Prof. Anindya Chakrabarti

Financial markets exhibit non-trivial comovement and dependency structure. The standard approach in the finance literature is to consider the market in its aggregate form. A more recent 'data'-oriented approach emphasizes a more granular decomposition of the market so that the aggregate dynamics can be broken down into contributions arising from individual assets. This leads to two analytical problems. First, one has to necessarily deal with a large amount of data such that the process scales with the volume of data (large N and large T where T>>N). Two, analyzing such a large volume of data requires toolkits which are at the intersection of econometrics and machine learning. In this project, the goal is to construct large scale financial networks based on multivariate time series data to capture the dynamics of the system. The main idea is to provide an algorithmic approach to convert time series into networks such that the properties of time series are also inherited by the resulting network. The spectral structure of the comovement network is known to capture, at least partially, the booms and busts in the markets. Here, we take up two specific problems. One, how reliably does the spectral structure reflect the system for the case where T~N. Two, a large chunk of the literature on networks construction depends on bivariate modelling which is subject to failure due to multiple hypothesis testing. Therefore, an imminent question is how to construct a network with a direct multivariate model.

Prof. Anirban Banerjee

Over the last decade, the Indian market has seen significant growth in algorithmic trading and more specifically, high-frequency trading (HFT) activity. During this period, we have witnessed a significant change in the trading landscape as presently close to half of the trading volume in the stock exchanges is contributed by algorithms. This rise has not always been smooth as there have been calls for regulations to restrict algorithmic trading activity due to the fear of probable market manipulation.

Latency is considered one of the most important market parameters for HFTs. Using a large novel dataset of order and trade level data from the NSE, we would like to inspect how the latency in the Indian market has changed and if that has caused any shift in the way HFTs operate. We would also like to observe how the different market quality parameters have evolved over this time.

Prof. Ankur Sinha 

Hyperparameter tuning in the area of machine learning is often achieved using naive techniques, such as random search and grid search that only lead to an approximate set of hyperparameters. Although techniques such as Bayesian optimization perform an intelligent search on the domain of hyperparameters, it does not guarantee an optimal solution. A major drawback of most of these approaches is that as the number of hyperparameters increases, the search domain increases exponentially, thereby increasing the computational cost and making the approaches slow. The hyperparameter optimization problem is inherently a bilevel optimization task, and there exist studies that have attempted bilevel solution methodologies for solving this problem. These techniques often assume a unique set of weights that minimizes the loss on the training set. Such an assumption is violated by deep learning architectures. Our study is on gradient-based bilevel optimization method for solving the hyperparameter optimization problem. The method is general and can be easily applied to any class of machine learning algorithms that involve continuous hyperparameters.

Prof. Anuj Kapoor

Conversational AI Chatbots are rapidly evolving as new technologies with both business potential and customer reactance. This study exploits large-scale field experiment data on thousands of customers who are randomized to receive highly structured outbound sales calls from chatbots. We vary features like the gender of the bot along with the formal or informal tone of the bot. In this paper, we propose a dynamic outbound call experimentation policy. The proposed approach extends multi-armed bandit (MAB) algorithms, from statistical machine learning, to include microeconomic choice theory. Our automated outbound call policy solves this MAB problem using a scalable distribution-free algorithm. Beyond the actual experiment, we plan to counterfactual simulations to evaluate a range of alternative model specifications and allocation rules in MAB policies.

Prof. Anuj Kapoor

Artificial intelligence(AI) assisted tools are increasingly being used in health care contexts to provide advice and motivation. But whether AI can be a good or even better substitute for human involvement in these contexts is an open question. We provide empirical evidence to answer this question specifically in the context of fitness tracking mobile applications (apps). In addition to facilitating the tracking of activity and food intake, such apps provide advice and motivation in the form of targeted messages to their consumers, and this can be done through human coaches or an AI coach. An AI coach allows these apps to scale their offerings to a larger number of consumers, available on demand to consumers, and potentially more finely targeted by leveraging vast amounts of data. On the other hand, human coaches might be better placed to show empathy, and consumers might also feel more accountable to humans. We compare human and AI coaches on their effectiveness in helping consumers achieve their weight-loss goals. Our empirical analysis is in the context of a large-scale mobile app that offers consumers different levels of subscription plans with human and AI coaches respectively, and specifically compares adopters of the two kinds of plans on their weight loss and goal achievement. We address the potential self-selection in plans by employing a matching-based approach. We find, for our sample of almost 65000 consumers that human-based plans do better than those in AI-based plans in helping them achieve their goals, but that this differs by consumer characteristics including age, gender and body mass index (BMI).

Prof. Hyokjin Kwak

Major empirical methods: web scrapping, data pre-processing, independent t-test, machine learning classification method (CNN and ResNet-32).

This project aims to study how anthropomorphized ‘brand new’ and ‘used’ toaster products affect consumer purchase or bidding behaviour. To do this, I scrapped all the toaster data information like product name, number of consumers watching the product, bidding details etc., from eBay website, toaster products were then manually labelled as Anthropomorphized "AB" or Non- Anthropomorphized "Non-AB." Exploratory data analysis (EDA) was used to look at the attributes of the data. IBM SPSS software is also used to analyse the Independent sample t-test. This test compares the means of two independent groups, AB and Non-AB, to see if there is statistical evidence that the relevant attribute means are significantly different.

Also, multiple deep learning approaches were used to classify the AB and NAB toaster images.

Prof. Hyokjin Kwak

Major empirical methods: web scrapping, data analysis, sentimental analysis.

The objective of this study is to check the purchasing behaviour of consumers in response to creative advertisements that have been posted on Amul's twitter handles and to also apply sentimental classification techniques to the comments that have been posted on Amul's twitter handles. Initially, I compiled all available data on Amul's most successful advertisements from 2019 to 2021, which can be found on the company's website and neatly labeled according to year. Information about advertisements was then culled from the Amul Twitter account and matched with captions taken directly from the company's website. The properties of the data were investigated using exploratory data analysis (EDA) methods. IBM SPSS is also used to analyse the results of the independent sample t-test, which compares the means of two groups to determine whether or not there is statistical evidence of the relevant attributes are significantly different. In addition to that, a sentimental analysis was carried out on the user comments left on the advertisement that was uploaded to Amul's Twitter handle

(https://twitter.com/Amul_Coop)

Prof. Hyokjin Kwak

Major empirical methods: web scrapping, data pre-processing, sentimental and emotions analysis.

This research will compare the emotional and sentimental analysis of book reviews that were posted on Amazon.in (India) and Amazon.com (US) by taking into consideration of other characteristics such as book ratings, book cost, discounts available on the book, etc. This study will focus mostly on the sentiments and emotions expressed in Amazon USA and India book reviews. I explored various algorithms like VADER (Valence Aware Dictionary and Sentiment Reasoner), Textblob, SentiBERT etc. to detect the sentiments present in the product reviews also explored emotions like happy, fear, disgust, anticipation, joy, sadness, surprise, and trust.

Prof. Hyokjin Kwak

Major empirical methods: web scrapping, data pre-processing, text analysis which includes identifying bi-gram and trigram words

In this study, I will be performing an analysis of the textual contents, beginning with the reviews of the product and continuing all the way through to the purchase recommendation. In addition, a semi-automatic method will be utilised to extract terms from the text in reviews, and a knowledge graph will be utilised. The extracted phrases were connected to the various technical aspects of the items. After that, vector representations of the graph elements will be trained, which will result in a significant improvement in the overall quality of the recommendations. I intend to look into Adaptive Text Rank, which is based on a set of technical characteristics and a collection of sentiment words; the SOTA BERT model, which matches terms with the technical features of the products; the TransE method, which trains vector representations of graph elements; and the ABAE method, which highlights important characteristics for products based on a collection of reviews.

Prof. Hyokjin Kwak

Major empirical methods: web scrapping, data pre-processing, text analysis which includes identifying bi-gram and trigram words, collection of large labelled dataset of fake and non- fake reviews from official sources.

The most recent AI text generation models, such as GPT-2 and GPT-3, along with a variety of transformer models, can be utilised to generate fake computer-generated reviews. Researchers in the field of data science have demonstrated how artificial intelligence reviews may be generated and identified with the use of machine learning. Additionally, fake reviews can be purchased in bulk from various sources. Researchers have uncovered a wide range of potential traits that can help humans and models determine the difference between a fake review and a genuine one. These features were discovered after studying fake reviews in great detail. Because fake reviews frequently utilise the same language, the review content itself is generally considered to be the most important aspect. This is especially true in cases where multiple reviews were written by the same individual, firm, or other sources. Review length, Sentiment, Helpfulness, Reviews per user, and Verified review are some of the often-used features that can be noticed in studies on fake review detection. In order to create a model that is able to differentiate between the fake reviews that are posted on Amazon.com and Amazon.in, I will be conducting research into text-classification algorithms and NLP pre-processing techniques, such as count vectorization and TF-IDF, as well as machine and deep learning techniques

Prof. Hyokjin Kwak

Major empirical methods: web scrapping, data pre-processing, text analysis, Latent Dirichlet Allocation (LDA)

The technique of automatically identifying topics that are present in a text item and deriving hidden patterns that are represented by a text corpus is referred to as topic modelling. The usage of topic models is beneficial for a variety of tasks, including the clustering of texts, the organisation of huge blocks of textual data, the recovery of information from unstructured text, and the selection of features.

My goal here is to extract a set number of relevant word groups from the reviews based on sentiment, i.e., positive, negative, or neutral. These word groups are essentially the issues that will aid in determining what the customers are actually talking about in the reviews. This will inform us which subjects are frequently addressed by Indian and American reviews when they favour or dislike the product. This can be expanded upon in terms of emotions as well.

Prof. Jeevant Rampal

Auctions are often used to sell property rights for liquor licenses, spectrum licenses, land and mineral rights, and construction projects etc. This project investigates potential improvements in these auctions using a computational data-driven approach. The first part of this project will be to collect primary data of the participants and their choices in auctions. Subsequently, using the game-theoretic properties of the chosen auction design, we will computationally estimate the true (unobservable) value distribution across players of the object(s) being auctioned (e.g., liquor licenses). The estimation method used will be non-parametric “distance minimization” between the observed out-of-sample distribution of bids, and the predicted out-of-sample distribution of bids using optimally calibrated parameter values. E.g., Athey, Levin, and Seira (QJE 2011) use their estimated model to make comparative static predictions and test that for fit against data from timber auctions.

Finally, to analyse which auction design would have best met the various aims of the auction designer, we will use the calibrated model, parameters, and the estimated valuations of the bidders. In particular, using these we will simulate the revenue, efficiency, and other metrics of importance for different auction designs. In addition to the use of simulation described above, to analyse alternate auction designs, we will use simulations of variations of the estimated model, parameters (like risk aversion, budgets etc.), and value distributions to analyse the different rates with which different auction designs can meet the various possible aims of the auction designer.

Prof. Sachin Jayaswal 

We propose to study a multiperiod interdiction problem, in which the leader (attacker) with a limited interdiction budget decides the sequence of facilities to interdict (destroy) over time so as to inflict the maximum cumulative damage to the follower. The follower's objective is to serve a given set of demand points from the surviving subset of facilities the minimum cumulative cost across all periods. for this, his decisions include the assignments of demand nodes to the surviving facilities and the allocation of his limited budget to the revival of interdicted facilities and the protection of the surviving facilities against their interdiction in the future periods. The multi-period version of the problem, which is the focus of the proposed study, presents additional complexity due to the leader's interdiction decisions constrained by the follower's protection decisions. The objective of the proposed study is to design efficient exact solution methods for this challenging bilevel integer program.

Prof. Samrat Gupta 

The Information and Communication Technology (ICT) provides users unparalleled access to information from around the globe. In spite of demographic differences, people can communicate, express and evolve their opinions on topics ranging from politics to culture. The wide-ranging information exchange on digital media can lead to two scenarios viz. formation of public sphere or formation of echo chambers. While the public sphere, which promotes greater diversity, is a well-researched domain, substantially less research has been conducted on echo chambers in relation to socio-cultural events. Despite the huge affirmative impact of socio-cultural events on society, the proliferation of controversies around them and reinforcement through echo chambers is collectively having malefic effects on societies. Recent controversies around socio-cultural products such as movies, painting, books, cartoons, etc. resulted in serious outcomes. For example, Indian movie Padmavat brought polarization of public perception which further reinforced through echo chambers and escalated into widespread agitations. It led to mass destruction of property and human suffering during agitation. We believe this represents a mounting problem for society, one that is likely to intensify in the era of social media. Thus, understanding the causes, symptoms and consequences of socio-cultural polarization is critical and would be valuable for developing interventions to reduce unhealthy societal and organizational polarisations.

Prof. Sobhesh kumar Agarwalla and Prof Vineet Virmani

The proposed research project on modeling implied volatility (IV) and understanding the information content of option prices is part of our larger research agenda on studying ways to quantify uncertainty in financial markets, focusing on India. Traders in options markets do not usually quote option prices, but the volatility implied by them. IV is that volatility input to the famous Black-Scholes option pricing formula such that the Black-Scholes prices match the market price of the options. It has been observed that IV is not a constant but varies systematically with strike/delta and expiration date. The shape of the observed relationship between implied volatility and strike is called volatility smile or skew. In this project, we plan to explore various ways of modeling the dynamics of volatility smile using variants of state-space models and the Kalman Filter.The proposed research project on modeling implied volatility (IV) and understanding the information content of option prices is part of our larger research agenda on studying ways to quantify uncertainty in financial markets, focusing on India. Traders in options markets do not usually quote option prices, but the volatility implied by them. IV is that volatility input to the famous Black-Scholes option pricing formula such that the Black-Scholes prices match the market price of the options. It has been observed that IV is not a constant but varies systematically with strike/delta and expiration date. The shape of the observed relationship between implied volatility and strike is called volatility smile or skew. In this project, we plan to explore various ways of modeling the dynamics of volatility smile using variants of state-space models and the Kalman Filter.

Do conventional legal principles work, or may they need to adapt, when traders hand their affairs to computers operating by algorithm?

This question was posed specifically by Justice Jonathan Hugh Mance of the Singapore International Commercial Court (SICC) in a strong dissent in the seminal case of Quoine Pte Ltd v B2C2 Ltd [2020] SGCA(I) 02 (“Quoine v B2C2”). The case is about contracts formed in the context of algorithmic trading.

Broadly, two schools of thought have emerged in this debate.

One group argues that unenforceability of AI-negotiated contracts under current doctrines of contract law. The reason is on account of consent. Without consent a person cannot be bound by terms of the contract. The AI led contract formation without element of human involvement means there is no consent and thus could not be enforced legally. To overcome this, scholars argue “treating AI contracting programs as the legal agents of the people who operate them”. Explanation to this is that “A legal agent may negotiate and agree to a contract that binds their principal, even though the principal does not know or consent to the specific terms of that contract”.

Others argue that “agency is not the right way to understand the legal relationship between an AI program and the person who uses it. Agency law is a poor foundation on which to build the law governing artificial intelligence, because AI programs are not legal persons that can be responsible for themselves. Applying agency law to AI programs would create perverse incentives for the oversight of AI programs and would allow the people who benefit from AI programs to avoid liability for some of the risks created by those programs”.

The question that can be posed here and in the context of broader development of AI driven contracts as one understands, are two:

1. Does AI interaction in contracts leads to development of new legal principles or does the existing principle of contract is sufficient to accommodate AI related technology intervention in contracts law?
2. What are the legal changes/challenges one can expect in future contracting when dealing with AI?

Prof. Swanand Deodhar and Prof. Samrat Gupta

The project "Effects of Dynamic Rewards on Solver Efforts in Creative Crowdsourcing" aims to examine the effects of dynamic reward as an incentive mechanism in the creative contest landscape. Prior work has examined the role of reward, articulating a variety of effects. However, the typical approach is that the reward gets fixed at the start of the contest and does not change during the contest period. Our study aims to empirically relax this boundary condition. To this end, we plan to conduct a randomized field experiment in which rewards are announced at the start of the contest, and we offer different incentives during the contest. The findings from this study will contribute to the existing literature on crowdsourcing, creativity, and motivation, offering practical insights for both platform administrators and organizations leveraging creative crowdsourcing. The study will benefit the seekers (organizations) and the solvers (individuals). Ultimately, this research seeks to inform the design of various reward structures that enhance solver engagement, foster creativity, and optimize outcomes in the context of creative CCs.