Institute for Outlier Research in Business

Joe Raffie, Department of Management and Organization

Nan Jia, Department of Management and Organization

This project will study the evolution and emergence of gender equality (inequality) in the American legal services industry. To do so, we will build and examine a novel database that we are constructing from historical law firm directories. This database spans a near century of time,starting in the year 1870, a time when women in the USA did not have the right to vote and when virtually no women were licensed attorneys, and ending in the year 1962, at which time significant changes and advancements had been made (albeit not enough). The historical nature of this data and the time span it covers (92 years), provides an unusual and unique opportunity to examine the factors which contributed to (or detracted from) changes in gender equality, starting at a point in time when female labor force participation in this industry was effectively zero. Our goal is to learn from history in a way which can be applied to the plethora of equality issues we currently face and to generate knowledge which can be used to promote and enhance gender equality, diversity, and inclusion, within modern firms.

Rodney Ramcharan, Department of Finance and Business Economics

The motivation for the data collection begins with the fact that influential theories in macroeconomics and finance predict that the structure of financial intermediation and the balance sheet of intermediaries can have a powerful effect on economic fluctuations and asset prices (Brunnermeier and Sannikov (2014), (He & Krishnamurthy, 2013). These effects can however be ambiguous. Greater competition in credit markets can generate more efficient intermediation, reduce borrowing costs and relax credit constraints for marginalized borrowers. This can lead to faster economic growth in the long run—the “credit is good for growth” hypothesis. But more competition in the financial system can also erode the profitability of incumbent financial institutions, leading to riskier lending and more unstable banking. In the most extreme case, increased competition in credit markets can foster an ex-post misallocation of credit to riskier borrowers, producing asset price booms and busts, as well as profound and longlasting shifts in financial regulation (Mian and Sufi (2009), (Favara & Imbs, 2015), and (Raghuram Rajan & Ramcharan, 2015)).

Jason Lee, Department of Data Sciences and Operations

In the past decade, neural networks have made a remarkable impact on application domains including computer vision, robotics, and natural language understanding. Specific advances from Deep Learning include AlphaGo, a superhuman Go-playing AI, strategy game playing, and realistic chat-bots. At their core, these applications all involve learning the parameters of a neural network via the stochastic gradient descent algorithm. This proposal lays out a program to study provable learning methods in neural networks by separating the study into two interrelated problems:

(i) Optimization of Neural Networks. The loss function of neural networks is highly non-convex, yet standard SGD and its variants attain near global minima, as evidenced by zero training error. What properties (e.g. overparametrization) allow for gradient methods to converge to local or even global minima? Can we design the landscape by modifying the architecture and loss to allow for efficient training?

(ii) Generalization and Regularization in Neural Networks. Since commonly used neural networks are overparametrized, the model is able to perfectly interpolate the training data. In fact, there are innately many global minima that perfectly interpolate the training data. Common statistical wisdom suggests that most of these models will incur high generalization error; however, empirically SGD nds (near) global minima that generalize. Why does the stochastic gradient algorithm nd global minima that do not overt? Does the stochastic gradient algorithm induce an implicit regularizer? Can we isolate the regularizer and use it as an explicit regularizer to further improve generalization?

Joao Ramos, Department of Finance and Business Economics

Ala Avoyan, Indiana University

Economic situations often require agents to coordinate their actions, and coordination failures leading to under performance are pervasive in society. For instance, firms have to make investment decisions with uncertain returns that depend on the amount invested by other firms; thus, coordination failure may limit economic activity, with firms choosing suboptimal investment levels (see,for instance, Rosenstein-Rodan (1943)). Focusing inside the firm, consider a team that has to hand in a joint report before a deadline. Each team member is responsible for a section, and the report is complete only when all sections are delivered. A team member would be willing to put extra effort to hand in her part on time, but only if she was certain all others would do it as well.

Coordination failure may occur because, although group members have common preferences over outcomes, to achieve the best outcome requires taking an action they are unwilling to take, unless others do it as well. This makes strategic uncertainty a relentless feature: although superior outcomes (e.g. all putting extra effort) can be reached, given the uncertainty about others’ strategy, the risk of mis-coordination makes those outcomes unattainable.

Although the benefits of overcoming coordination failures are obvious, it is not clear how institutions can mitigate strategic uncertainty. Given that institutions are rarely random, this is a natural question to be answered from an experimental perspective. Recent experimental research focuses on communication institutions—comparing players sending public to private messages before choosing actions, or one-round of messages to many rounds—with mixed results for efficiency. Given the lack of theoretical implications for many of the institutions studied, it is unclear how a particular institution could help agents coordinate on higher outcomes. Furthermore, even if an institution improves coordination in the lab, it is unclear which institutional feature is responsible for the success.

WORKING PAPER

Clive Lennox, Leventhal School of Accounting

Due to a lack of publicly available data, there is relatively little evidence on how partners are incentivized and what audit partners do during an audit. In this project we will use three unique and proprietary data sources to peer into this black box:

1) the equity ownership stakes of audit partners
2) the identities of review partners and engagement partners, and
3) the audit adjustments that are made to reported earnings during the audit.

We expect to generate two papers using these data. First, we will examine how the ownership stakes of audit partners affect audit quality. Second, we will examine why internal control audits can have adverse consequences for financial reporting quality. As explained below, these two research questions are very important for regulators and practitioners, as well as the academic community.

Adel Javanmard, Department of Data Sciences and Operations

As we often hear, we are living in the era of data deluge. `Big data' technologies have allowed the acquisition of vast amount of fine-grained data and their accumulation into large scale databases, at an unprecedented speed. Powerful hardware and software systems have also been developed to crunch these data and extract information from them. Due to all these developments, the data-driven approach has become de rigueur nowadays in almost every field. Given the dataset, one of the off-the-shelf software packages is used to fit a statistical model which is then used for prediction, discovering new associations between variables (e.g. a specific demographic variable with the future income), clustering, policy design, decision making, and so on. However, this trend is a double-edge sword; The increasing complexity of these data and of the algorithms used has made statistical models significantly less interpretable. Employing the derived models without a proper understanding of their validity can lead to large number of false discoveries, wrong predictions and massive costs. Consider a concrete example where the medical records of patients are used to develop a model for providing personalized risk score for a chronic disease. A high risk score can trigger an intervention, such as incentive for healthy behavior, additional tests, and medical follow-ups which are all costly. Now, the question is how certain are we from the predictions made by this statistical model? What is their limit of validity? How biased is the resulting model? Closely related is the concern of reproducibility of the results. Researchers would like to know if their findings in a study can be successfully replicated in another study under the same conditions, not exactly but up to statistical error. Answering these questions for modern high-dimensional data has created a need for novel foundational perspectives on inferential thinking.