Search

This study constructed a Japanese chat dataset for tuning large language models (LLMs), which consist of about 8.4 million records. Recently, LLMs have been developed and gaining popularity. However, high-performing LLMs are usually mainly for English. There are two ways to support languages other than English by those LLMs: constructing LLMs from scratch or tuning existing models. However, in both ways, datasets are necessary parts. In this study, we focused on supporting Japanese in those LLMs and making a dataset for training or tuning LLMs in Japanese. The dataset we constructed consisted of various tasks, such as translation and knowledge tasks. In our experiment, we tuned an existing LLM using our dataset and evaluated the performance qualitatively. The results suggest that our dataset is possibly beneficial for LLMs. However, we also revealed some difficulties in constructing LLMs in languages other than English., Comment: 12 pages

A lead-lag effect in stock markets describes the situation where one (leading) stock return is cross-correlated with another (lagging) stock return at later times. There are various methods for stock return forecasting based on such a lead-lag effect. One of the most representative methods is based on the supply chain network. In this research, we propose a stock return forecasting method with an economic causal chain. The economic causal chain refers to a cause and effect network structure constructed by extracting a description indicating a causal relationship from the texts of Japanese financial statement summaries. We examine whether the lead-lag effect spreads to the 'effect' stock group when there is a large stock fluctuation in the 'cause' stock group in the causal chain. We confirm the profitability of the proposed strategy and the evidence of stock return predictability across causally linked firms in the Japanese stock market.

Organizingcompaniesbyindustrysegment(e.g.artificial intelligence, healthcare or fintech) is useful foranalyzingstockmarketperformanceandfordesigning theme base investment funds, among others. Current practice is to manually assign companies to sectors or industries from a small predefined list, which has two key limitations. First, due to the manual effort involved, this strategy is only feasible for relatively mainstream industry segments, and can thus not easily be used for niche or emerging topics. Second, the use of hard label assignments ignores the fact that different companies will be more or less exposed to a particular segment. To address these limitations, we propose to learn vector representations of companies based ontheirannualreports. Thekeychallengeistodistill the relevant information from these reports for characterizing their industries, since annual reports also contain a lot of information which is not relevant for our purpose. To this end, we introduce a multi-task learning strategy, which is based on fine-tuning the BERT language model on (i) existingsectorlabelsand(ii)stockmarketperformance. Experiments in both English and Japanese demonstrate the usefulness of this strategy.