My research agenda is to answer research questions in the areas of regulations, political economy and gender norms by developing novel artificial intelligence techniques such as word embeddings, supervised learning, deep learning etc.
Please download my CV here.
Regulatory Costs and Market Power (Job Market Paper)
[ Abstract ]
Market power in the US has been rising over the last 40 years. However, the causes remain largely unknown. This paper uses machine learning on regulatory documents to construct a novel dataset on compliance costs to examine the effect of regulations on market power. The dataset is comprehensive and consists of all significant regulations at the 6-digit NAICS level from 1970-2018. We find that regulatory costs have increased by $1 trillion during this period. Moreover, small firms face higher costs than large firms despite attempts from regulators and politicians to limit the burden on small firms. We document that an increase in regulatory costs results in lower (higher) sales, employment, markups, and profitability for small (large) firms. Regulation driven increase in concentration is associated with lower productivity and investment after the late 1990s. We estimate that increased regulations can explain 31-37% of the rise in market power. Finally, we uncover the political economy of rulemaking. While large firms are opposed to regulations in general, they push for the passage of regulations that have an adverse impact on small firms.
The Political Economy of Financial Regulation
Rainer Haselmann, Arkodipta Sarkar, Shikhar Singla and Vikrant Vig
[ Abstract ]
Using the negotiation process of the Basel Committee on Banking Supervision (BCBS), this paper studies the way regulators form their positions on regulatory issues in the process of international standard-setting and the consequences on the resultant harmonized framework. Leveraging on leaked voting records and corroborating them using machine learning techniques on publicly available speeches, we construct a unique dataset containing the positions of banks and national regulators on the regulatory initiatives of Basel II and III. We document that the probability of a regulator opposing a specific initiative increases by 30% if their domestic national champion opposes the new rule, particularly when the proposed rule disproportionately affects them. We find the effect is driven by regulators who had prior experience of working in large banks - lending support to the private-interest theories of regulation. Meanwhile smaller banks, even when they collectively have a higher share in the domestic market, do not have any impact on regulators’ stand - providing little support to public-interest theories of regulation. Finally, we show this decision-making process manifests into significant watering down of proposed rules, thereby limiting the potential gains from harmonization of international financial regulation.
Capital Requirements, Market-Making, and Liquidity
Rainer Haselmann, Thomas Kick, Shikhar Singla and Vikrant Vig
[ Abstract ]
We employ a proprietary transaction-level dataset in Germany to examine how capital requirements affect the liquidity of corporate bonds. Using the 2011 European Banking Authority capital exercise that mandated certain banks to increase regulatory capital, we find that affected banks reduce their inventory holdings, pre-arrange more trades, and have smaller average trade size. While non-bank affiliated dealers increase their market-making activity, they are unable to bridge this gap - aggregate liquidity declines. Our results are stronger for banks with a higher capital shortfall, for non-investment grade bonds, and for bonds where the affected banks were the dominant market-maker.
Rainer Haselmann, Shikhar Singla and Vikrant Vig
[ Abstract ]
We exploit the establishment of a supranational supervisor in Europe (the Single Supervisory Mechanism) to learn how the organizational design of supervisory institutions impacts the enforcement of financial regulation. Banks under supranational supervision are required to increase regulatory capital for exposures to the same firm compared to banks under the local supervisor. Local supervisors provide preferential treatment to larger institutes. The central supervisor removes such biases, which results in an overall standardized behavior. While the central supervisor treats banks more equally, we document a loss in information in banks’ risk models associated with central supervision. The tighter supervision of larger banks results in a shift of particularly risky lending activities to smaller banks. We document lower sales and employment for firms receiving most of their funding from banks that receive a tighter supervisory treatment. Overall, the central supervisor treats banks more equally but has less information about them than the local supervisor.
Measuring Institutional Strength and Efficiency: Evidence from 200 Years of Legal Cases
Mayukh Mukhopadhyay, Arkodipta Sarkar and Shikhar Singla
[ Abstract ]
We apply machine learning techniques to 200 years of US legal judgements to measure two major components of institutional quality: the strength of creditor and property rights and the efficiency of courts (measured by the average waiting time per case). Our paper is the first to measure the distribution of institutional quality across states as well as its evolution across time. We show that creditor and property rights have declined steadily across time, which supports narrative evidence from legal scholars on this topic. Court efficiency on the other hand has steadily improved with a 64% decline in waiting times in US Federal Courts from 1900 to 2015. We also provide a framework to analyse the strength of other institutions (e.g., labour regulation) using text data from legal cases.
Gender Norms Do Not Persist But Converge Across Time
Shikhar Singla and Mayukh Mukhopadhyay
[ Abstract ]
We investigate the evolution of gender norms for 160 years in the US. Socioeconomists have posited two fundamental and widely debated theories on the evolution of cultural norms across time. One argues that cultural norms should converge across time as economies become more advanced and integrated, whereas the other states that cultural traits are highly persistent, passed down from generation to generation. These theories remain untested due to a lack of granular and high-frequency data over a longer time period. We develop a novel unsupervised machine learning methodology and apply it to 193 million pages of local newspaper text to produce localised attitudes towards women on four dimensions: career vs family, attitudes towards abortion, attitudes towards feminism/suffrage, and violence against women. We establish novel facts on the evolution of attitudes across time. First, attitudes are less persistent than the existing literature hypothesises. Second, the persistence varies considerably across regions and dimensions. Third, attitudes exhibit cyclical patterns. Fourth, regional variation in attitudes decreases considerably over time and has fallen between 64% to 79%. Fifth, a decrease in transport costs that allows for easier information sharing is associated with a homogenisation of the norms.
Machine Unlearning Human Biases: Inclusive Word Embeddings by Excluding Biased Text
[ Abstract ]
Word embeddings exhibit biases such as racial and gender biases due to the presence of these biases in the training corpus. Usage of these algorithms can increase the stereotypes in various contexts. We present a simple and generalizable approach of detecting the parts of a corpus that affect the bias and show how removing those parts can debias the word embeddings. The approach finds words that link the target words for a group and biased or attribute words (indirect bias). Unlike prior work, our approach a) removes the biases completely, b) removes indirect bias, and c) can be generalized to any type of bias, downstream task or word embedding model. We apply our methodology on Wikipedia and American National Corpus (ANC) for Word2Vec and GloVe models on the racial and gender biases. It is highly accurate in removing the biases without affecting the performance of the models in capturing semantic information.