Research in both applied and theoretical concepts play a significant role within the Dataism Laboratory for Quantitative Finance (DLQF). Our work delves deeply into practical applications, bridging the gap between complex quantitative theories and real-world financial scenarios. This dual approach not only advances our understanding of market dynamics and financial instruments but also enhances our ability to develop innovative tools and methodologies. Through rigorous theoretical exploration and hands-on application, we strive to push the boundaries of what’s possible in quantitative finance, ensuring that our research remains both cutting-edge and highly relevant to the evolving needs of the industry.
Undergraduate Research
Apurv Sanjay Deshpande, Brett Noneman, Nicole Raphael,
Eeshan Umrani, Syon Ravala
Neural Network-driven Hypothesis Testing and Nonlinearity Analysis
In the last decade, there has been great improvement in enhancement regarding the analysis of data, enabled through rapid development in the field of machine learning and neural networks. Nevertheless, a noticeable gap remains in integrating these powerful tools into traditional statistical methods; this includes hypothesis testing and nonlinearity analysis in models. This project aims to bridge this gap by developing a comprehensive Python library that leverages the capabilities of neural networks to enhance hypothesis testing and nonlinear pattern detection. Central to our project is the understanding that a large number of datasets in the real world, particularly financial and time series data, contain complex nonlinear relationships that traditional linear models often fail to capture adequately. By integrating neural network modeling with robust statistical techniques, we will provide a truly powerful tool to those researchers and analysts who wish to test hypotheses about data distributions and seek elusive nonlinear patterns in data. The results emerging from this new approach will definitely increase the accuracy of predictive models and decision-making processes in quantitative finance and related fields by at least an order of magnitude, since their success strongly depends on successfully understanding the underlying structure of the data. The discussed library would not only be able to apply even more sophisticated data analysis methods but contribute to the changing landscape of data-driven research methodologies for financial data.
Raahul Esakiraja, Khyati Goyal, Trent Raymart Milagroso, Adhi Shankar, Benjamin Scott Sullivan
Enhancing CNN Efficiency for Time Series Forecasting
Time series forecasting plays acritical role in various domains, from finance to supply chain management. In recent years Convolutional Neural Networks (CNNs) have become more prevalent in time series forecasting due to their ability to understand “spatial” structure of continuous coordinates. A time coordinate works just as well as a spatial coordinate thus allowing CNNs to be leveraged in this field. However, deep learning models like NNs often face deployment challenges due to their high memory and computational demands. This paper explores a magnitude-based weight pruning technique applied to a pre-trained CNN, reducing model complexity by sparsifying low-magnitude weights and reapplying to model structure to create a compact model. We evaluate the method using metrics like memory usage, execution time, and predictive accuracy. Our results reveal that significant model compression can be achieved without substantial loss inaccuracy, offering a practical solution for real-world forecasting applications.
Varun Budati, Alexander Ardaiz, Sahana Sarathy
Order Execution and Optimization
Our research focuses on traditional execution models that combine neural networks and reinforcement learning methodologies, primarily built upon the hybrid approach to the Almgren-Chriss model, presented by Hendricks and Wilcox in their paper, ’A Reinforcement Learning Extension to the Almgren-Chriss Framework for Optimal Trade Execution’. In order to create and utilize a non-parametric execution strategy, we choose a model suggested by Macr`ı and Lillo in ’Reinforcement Learning for Optimal Execution when Liquidity is Time-Varying’, applying reinforcement learning (RL) techniques such as Deep Q Learning (DQL), Double Deep Q Learning (DDQL) and Proximal Policy Optimization (PPO). These models are then tested on historical data to analyze their performance and compared to the ’Stochastic Impact Model’ by Barger and Lois, as well as baselines Time Weighted Average Price (TWAP) and Volume Weighted Average Price (VWAP). Finally, we benchmark these results against
the Almgren-Chriss model to demonstrate how reinforcement learning can contribute to improvements in trade execution and market making for equities.
Graduate Research
Jamshid Ardalankia
Analysis of the Global Banking Network by Random Matrix Theory
Since 2008, the network analysis of financial systems is one of the most important subjects in economics. In this paper, we have used the complexity approach and Random Matrix Theory (RMT) for analyzing the global banking network. By applying this method on a cross border lending network, it is shown that the network has been denser and the connectivity between peripheral nodes and the central section has risen. Also, by considering the collective behavior of the system and comparing it with the shuffled one, we can see that this network obtains a specific structure. By using the inverse participation ratio concept, we can see that after 2000, the participation of different modes to the network has increased and tends to the market mode of the system. Although no important change in the total market share of trading occurs, through the passage of time, the contribution of some countries in the network structure has increased. The technique proposed in the paper can be useful for analyzing different types of interaction networks between countries.
Tushar Deshpande
Dynamic Options Pricing: Machine Learning approaches
Options pricing is a vital aspect of financial markets. It is used in risk management, investment strategies, and market efficiency along with applications in many different areas as well. Traditionally options are priced using the famous Black Scholes model which is used mostly for European options, the Monte Carlo simulation and the Finite difference methods which is used for American options, among many others. These models often try to simplify the process, which may not fully capture the real-world complexities. The rapid advancements in machine learning and deep learning have sparked significant interest in applying the sophisticated techniques to complex financial challenges as well, particularly in the domain of options pricing. These techniques present a promising avenue to address shortcomings like finding
non-linear patterns in economic data, efficiently processing and analyzing large multi-feature datasets, and real-time dynamic responses to changing and volatile market dynamics. This paper aims to study the potential applications of machine learning techniques in options pricing and try to create an optimal model that can price options accurately.