Robert Roloff

Based on an obfuscated data set provided by the client, a quantitative asset manager, I researched whether new machine learning methods could improve the existing trading systems. Since the outcome of the initial research project had been very promising, the client decided to use the new approach in production. In the following steps, I prepared a roadmap for the implementation, and we gathered a team of other researchers and developers. Together with a freelance developer, I created a new machine learning back end that allowed us to leverage cloud computing resources from AWS. The solution is now successfully used by the client for the trading strategy of his main fund.

Training machine learning models for financial time series data often requires specific cross-validation schemes that standard libraries, such as scikit-learn, do not support. I developed a hyperparameter optimization infrastructure including Bayesian, random, and grid search that takes the client's cross-validation schemes into account and uses cloud computing resources from AWS. This enabled the client to search a much larger parameter space and train better machine learning models for his trading strategies.

Recently, commodity prices have become increasingly volatile, and sometimes they show very strong trends. As a consequence, producers and consumers of these commodities seek ways to optimize their hedging against adverse price movements. The client I was working for provides software that generates price and volatility predictions for commodities via modern machine learning tools. I researched and developed several models for a range of industrial and precious metals, allowing the client to extend their product universe.