Since Spring 2024, Bonsai ACG’s Neuroscience arm has been developing a computer vision pipeline to automate the classification of dendritic spines from Golgi-stained images. With neurodegenerative diseases such as Alzheimer’s Disease, Parkinson’s Disease and Huntington’s Disease affecting millions worldwide, a deeper understanding of how neurons communicate — or rather fail to do so — has become a very active field of research in recent years. It is for this reason that much significance has been placed on the neuronal structure that is responsible for receiving information from other neurons: dendritic spines. While both losses of and morphological changes in dendritic spines have been associated with neurodegeneration, existing methods to quantify dendritic spine morphology are labor intensive, subjective, and require specialized imaging technology.

Bonsai ACG’s software EZspine is being developed to address this need for an objective, easily integrable spine-analysis method. EZspine reads Golgi-stained Z-stacks, the most accessible and widely-used approach to image neurons for spine analysis. While Golgi-stained images are often noisier than images captured using alternative techniques, we have integrated numerous image-processing techniques with custom computer vision algorithms to reliably prepare even the most artifact-ridden images for spine analysis. By extracting only in-focus regions of these cleaned Z-planes, EZspine is effectively able to reconstruct dendrites in 3D-space using only a Z-stack of 2D Golgi-stained images. Following this reconstruction, EZspine segments spines from the dendritic shaft before classifying them into established morphological bins: “mushroom,” “stubby,” “thin,” etc. While existing approaches take anywhere from hours to weeks, this pipeline classifies dendritic spines in only a few minutes with minimal user intervention.

With this method being reliable, efficient, and using only the same Golgi-stained images that most researchers in this field capture, we hope to provide a spine analysis method that can be integrated into labs — and used to accelerate investigations into neurodegenerative disease — immediately. We plan to publish this method and release the open-source EZspine software this year.

Coming Soon.

Bonsai’s Quantitative Finance team had an exciting opportunity to participate in IMC’s Prosperity-2, a highly competitive, 15-day long trading simulation. This event provided a unique platform for our team to apply quantitative analysis in a fast-paced, simulated marketplace, where real-time decision-making was key to maximizing returns. The competition was not just about making profits; it was about honing our ability to develop and execute advanced trading strategies using data-driven insights.

Throughout the competition, we focused on statistical arbitrage, a strategy that involves identifying pairs of assets whose prices move together but temporarily diverge. By using various forms of statistical analysis, our team was able to identify the ideal pairs for trading. We employed comprehensive backtesting techniques to evaluate the effectiveness of our strategies across historical data, assessing performance metrics such as realized profit and loss, Sharpe ratio, and drawdowns. These analyses provided us with crucial insights into how the strategies performed under different market conditions.

In addition to statistical arbitrage, we implemented machine learning techniques to enhance our decision-making. A key innovation in this project was the development of a Jupyter notebook to identify optimal pair-candidates using k-means clustering, a method for grouping similar data points. To refine our entry and exit points, we utilized a Long Short-Term Memory (LSTM) model, a type of recurrent neural network that is particularly effective for forecasting time-series data. This allowed us to predict price movements and optimize our trading positions, giving us a competitive edge.

Our participation in the Prosperity-2 competition was a rewarding experience that expanded our understanding of quantitative finance, statistical analysis, and machine learning in trading. The strategies we developed and the lessons we learned during this competition will inform future projects as we continue to explore the intersection of finance and data science.

Coming Soon.

In the Spring semester of 2024, Bonsai ACG’s marketing and business team undertook a significant Twitter data analysis project in partnership with the University of Miami’s Business School. The project, funded with a $5,000 budget, aimed to analyze and visualize geospatial trends across over 700,000 tweets. This large-scale data analysis effort provided deep insights into social media behavior and trends, offering invaluable information for both academic research and business applications.

Our team began by managing the budgeting, data collection, and technical infrastructure needed to handle this vast dataset. We utilized Twitter’s professional API to gather the data, which included several millions of data points such as tweet content, tweet metadata, and user metadata. One of the main challenges we faced was that a significant portion of tweets lacked explicit location data, which was crucial for the geographic analysis we were tasked with conducting.

To overcome this, we implemented geolocation inference using the Pigeo repository. This tool allowed us to infer the location of tweets even when explicit geotags were missing. However, one of the key challenges we encountered was understanding the dependencies of this older tool and learning how to effectively run subprocesses to make it function within our modern data pipeline. This required extensive research and trial-and-error as our team delved into documentation and community forums to troubleshoot compatibility issues and ensure smooth integration. By analyzing tweet metadata, including language and timezone information, we were able to estimate geographic data, enabling a robust analysis of social media trends by region.

The data processing and analysis phase was extensive, involving the cleaning, sorting, and visualization of the tweet data using Python. Our team conducted an in-depth examination of the error rates associated with geolocation and data collection methods, ensuring the integrity of our findings. The final product was a series of visualizations and reports that illuminated social media trends across various geographic regions, offering both academic researchers and business analysts a powerful tool for understanding the dynamics of online communication. This project tested Bonsai Analysts’ ability to manage large-scale data collection and analysis efforts.

Coming Soon.

Coming soon.