In recent years the risk of mutual funds has become difficult to assess. This is because mutual funds have come to choose various assets, some of which may have high risk, and there may be some funds whose performances depend on dangerously much those of a part of the assets that the funds hold, and so on. Companies managing mutual funds are expected to perform risk control to prevent investors from taking unforeseen risk of funds. A related revision to the investment fund legal system in Japan led to establishing what is known as"the rule for investment diversification." Some papers discussed that the rule for investment diversification affected price formation at the time of a market crash; however, we could not find that it affects price formation at the time of a soaring market. In this paper, we investigated that the rule for investment diversification affected price formation in financial markets where two types of investors who followed the rule and did not follow it participated at the time of a soaring (crashing) market that was caused by the bulge (collapse) of an asset fundamental price using agent-based simulations and discussed the difference between these effects. As results, we found that, in two-asset markets where two types of investors who followed the rule and did not follow it participated, when one asset fundamental price soared, the rule for investment diversification prevented its price rising, and when one asset fundamental price collapsed, the rule promoted the other's price to decline.

Recent popularity in algorithmic trading has spurred on researchers to investigate the variety and the evolution of the trading strategies. In this talk, we present our recent study (under review in PLOS ONE), in which the strategy distribution of limit orders is analyzed by using the high frequency data set including anonymized trader IDs. We first identify timescales for each trader to measure market-price trends by the multi-regression analysis. Clustering the timescales into several clusters, we then show the frequencies of the submissions and transactions for each cluster. Furthermore, we provide the microstructure insight to their frequencies in terms of the average shape of limit orders. Finally, we quantify the activity level of each cluster, and show that some clusters are unique to the local time in Tokyo or New York.

In recent years, investment strategies using artificial intelligence have attracted a significant amount of research attention. However, it is difficult to construct an efficient investment strategy using artificial intelligence owing to the variable factors in market prices. Therefore, this study aims to focus on a trading method called the NT ratio transaction to reduce the number of price-variable factors. This transaction is an arbitrage transaction, which utilizes the difference in the price movements between Nikkei 225 futures and TOPIX futures. These futures generally exhibit similar price movements and even if the price differences expand, they tend to return to their original separation. Using this transaction, we can target profits from this price difference while offsetting a considerable number of price-variable factors. Therefore, in this study, we construct a model to acquire an investment strategy based on NT ratio transactions via deep reinforcement learning and confirm the effectiveness of this model.

In recent years, many researchers have studied stock trading using technical analysis. However, it is necessary to have deep knowledge to use such technical analysis and it is difficult to make a profit using such techniques. Therefore, we construct an evolutionary model to create a profitable investment strategy using technical indicators.

In This paper, I propose a method to use Artificial Intelligence for early warning in TOPIX trading. I construct an early warning system with AI that includes AI traders who preliminarily learned TOPIX market data, and by using forecast of AI traders as early warning indicators, efficiency of scoring the indicators has been improved. In addition, I conducted a TOPIX trading simulation using the early warning system with AI, which resulted in a higher Sharpe ratio than that of TOPIX.

In this research, I have developed a prediction model of long-term interest rate (long-term government bond yield) using machine learning method (SVM, nonlinear SVM, decision tree, RF, logistic regression, LSTM). As a result, it was confirmed that the accuracy of the LSTM-based model is relatively higher in the long-term interest rate prediction than the other models. Furthermore, long-term interest rate is influenced by interest rate fluctuations in the surrounding maturity due to the influence of arbitrage transactions. Therefore, I constructed a fluctuation model of the yield curve incorporating the relationship between long-term interest rate and the other maturity rates in the form of extending the above-mentioned LSTM-based prediction model. As a result, when using the yield curve fluctuation model for predicting long-term interest rate, some improvement was seen in the prediction accuracy of long-term interest rate.

In foreign-exchange (FX) dealing, FX brokers basically cancel out the orders from their customers to prevent the price fluctuation risk by cover transactions with global megabanks called Counter Party (CP). Each CP has huge amount of money to play a role of market reader, and might have proprietary know-how to foresee future price movements. From this viewpoint, we try to extract their knowledge by a machine learning approach, and therefore we apply the stacking method that aggregates some predictors to extract the ensemble knowledge. If CP's price quotations are decided by foreseeing future price possibilities, their quotations can be considered as predictors. From this concept, we apply the stacking method to their quotations and obtain the ensemble knowledge from them. Through some simulations using real price data, we could confirm that the given ensemble knowledge improves the prediction accuracy of FX price movements compared to the machine learning using a single CP's price quotation.

In this research, we propose a method of extracting business segments from securities reports and extracting sentences including causal and result information concerning business performance for each extracted business segments. For example, our method extracts "In the aluminum rolled products business, shipments of high purity foils for aluminum electrolytic capacitors for industrial equipment and automotive use increased and sales increased." as a sentence including causal information concerning business performance. Moreover, our method estimates that the sentence belongs to business segment "aluminum". Our method extracts "As a result of this segment sales in this segment were 105,439 million yen, operating profit was 6,697 million yen." as the result information belong to "aluminum" segment.

In this paper, we propose a method for classifying timely disclosures of company information with risk to the corporate performance by using deep learning. For example, our method classifies a timely disclosure "A notice of recording a extraordinary loss" as "Including risk". Moreover, our method classifies the timely disclosure as "Extraordinary loss". Our method makes timely disclosures of company information easy to see and investors will be useful to make investment decisions. We evaluate our method and it attained 87.4% accuracy.

In this paper, we propose a method for extracting causal relation between performance factors and performance results from summaries of financial statements by deep learning. For example, our method extracts "Sales of copper tubes for air conditioners declined due to high demand due to the hot summer heat caused by the hot summer of the year" as a performance factor, and "Consolidated operating profit was 4,859 million yen" as a performance result corresponding to the performance factor. By extracting such causal relation, it is possible to analyze what kind of factors have changed the performance.

In this paper, we examined a new method to predict fluctuation of foreign exchange rate. We input information extracted from historical data into PointNet++, which is proposed by C.R.Qi et al.[3] and predicted US Dollar to Yen Exchange rate 5 minutes later. The results implied that although there was tendency of overfitting, our method might capture a part of some structured factors of foreign exchange fluctuation. It is suggested that various range of approach related to machine learning could be useful for financial problems.

Foreign-exchange trading (FX) is well known as an asset management method like stock investment. Individual traders basically send their orders to an FX broker, and the FX broker executes cover transactions with global megabanks to prevent the price-fluctuation risk. If it is possible to foresee the future price movement, FX brokers can make their cover transactions more efficient. Fortunately, FX brokers can see the trading positions of their customers. If each customer has a little predictive power, the aggregation of all customers' positions might improve the predictive power in terms of the wisdom of crowds. From this viewpoint, we tried to extract the collective knowledge from all customers and applied it to improve cover transactions. As a result, our idea worked well to make FX brokerage business less risky and more profitable.