Journal of Finance and Data Science最新文献

We document frequent occurrences of negative conversion premium (NCP) events in the Chinese convertible bond market, when the bond is convertible and the underlying stock can be freely sold. This implies that when an NCP event occurs, existing stock holders can earn a riskless profit through a long-short strategy which sells the underlying stock and buys the convertible bond at the same time, then converts the bond into stocks. Facing short sale constraints, traders not holding any position in the underlying stock can still profit from an overnight trading strategy which buys the convertible bond at the NCP event day t, then sells the converted stock on day t + 1. We also find that the next-day opening prices following NCP events are significantly lower, which is evidence for the stock selling from the overnight trading strategy. Overall, our findings show that investors in China are aware of the NCP events and they earn abnormal returns through active trading. However, it remains as a puzzle why existing stock holders such as institutional investors do not trade away the negative conversion premium through the riskless long-short strategy.

Predicting future pairs of the acquirer and acquiree companies is important for acquisition or investment strategy. This prediction is a challenging problem due to the following requirements: to incorporate various non-financial factors and to address the lack of negative samples. Concerning the former, we proposed including a network feature that represented the importance of an acquirer and an acquiree in the investment and category networks, as well as a company relation feature associated with their similarity and closeness. Considering the latter requirement, as negative examples, we set the pairs of acquirers and acquirees with the features that were similar to those of positive examples. This allowed learning minor differences between the companies selected for acquisition and the candidate ones. We evaluated our proposed prediction model using 2000–2018 acquisition logs collected from CrunchBase. Based on the analysis of the high SHapley additive explanation (SHAP) value features, we found that the newly considered network and company relation features had high significance (10 out of 22 top key features). We also clarified how these novel features contributed to the prediction of acquisition occurrence by interpreting the SHAP value.

In this paper, we find that the conflict of interest between loan holders and bondholders is positively related to bond IPO underpricing, which serves as a compensation to the initial bond investors. We construct four proxies for the conflict between loan holders and bondholders, namely, a loan covenants index, the outstanding loan amount, the number of lead banks, and the loan remaining maturity. Our empirical tests show that all four variables are positively related to bond IPO underpricing, indicating that the loan structure of firms has a real impact on the pricing of their bond IPOs.

Can deep reinforcement learning algorithms be exploited as solvers for optimal trading strategies? The aim of this work is to test reinforcement learning algorithms on conceptually simple, but mathematically non-trivial, trading environments. The environments are chosen such that an optimal or close-to-optimal trading strategy is known. We study the deep deterministic policy gradient algorithm and show that such a reinforcement learning agent can successfully recover the essential features of the optimal trading strategies and achieve close-to-optimal rewards.

We construct the French nominal yield curve using Svensson³³ methodology and all available public data of French nominal government debt securities—Obligations Assimilables du Trésor (OATs). Our sample period starts in October 1987 and ends in April 2018. We find that the functioning of the French sovereign bond market has improved dramatically following the onset of the euro area and has been functioning reasonably well since then, with the exceptions of the Global Financial Crisis period and the European sovereign crisis period. We also find that, the French nominal on-the-run securities have, on average, a negligible liquidity premium, in sharp contrast to the U.S. nominal Treasury market, where such a premium is sizable. On average, the level and the slope of the French zero-coupon rates have been decreasing since the Global Financial Crisis.

This paper presents the first-ever comparison of the forecasting power of two types of narratives: articles in a major daily newspaper and regular research reports released by professional forecasters. The applied testing methodology developed in ²² and extended in this paper includes two natural language processing (NLP) techniques – the sentiment analysis and the wordscores model – that are used to convert the text corpora into the NLP indices. These indices are explanatory variables in linear regression, Granger causality test, vector autoregressive model and random forest model. The paper extends this methodology by applying Latent Dirichlet Allocation (LDA) to the newspaper corpus to filter out articles that discuss topics not relevant for economic and financial analysis. The forecasting test is conducted for two major banks in Poland – BZ WBK and mbank and for major daily newspaper Rzeczpospolita, in Polish. It is shown that mbank narratives have the best forecasting power, while BZ WBK and Rzeczpospolita trade second and third place depending on the model applied. In the vast majority of analyzed cases adding an NLP index to the model improves the forecast accuracy. The answer to the title question is – it depends. Before paying for economic research asset managers are advised to apply methods such as presented in this paper to evaluate whether sell-side research offers any forecasting value in comparison with a newspaper.

This paper investigates the asset pricing implications of rollover risk, i.e., the risk that firms might not be able to refinance their due liabilities. I find that firm-specific rollover risk coupled with deteriorating credit market conditions significantly increase firms' credit spreads. During the 2008–2009 financial crisis period, the one-year CDS spreads for high rollover risk firms are 32–72 basis points higher than the spreads of low rollover risk firms. Longer maturity CDS spreads show similar patterns with smaller magnitudes. During normal periods, however, CDS spreads are mostly explained by fundamental variables and rollover risk is not a significant determinant. Similar rollover risk effect is also observed in other financial markets, including corporate bond, stock, and options markets.

In this paper, the daily returns of the S&P 500 stock market index are predicted using a variety of different machine learning methods. We propose a new multinomial classification approach to forecasting stock returns. The multinomial approach can isolate the noisy fluctuation around zero return and allows us to focus on predicting the more informative large absolute returns. Our in-sample and out-of-sample forecasting results indicate significant return predictability from a statistical point of view. Moreover, all the machine learning methods considered outperform the benchmark buy-and-hold strategy in a real-life trading simulation. The gradient boosting machine is the top-performer in terms of both the statistical and economic evaluation criteria.

We propose a machine learning toolkit applied to the detection of rare events, namely banking crises. For this purpose, we consider a broad set of macroeconomic series (credit-to-GDP gap, house prices, stock prices, inflation rates, long-term and short-term interest rates, etc.), in combination with their leads and lags, various filtering methodologies, and datascience models that complement time series analysis. The main advantages of the approach are its robustness, its flexibility and its prediction performance. Based on the best model specification, our methodology allows to compute an indicator for the probability of banking crisis along with an alert threshold up to 6 quarters ahead in real time for various developed economies.

There has been much debate about whether returns on financial assets, such as stock returns or commodity returns, are predictable; however, few studies have investigated cryptocurrency return predictability. In this article we examine whether bitcoin returns are predictable by a large set of bitcoin price-based technical indicators. Specifically, we construct a classification tree-based model for return prediction using 124 technical indicators. We provide evidence that the proposed model has strong out-of-sample predictive power for narrow ranges of daily returns on bitcoin. This finding indicates that using big data and technical analysis can help predict bitcoin returns that are hardly driven by fundamentals.