This study aims to understand the contagion effect of the major equity market sentiment events, defined as jumps in the VIX index, on cryptocurrency price jumps and the corresponding feedback effect on investors' sentiment. Using recent high frequency intraday data with multivariate Hawkes processes, we find that several noteworthy contagion effects exist between Bitcoin and the market sentiment proxied by the VIX index. First of all, we find that positive market sentiment jumps tend to trigger a moderate level of cross-contagion on both positive and negative jumps in the Bitcoin market, whereas negative market sentiment jumps have no contagion effect on the Bitcoin prices. We also find that the Bitcoin market shows stronger positive self-contagion and cross-contagion effects than the equity market, in general. We also observe a lasting fear of missing out (FOMO) phenomenon in Bitcoin. It is supported by the evidence that the effect of positive jumps in the Bitcoin market lasts three times as long as that in both the equity market and the negative Bitcoin price jumps, and these positive jumps in Bitcoin may serve as a harbinger of equity market movement. These findings provide a better understanding of risk control and policy guidance for the cryptocurrency market.

Low-rank adaptation (LoRA) methods show great potential for scaling pre-trained general-purpose Large Language Models (LLMs) to hundreds or thousands of use scenarios. However, their efficacy in high-stakes domains like finance is rarely explored, e.g., passing CFA exams and analyzing SEC filings. In this paper, we present the open-source FinLoRA project that benchmarks LoRA methods on both general and highly professional financial tasks. First, we curated 19 datasets covering diverse financial applications; in particular, we created four novel XBRL analysis datasets based on 150 SEC filings. Second, we evaluated five LoRA methods and five base LLMs. Finally, we provide extensive experimental results in terms of accuracy, F1, and BERTScore and report computational cost in terms of time and GPU memory during fine-tuning and inference stages. We find that LoRA methods achieved substantial performance gains of 36% on average over base models. Our FinLoRA project provides an affordable and scalable approach to democratize financial intelligence to the general public. Datasets, LoRA adapters, code, and documentation are available at https://github.com/Open-Finance-Lab/FinLoRA

We introduce FinSTS, a novel dataset for financial semantic textual similarity (STS), comprising 4,000 sentence pairs from earnings calls and SEC filings. To improve models for the Financial STS task, we propose an active learning (AL) algorithm that efficiently selects informative sentence pairs for annotation by GPT-4 and creates high-quality training data. Using this approach, we train FinSentenceBERT, a model that generates semantic embeddings specifically for financial text. FinSentenceBERT establishes a new performance benchmark on FinSTS, outperforming models that use basic pooling strategies or are fine-tuned on general datasets. Surprisingly, a general SBERT model trained using our AL approach surpasses even models based on FinBERT, a language model pre-trained on financial text. Our research contributes a specialized dataset, model, and methodology that advance semantic understanding in the financial domain, with potential applications to other specialized domains.

News sentiment is different from the true investor sentiment, and there is a conductive process of information flow from news sentiment to the latent investor sentiment and vice versa. This study aims to develop a methodology to estimate the latent effect between the investor sentiment jumps and the market return jumps using a multivariate Hawkes process along with a deep reinforcement learning algorithm. We achieve this goal through a three-step process: (i) identify the baseline intensity among the events of news sentiment and market return by a multivariate Hawkes process; (ii) estimate the hidden effect that drives the movement of events of news sentiment and market return from the baseline intensity via deep reinforcement learning; (iii) reveal the interaction mechanism among the true investor sentiment and the market return that is responsible to the latent investor sentiment. This approach can be broadly applied to analyzing many phenomena in finance and economics where latent events are non-stationary and can not be observed directly.

eXtensible Business Reporting Language (XBRL) has attained the status of the global de facto standard for business reporting. However, its complexity poses significant barriers to interpretation and accessibility. In this paper, we present the first evaluation of large language models’ (LLMs) performance in analyzing XBRL reports. Our study identifies LLMs’ limitations in the comprehension of financial domain knowledge and mathematical calculation in the context of XBRL reports. To address these issues, we propose enhancement methods using external tools under the agent framework, referred to as XBRL-Agent, which invokes retrievers and calculators. Extensive experiments on two tasks - the Domain Query Task (which involved testing 500 XBRL term explanations and 50 domain questions) and the Numeric Type Query Task (tested 1,000 financial math tests and 50 numeric queries) - demonstrate substantial performance improvements, with accuracy increasing by up to 17% for the domain task and 42% for the numeric type task. This work not only explores the potential of LLMs for analyzing XBRL reports but also augments the reliability and robustness of such analysis, although there is still much room for improvement in mathematical calculations.

We examine the risk factors disclosed in the 10-K financial statement section 1A across 9 years with over 500 hundred companies. We propose a financial disclosure risk factor to extend the Fama-French 3 factor model and Fama-MacBeth cross-section regression. Using the risk factors data from 2015 to 2023, we find the average risk-return premium across nine sectors is significant after controlling for other risk factors from the Fama-French 3-factor model. The premium is measured by monthly return series on risky-minus-less risky stocks or by the coefficient of stock risk factor estimated from cross-section Fama-MacBeth regressions. These text risk factors can potentially be used to construct portfolios that can generate significant returns across different sectors.

This paper investigates how counterparty risk and precautionary liquidity concerns can lead to dry-up of liquidity in the interbank market with adaptive bank agents sharing common objectives of collecting as much interest payment while minimizing the insolvency risk. The agents are calibrated based on the US bank balance sheet data and are allowed to make autonomous decisions under different simulated market conditions. Through simulation studies, we observe an endogenously formed multi-layer lending network exhibiting the well-known core-periphery structure. Moreover, we find the adaptive learning banks would endogenously form liquidity hoarding phenomenon under an exogenous shock to a subset of the banks in the system. We also find that in an adaptive learning environment, fire sales would lead to a decrease in interbank liquidity and increase the probability of contagion through interbank lending networks.

We propose a life-cycle model, where individuals facing uninsurable labor income risks choose whether to participate the stock market and make decisions on home-ownership, in an environment with social safety net and the retirement savings system. The model is motivated by the empirical finding that active stock market participation is associated with higher level of education and employment experience in the finance sector. The model exhibits good fit on portfolio choice, home-ownership, consumption pattern in the cross- section and through life-cycle. We find that the lack of stock market access, proxied by a fixed entry cost and variable costs, plays a key role in generating heterogeneous outcome in agent’s wealth accumulation, and therefore, increases wealth inequality.

With the increasing use of big data, cloud computing, and machine learning in high-stake domains such as justice systems, financial institutions, and healthcare, concerns about fairness have become more prominent. This paper presents a novel approach to foster fair decision-making by tackling social bias and enhancing transparency in machine learning models. The proposed framework leverages quantum-inspired complex-valued neural networks and attention-based networks, offering improved transparency in modeling the decision process for interpreting feature importance and dependency. Furthermore, our approach tackles the challenges posed by imbalanced data through the incorporation of focal loss and oversampling techniques, resulting in reduced prediction errors. Through extensive experiments conducted on real-life datasets encompassing criminal charge prediction, financial fraud detection, and credit card default payment prediction, our approach consistently demonstrates reliable prediction precision and recall. Notably, our analysis of feature significance highlights the statistical importance of task-related features such as historical records of bank transactions or criminal charge history, while socially biased identifiers like race, gender, and age exhibit minimal significance. By excluding these biased features, our approach enhances fairness without compromising prediction accuracy, thereby contributing to the advancement of fair decision-making in big data and cloud computing across various high-stake domains.

This study investigates comparability of firms' financial statement presentation, that is, the display of line items on financial statements. The line items that are displayed on financial statements are the result of disclosure choices such as placement, formatting, and aggregation. Given the emergence of machine-readable data (for example, XBRL-tags used in this paper in creating our test variable), a question arises about the relevance of items' presentation on the face of the human-readable financial statements. This paper examines the relation between line-item comparability (based on the face of financial statements) and financial statement users' judgments as proxied by attributes of analysts' forecasts. We posit that greater line-item comparability likely lowers analysts' processing costs and enhances analysts' ability to evaluate firms' economic performance. Our results indicate that greater line-item comparability is associated with more analyst coverage, increased forecast accuracy, reduced forecast dispersion, and greater timeliness of forecasts. Overall, these results suggest that comparability of line items on the face of financial statement enhances informativeness.

Compared with investing in individual stocks, ETF investment is capable of diversifying the non-systematic risk or exposure to broad market or industry sectors. The aim of this paper is to develop a jump contagion modeling framework to understand the contagion effect of market jump events of energy sector ETFs using multivariate Hawkes process modeling approach. Through analyzing intraday high-frequency market data, we find that negative index jumps lead index price discovery processes, and their influences disappear faster than the positive index jumps in both the S&P500 and the crude oil futures. And on average, the self contagion in negative jumps is stronger than the self contagion in the positive jumps across all ETF groups. However, the ETFs focused on the master limited partnership (MLP) segment show less negative self contagion and relatively stronger positive self contagion than the other energy ETFs. Overall, the influence of negative jumps on ETFs from both the equity index and the energy future index is stronger than that of the positive jumps. And the influence of the equity index (S&P500) jump on ETFs lasts longer than that of the crude oil futures index (CLC1).

In response to recent financial crises, financialmarkets have experienced rapid and profound changes. For example, the belief that global banks are ‘too big to fail’ has awoken the Federal Reserve to lean toward passing the Dodd–Frank Wall Street Reform and Consumer Protection Act1 (https://www.congress.gov/111/plaws/publ2 03/PLAW-111publ203.pdf). The deep reasons that led to the 2008 financial crisis following overzealous lending to an overheated housing market remain an unresolved long-term problem. The creation of new insurance instruments against risky mortgage products and mortgage-backed securities further increased the complexity of the market. The Troubled Asset Relief Program (https://home.treasury.gov/data/troubled-assets-relief-pro gram) was proposed to deal with the aftermath of the crises. Classic issues in finance, such as systemic risk modelling, inevitably become more complex. Further, the advance in financial technology has enabled trading to take place at the microscopic level (see Huang and Li 2017). High-frequency trading means that market structure is presented in much more profound ways: the price formation processes and price discovery of financial assets are driven by information flows that are highly interactive and very fast moving. Other market features such as order book dynamics, liquidity provisions and resilience also affect price movements. Thus, trading has become both extremely complicated and dynamic. Markets that are interconnected through the new trading dynamic have become more complex and are exposed to a greater number of and more variable types of extreme events. For example, the rampant risk inherent in proprietary trading across Wall Street firms has prompted the regulators’ proposal of the Volcker Rule (https://www.sec.gov/divisions/marketreg/faq-volcker-rule-section13.htm) to potentially legislate to prevent banks from taking on too much risk and incurring the consequent risk of default. A major lesson learnt from these market events is that more rigorous scientific approaches, including modelling techniques, are urgently needed to understand and interpret complex market and financial phenomena. Despite their simplicity and past efficiency, classical asset pricing theory and risk modelling have limitations in their ability to capture the new features of contemporary finance in today’s markets. For instance, stock prices almost certainly do not evolve as a random walk and return distributions often deviate far from the normal. Features observed in return time series such as fat tails, high peaks and extreme skewness are accepted stylized facts of financial markets in the post-crisis era. The occurrence of mini-flash crashes or so-called black swan events is no stranger to traders. They result in many practitioners seeking to neutralize their positions before the closing bell of a trading day. Price movements are also deeply influenced by news that often dominates investment sentiment and leads to either excessive risk-taking or contagion.

Spoofing has been identified a form of market manipulation, and it is harmful to the stability of the financial market. However, the effect of spoofing activity is hard to analyze due to its complex interactions within the market and lack of data. This paper presents an agent-based simulation model of the continuous double auction market to replicate and analyze the market dynamics under spoofing conditions. The simulated market consists of fundamentalist, chartist, zero intelligence agents, and spoofing agents where several existing market stylized facts are validated. The results show that in the presence of the spoofing agents and their market manipulation activities, the market volatility would increase, and spoofing activities would exacerbate the price variations. The fundamentalist agents would suffer a loss during the spoofing period but would be able to make profit during the price recovery phase. The chartist agents would suffer a loss when the spoofing agent realized its profit and the price recovery process start, at which they falsely believed the price movement trend would continue. The Sharpe ratio analysis also indicates the market manipulation activities of the spoofing agent would give themselves an unfair advantage resulting in a significantly higher Sharpe ratio than the other agents.

The Efficient Market Hypothesis has been well explored in terms of daily responses to market movements and financial reports. However, there is lack of evidence about information efficiency after the popularization of intraday trading. We investigate the time series properties of information adopted in the intraday market, in particular the causality effects. We use 30-min market price and news data to represent the past market data and the public information respectively, so that our analysis is in line with the EMH framework. Traders’ responses to such information are associated with the financial crisis. There was strong overreaction to market data right before the 2008 crisis and traders tend to rely more on news data during the crisis. We confirm that, in terms of the intraday information efficiency, it is worthwhile to adopt both types of information. Furthermore, there is still room for improving the price discovery process to reveal such information more effectively.

In this study, we use entropy-based measures to identify different types of trading behaviors. We detect the return-driven trading using the conditional block entropy that dynamically reflects the “self-causality” of market return flows. Then we use the transfer entropy to identify the news-driven trading activity that is revealed by the information flows from news sentiment to market returns. We argue that when certain trading behavior becomes dominant or jointly dominant, the market will form a specific regime, namely return-, news- or mixed regime. Based on 11 years of news and market data, we find that the evolution of financial market regimes in terms of adaptive trading activities over the 2008 liquidity and euro-zone debt crises can be explicitly explained by the information flows. The proposed method can be expanded to make “causal” inferences on other types of economic phenomena.

The potential impact of interconnected financial institutions on interbank financial systems is a financial stability concern for central banks and regulators. A number of algorithms/methods have been developed to extrapolate latent interbank risk exposures. However, most use highly stylized network models and reconstruction methods with global optimality lending allocation approaches such as maximizing entropy or minimizing costs. This paper argues that U.S. bank lending and borrowing decisions are largely suboptimal and performance-driven. We present an agent-based model to endogenously reconstruct interbank networks based on 6,600 banks' decision rules and behaviors reflected in quarterly balance sheets. The model formulation reproduces dynamics similar to those of the 2007-09 financial crisis and shows how bank losses and failures arise from network contagion and lending market illiquidity. When calibrated to post-crisis data from 2011-14, the model shows the banking system has reduced its likelihood of bank failures through network contagion and illiquidity, given a similar stress scenario.

This paper studies the impact of interbank risk exposures due to fire sales in the interbank lending market. We incorporate fire sale behaviors in a large-scale agent-based model that represents dynamic U.S. interbank lending system. We compare the simulated interbank networks with and without fire sales. The model shows that fire sales reduce the number of bank defaults due to financial shocks, while increasing the likelihood of contagion. Moreover, we investigate the implications of the capital requirements imposed by the Basel regulation by conducting a sensitivity test on capital ratios. We find the higher capital requirements may result in larger number of bank defaults.

This paper investigates how learning agents perform in inter-bank lending markets. Using policy gradient reinforcement learning agents, we demonstrate the adaptive learning banks can endogenously form liquidity hoarding phenomenon under an exogenous shock to the system. We also simulate the relationship lending using temporal difference, finding enhanced inter-bank lending relationship during financial crisis. Then we incorporate fire sales and study how fire sales would affect inter-bank lending market dynamics, and the results show that fire sales amplify liquidity hoarding under a distressed shock.