Speaker 1: Carlos Barajas

Title

A High Performance Computing Framework for Hyperparameter Searching Storm Data with GPUs

Abstract

Predicting violent storms and dangerous weather conditions in a timely fashion can be difficult with current models due to the immense complexity associated with weather simulation. But for public warnings, for instance, of tornados, time is critical. Machine learning has the potential to classify weather patterns more quickly than other simulation methods, and thus to provide quicker alerts to the public, leaving more time to respond to the alert.

We use various parallel frameworks for hyperparameter tuning in a high performance computing environment. This allows us to examine the wall time performance of various hyperparameter configurations produced with great detail. We compare CPU and GPU based learning while also examining the impact of varying the number of GPUs used for training.

Speaker 2: Qing Ji

Title

Creating Stock Portfolios Using Hidden Markov Models

Abstract

Hidden Markov models (HMM) have been widely used to analyze stock market data in the statistical literature. Due to hidden market trends, the structure of HMM fits well with stock data. By utilizing historical stock closing values over a fixed training period, we evaluate stock performances in terms of capital gain using HMM. Stocks are selected into a yearly portfolio based on the model. We used out-of-sample testing to investigate our portfolio selection method and showed annual capital gains from 2010 to 2018. The performances of proposed portfolios were compared to the S&P 500 index.