Amey Joshi

It is a well-known fact that the team of portfolio managers influences the performance of a mutual fund. Investment professionals are known to hop organizations together. If a team member leaves one company to join another, the rest of the team does the same thing within a few months.

I used investment professionals' employment data to construct a time series of graphs and use them to discover cliques and professionals who always end up being in the same organization.

Certain happenings in a company significantly affect the price of its stock or bonds. Such occurrences could be a leadership change, new patent, lawsuit, fraud, or industrial dispute. Frequently, before the news is published in public media, the premonition is felt in the stock prices. We term these happenings as events. They happen infrequently, but if we can detect their presence, we can take advantage of the price movements to make investment decisions.

We used spectral analysis to identify the signature of rare events and correlated them with a news item announcing a significant event related to the company. It was impossible to accurately measure the time lag between the signal and the item's appearance. However, we confirmed that the signal was detected before the news appeared in the public domain.

Built a program that uses the blossom algorithm to help users find better matches on a dating platform. This algorithm computes a maximum matching in an undirected graph to find the most number of pairs. The program continues in several rounds, and the algorithm ensures that a pair from a previous round is not repeated. The participants just need to fill in information about themselves and describe desired partners in a speed-dating program.

Recurrent neural networks (RNN) can be viewed as numerical algorithms to solve ordinary differential equations.
I conducted a mathematical analysis of the RNNs using the framework of differential equations and showed that adding an attention layer to gated recurrent units does not alter their stability. The study was supplemented with Python code, demonstrating the same conclusion.

Built a time series of the incoming traffic in a call center to predict the number of agents needed to handle it and the probability of a new customer having to wait before their call is answered. I also predicted the average speed to answer because the regulators tend to focus on it.

A school wanted a schedule that conforms to the government-mandated hours for every subject, using the available classrooms, laboratories, and other facilities. When formulated as a constraint optimization problem, it had over 80 decision variables and over 700 constraints. I solved the problem using a CPLEX constraint optimizer.

Used the finite-difference time-domain (FDTD) technique to solve Maxwell equations to simulate the famous Young's double-slit experiment in Python. The simulation allows the users to change the geometrical parameters of the experiment and the nature of the source of electromagnetic radiation.

Built a dashboard in R (Shiny) to monitor the returns from a portfolio with alternative investments. The raw data was pulled from an Oracle database. It was used to compute financial metrics. The dashboard showed a trend of the metrics over a period whose extent could be adjusted.