The aim of this research is to build an ontology of properties and manufacturing processes of inorganic materials for supporting engineers in analyzing patent documents. Recognizing that the properties of inorganic materials are context-dependent, the target concepts are defined in relation to their specific contexts. Firstly, properties are defined as fundamental properties that depend on objects or manufacturing processes. Secondly, manufacturing processes are defined as concepts that imply either what is achieved only or how it is achieved as well. Thirdly, the concepts of improvement activity, frequently encountered in patent documents, are defined. The potential utilization of the ontology in patent analysis is discussed, followed by the development of an ontology-based patent analysis system. This system can capture object-property-value triples and restore missing information. Its test use demonstrates its usefulness in facilitating patent analysis.

While securities reports contain valuable information to investors, the non-financial information contained therein is not well-structured, and it is not easy for investors to access desired information. Some necessary information cannot only be interpreted from the text but from tables as well, but since the format of such tables is not standardized, it requires considerable effort to find the rationale from the tables. In this paper, we propose a TTRE (Text-to-Table Relationship Extraction) task regarding the numerical values contained in securities reports, assuming that by linking the text with the relevant cells in the tables, we can assist investors in finding such information. We propose the task, build a dataset, which is then published, and host a shared task using it. By analyzing the evaluation results of the baseline methods we implemented, we discussed the difficulty level of the TTRE task and the future challenges to improve its performance. The results showed that the TTRE task requires consideration of not only the text of sentences and cells but also the structure of tables, making the task very challenging. It was also found that regarding the dataset construction, annotators’ judgements contained fluctuations, suggesting that the annotation method and the dataset size need to be improved.

The number of funds remains high and the amount of economic and social news information increases rapidly, increasing the burden on asset management companies in preparing disclosure documents. These disclosure documents are important for mutual fund holders, and in particular, market comments and outlooks are essential to understanding the current and future investment environment. Writing these documents takes a lot of time, adding to the workload of asset management companies. Recently, advancements in Large Language Models (LLMs) have expanded their use in various tasks. However, LLMs struggle to easily learn new information due to computational resources and costs, making it challenging to generate market comments and outlooks reflecting the latest economic data. Retrieval Augmented Generation (RAG) is a solution to this problem. In this study, we use ChatGPT, a type of LLM, to develop a tool that automatically generates market comments and outlooks. This tool can incorporate the latest news information and generate comments based on the information while suppressing hallucinations. We propose two methods: few-shot learning, which uses past market comments as examples, and zero-shot learning, which does not use past market comments. We collected actual market comments from publicly available mutual fund disclosure documents and conducted both qualitative and quantitative evaluations in comparison with the generated comments.

Financial results briefings are a bidirectional communication channel between companies and stakeholders, complementing unidirectional communication through annual reports and integrated reports, among others. Financial results briefings provide a platform for companies to promptly explain their financial performance, business status, and strategies. They typically consist of a presentation by management and a QA session, enabling participants to directly inquire about financial performance and resolve concerns. We analyze the information value of financial results briefings from both the company and investor perspectives in the Japanese market. For companies, the value lies in determining whether information disclosed during these financial results briefings reduces the cost of capital. Higher-quality information disclosure is theorized to mitigate information asymmetry among market participants and, consequently, lower the cost of capital. Therefore, we assign sentiments to text data from financial results briefings, considering the corresponding financial results. We then examine the correlation between text sentiments, text lengths, and the cost of capital. For investors, the question is whether information disclosure during these financial results briefings influences post-disclosure abnormal returns. We examine the relationship between sentiment and postdisclosure abnormal returns. As a result, from a company’s perspective, we find that the cost of capital tends to be lower when the text length in the QA session is large. This suggests that sufficient information disclosure contributes to reducing the cost of capital, and is consistent with existing empirical studies and theoretical models. From the investors’ perspective, we find that the sentiment of the QA session and the excess sentiment beyond the expected from the explanation are associated with short-term abnormal returns. This result confirms that QA is a valuable source of information for investors. We underscore the significance of financial results briefings in both company IR activities and from the perspective of investors.

Since the emergence of the COVID-19 pandemic, disruptions in supply chains have significantly impacted both the global economy and asset markets. Despite a rising interest in supply-related data among policymakers, researchers, and financial market participants, existing indicators often wrestle with pervasive issues of low frequency and coarse granularity. In this carefully crafted paper, we ambitiously propose new, robust indices for the highfrequency nowcasting of disturbances specifically within the automotive supply chain. Firstly, by judiciously utilizing inter-factory transition data alongside time series anomaly detection methodologies, we have successfully created the Supply Chain Turbulence Index (SCTI). To further augment the SCTI, we introduce a novel, sophisticated technique, grounded on the principles of enhanced Variational Autoencoders, to diligently isolate supply factors contributing to bottlenecks in the supply chain, subsequently creating a nuanced subindex, dubbed SCTI-supply. The SCTI exhibits a strong correlation with existing statistics concerning supply chain delays and demonstrates the remarkable capability to detect micro-level production interruptions across various car manufacturers’ plants. On the other hand, SCTIsupply correlates effectively with low-frequency supply chain indicators we developed from established statistics and proves exceedingly effective in identifying supply shocks under rigorous event study analysis.

In this paper, we propose a method for automatically evaluating integrated reports in terms of interpretability using a deep learning model with hierarchical attention to two types of information: sentences and pages. Specifically, we newly constructed a dataset labeled with investor evaluation labels based on the external evaluation of integrated reports by the GPIF. In addition, we proposed a deep learning model with a Bi-LSTM layer for learning page order and two attention layers for sentences and pages. In the evaluation experiments, the proposed model achieved a classification performance of F1-score 0.847 for integrated reports of companies that do not appear during training. In the discussion, by visualizing the weights of the attention layer of the model, we confirmed that the information of interest in the model is generally consistent with the evaluation criteria of investors. In addition, we examined the practicality of the proposed method by taking into account the bias of the dataset, and showed that the proposed method is able to automatically evaluate both learned and unknown integrated reports of companies.

Transactions in Contract for Difference (CFD) markets are increasing. However, because of the limited information and data, it is difficult to grasp the characteristics of the market as a whole. In Japanese CFD markets, the market maker system has been adopted. In market maker system, brokers called market makers act as intermediaries. Specifically, Market makers indicate both buy and sell quotes and guarantee execution on investors’ orders. Although market makers play an important role in providing liquidity to the market, there are few studies on their behaviors due to the limited data available.

The main source of profit for market makers is“ spread ”, the difference between the limit prices of both buy and sell orders to the market executed in equal units. On the other hand, the position generated when only one side of the orders is executed is exposed to the risk of price change of the underlying asset. Such a position held by a market maker due to his or her bias in buying and selling is called“ inventory ”. Market makers sometimes manage their inventory by hedging, which is referred to as counter-orders to the underlying asset market. They can reduce inventory more quickly and reliably by hedging than adjustment of the quotes, but are usually forced to bear hedging cost because they are price taker in undelying asset market. Therefore, market makers need to manage the spread profit, inventory risk, and hedging costs.

In this study, we reproduce CFD market on an artificial market simulation. We derive a Markov decision process (MDP) of CFD market makers over a discretized time horizon and train the CFD market maker, a deep reinforcement learning agent, using an artificial market simulation and analyze the details of the strategy using the proposed method. The proposed method is based on the design of reward function that contains the factors market makers should consider: spread, hedging costs, and inventory, and it simultaneously achieves risk management and profit maximization while providing liquidity as a market maker. Our model outperformed the baselines including the traditional market making strategy without hedging and deep reinforcement learning model that uses profit and loss (PnL) as a reward function. Also, our model was found to use two different risk management tools, limit price adjustment and hedging, based on the amount of inventory, market trend and volatility, and expected hedging cost.

Financial documents are increasing year by year, and natural language processing (NLP) techniques are widely applied to process these documents. Specifically, Transformer-based pre-trained models such as BERT have been successful in NLP in recent years. These cutting-edge models have been adapted to the financial domain by pretraining with financial corpora, while most financial pre-trained models are BERT-based and do not benefit from the mechanisms of recent state-of-the-art models. Many Japanese models need to perform word segmentation based on morphological analysis, which may reduce portability and processing efficiency. In this study, we show that models without pre-word segmentation have no performance disadvantage over models with pre-word segmentation in both the financial and general domains, while the models without pre-word segmentation may benefit from reduced computational complexity and increased model processing efficiency due to a reduced number of tokens. The length of tokens had little effect on the performance of the downstream classification tasks. We build both general models and financial models without pre-word segmentation on a large scale. We show that our financial pre-trained models perform better than conventional models on classification tasks in the financial domain, and that the general models can be good baselines to adapt to specialized domains. Our evaluation experiments show that additional pre-training is effective because it takes advantage of the performance of models constructed from large general corpora. We have released the pre-trained models at https://huggingface.co/izumi-lab.

Deep Hedging has garnered attention as a novel approach utilizing deep learning to address challenges in pricing and hedging in incomplete financial markets．However, the effectiveness of Deep Hedging when applied to multiple options has not been thoroughly examined．Additionally, financial market data is often noisy, making it challenging to train deep learning models effectively. Therefore, in this study, we aim to address these issues by evaluating the effectiveness of Deep Hedging for multiple options using data from the Bitcoin options market．We verify its effectiveness for multiple options and assess the impact of introducing smoothing techniques. Specifically, we introduce a technique called Deep Smoothing to reduce noise and prevent arbitrage opportunities when dealing with a portfolio composed of multiple European options with the same underlying asset, the same maturity, but different strike prices．We combine this smoothing technique with the structure of the Implied Volatility Smile（IVS）to propose a new framework of Deep Hedging for multiple options. We validate our empirical results with Bitcoin options market data, demonstrating that: （1） Deep Hedging outperforms traditional delta hedging, （2） when hedging multiple options, our method achieves performance equal to or better than conventional Deep Hedging targeting a single option, and （3） the application of Deep Smoothing to the input data leads to improved hedging performance．

One measure of financial markets is order book imbalance (OBI), defined as the number of buy orders minus the number of sell orders in the order book around the best quote. Thus, OBI is an indicator of an imbalance in supply and demand, which can potentially impact the price as price movements in the financial market. There is known to be a positive correlation between OBI and returns, and some investors attempt to exploit this characteristic to improve their investment performance. In financial markets, an investor sometimes want to place a large order. When the investor places the large order at one time, some other investors find out someone wants to trade large amount and trade ahead of time, or the market price gets rough and losses are incurred. Therefore, execution algorithms are used, by which a machine automatically divides orders into smaller lots to avoid distorting the market price. Execution algorithms could be expected to achieve improved performance when OBI is taken into account, but no such findings have been reported so far. One reason for the lack of research findings is that it is virtually impossible to measure the impact of OBI alone on the performance of execution algorithms, since there are many external factors that can affect financial markets. Multi-agent simulation is one way to solve problems that are difficult to analyse using empirical research and conventional methods. In a multi-agent simulation, individual actors in the world are regarded as agents and the behavioural rules and interactions of these agents constitute a model. In this context, a financial market constructed using multi-agent simulation is called an artificial market. In this study, we modelled an execution algorithm considering OBI, investigated how the model is affected by the market under several market patterns using artificial markets, and analysed the mechanism. The results showed that in stable markets, the performance of the execution algorithm with the OBI strategy varied with the number of orders placed. In contrast, in markets with unstable prices, the performance of the execution algorithm with the OBI strategy was higher than that of the conventional execution algorithm. Even in markets with manipulation by spoofing, the performance of the execution algorithm with the OBI strategy was not significantly less than that of the conventional execution algorithm, demonstrating that the model is not easily affected by spoofing.

In recent years, data generation techniques based on machine learning approaches have made significant progress and various generation models have been proposed in the fields of text and image. In the field of finance, there has also been research on the data generation of time series on stock prices. However, there are many problems in applying machine learning models to financial time series data. For example, the behavior of the machine learning model when given data that is not in the historical data is unclear, or if the data differs significantly from historical data under strong stresses such as the Lehman or the COVID-19 shock, the predictions by the machine learning model will be completely unreliable. This problem may be solvable by using data generation techniques that add virtual stress to the current state. In this study, we propose a new method for generating virtual time series data based on linear response theory (LRT). Using the LRT, we can obtain an approximate representation of the transition from the current equilibrium state to another equilibrium state by adding second-order fluctuation in the current equilibrium state and external forces (stresses). In other words, if the external force can be estimated in advance, we shall be able to obtain virtual time series data under external forces. As an application, we examined whether this method is effective with data augmentation of the stock prediction model using external forces estimated by historical data. As a result, the accuracy of stock price predictions was improved over the case without data augmentation.

In network analysis, traditional centrality measures are now complemented by the concept of curvature from differential geometry, which offers a refined understanding of network structures. Ricci curvature, an indicator of space distortion, is well-studied within the field of Riemannian geometry but has recently been adapted to discrete networks through the development of discrete curvature. Among the forms of discrete curvature, LLY Ricci curvature and Forman Ricci curvature are notable. LLY Ricci curvature, grounded in optimal transport theory, correlates with link density and clustering coefficients, offering a means to identify globally central nodes. Despite its effectiveness in diverse domains, including financial networks, its complex nature prevents its application to large networks due to computational constraints. On the other hand, Forman Ricci curvature, simpler and computationally more efficient due to its reliance on degree centrality, proves useful in identifying network bottlenecks but is limited by its sole dependence on network degree, constraining its applicability in constructing intricate graph algorithms. Therefore, this paper introduces a new graph algorithm that integrates both LLY and Forman Ricci curvatures, aiming to leverage their strengths while mitigating their limitations. To empirically validate the practical applicability, we applied our method to the financial market for stock screening, a critical process for investors. We conducted an empirical analysis using a well-known financial benchmark dataset to examine the performance of our proposed algorithm from a risk management perspective. The empirical results suggest that the stocks selected using our method are a robust set that maintains low correlation under normal conditions and responds effectively to events such as financial crises.