risk from firms' network. In short, our target is to numerically obtain additional credit risk from connected firms on network. Recently, commercial networks are available for investing and managing risk on professional information terminals like Bloomberg and Reuters. They enable us to check commercial connection of firms. We utilize them to expect positive and negative effect on observing firms from neighbor firms especially when the neighbor firms have any credit events. We propose a methodology to analyze/measure impact which observing firms potentially receive from their neighbors. We applied Merton model which is generally utilized for credit risk management to calculate additional risk and simplified the formula for practicability/usability. Also, it enables us to escape from having any difficulties in computation time. We introduce our approach with overviewing simple model guidance and explaining a few samples of numerical experiments.