In this study, we analyze the information value of extreme opinions on Twitter that are identified by the most positive and negative Twitter sentiments for each firm. We find that these extreme opinions predict stock returns without subsequent reversals. In addition, they contain incremental information regarding firm fundamentals that are identified by subsequent revisions in analysts' earnings forecasts and target prices. Finally, we find that the return predictability is attributed to the fundamental information contained in the extreme tweets. Our analysis sheds light on the role of extreme opinions on social media.

In This paper, the investment performances of the AI traders that predict fluctuations in Nikkei 225 Futures during the period under COVID-19 Crisis were measured. In addition, by comparing the data under the Lehman Shock with "learned AI traders" and "not learned AI traders", learning of past market crash could improve the investment performance of AI traders in future market crash. Furthermore, it was confirmed that AI traders who have learned the period of the Lehman shock were more likely to avoid risk when the market fluctuation range was small.

Various preceding market models can reproduce basic financial stylized facts such as volatility clustering, but most of those models require ad hoc tuning of parameters for the reproduction. Inspired by the idea of the sandpile model, we present a simple agent-based model of the financial market named Self-organized Speculation Game, where the number of traders is spontaneously tuned. While this model has high reproducibility of stylized facts, it holds similar behavioral properties to those of the sandpile model. The simulation results infer the possible contribution of self-organized criticality for the spontaneous emergence of stylized facts.

Recognizing risks that are descriptive information is important to judge how companies manage their risks, that is, the consistency of the strategy when evaluating companies. So, in this research, risk is defined as "uncertainty of future results that can be obtained by taking actions", and we tried to extract the details of the risks recognized by each company from the text of the securities report. We also tried to classify the extracted risk based on their respective contents. The extraction method is investigating and classifying the expression patterns (possibly, etc.) used when expressing the content of risk, and using them as clues.

In this paper, we examine whether it is possible to generate the government's "Assessment of the current state of the economy" by the Japanese government using several economic indicators that are believed to be consistent with the Japanese economic trend. In the Monthly Economic Report released by the Cabinet Office, the government publishes the government's "Assessment of the current state of the economy" along with expressions such as "recovering," "worsening," and "standstill", etc. In this paper, we developed a model to generate the government's "Assessment of the current state of the economy" and tested the accuracy of the predictions. In addition, we examined which economic indicators are likely to influence the government's "Assessment of the current state of the economy". It is expected that this analysis will be useful in predicting the government's policy decisions.

When predicting stock prices with a complex model using machine learning or artificial intelligence, overfitting sometimes occurs, and the prediction accuracy expected in actual operation cannot be obtained. In such a model, the cost function is presumed to be steep and multi-modal, while in a model that maintains stable prediction results, the cost function is considered to be gradual and single-peaked. In this study, we first compared the performance of several stock price prediction models, and then visualized the cost function for each model using t-SNE. As a result, the model using Lasso regression, which had the highest performance, showed a gradual unimodal cost function, while the linear regression, which had relatively low performance, showed a steep and multi-modal shape. Visualizing the cost function using t-SNE can be an important index for evaluating the stability and versatility of a stock price prediction model.

In recent years, investment strategies in financial markets using deep learning have attracted a significant amount of research attention. The objective of these studies is to obtain investment behavior that is low risk and increases profit. Although Distributional Reinforcement Learning (DRL) expands the action-value function to a discrete distribution in reinforcement learn- ing which can control risk, DRL has not yet been used to learn investment action. In this study, we construct a low-risk investment trading model using DRL. This model is back-tested on Nikkei 225 data and compared with Deep Q Network (DQN). We evaluate the performance in terms of final asset amount, standard deviation, and the Sharpe ratio. The experimental results show that the proposed method can learn low-risk actions with the increasing profit, outperforming the compared method DQN.

This paper examines the possibility of applying the novel likelihood-free Bayesian inference called BayesFlow proposed by Radev et al. (2020) to the estimation of agent-based models (ABMs). The BayesFlow is a fully likelihood-free approach, which directly approximates a posterior rather than a likelihood function, by learning an invertible probabilistic mapping that implements a Normalizing Flow between parameters and a standard Gaussian variables conditioned by data from simulations. This deep neural network-based method can mitigate the trilemma in the existing methods that all of the following three ?higher flexibility, lower computational cost, and smaller arbitrariness cannot be achieved at the same time. As a result of the experiments, BayesFlow certainly achieved the superior accuracies in the validation task of recovering the ground-truth values of parameters from the simulated datasets, in case of a minimal stock market ABM. The method did not involve any extensive search of the hyperparameters or handcrafted pre-selections of summary statistics, and took a significantly shorter computational time than an existing non-parametric MCMC approach.