Abelee is a data science company that, among other things, uses statistics and machine learning to make short-term predictions in the financial markets. Abelee has several prediction algorithms running in parallel. These algorithms are constructed to forecast on different time horizons with varying degrees of accuracy. Since financial markets are typically dynamic and highly non-stationary, it is difficult to combine predictions on a joint set of time horizons, and it is even harder to construct valid prediction sets for such aggregated predictions.
Conformal prediction refers to a set of algorithms used to construct valid prediction sets in statistics and machine learning. The prediction sets are valid in the sense that they guarantee correct (finite sample) coverage probability under very mild assumptions. This means that the method is very applicable for complex and non-standard methods (e.g. so-called "black box" models). Conformal prediction is constructed with on-line prediction problems in mind, which makes it suitable for the problems alluded to above, however, the general theory breaks down for most time series types of problems with autocorrelated or non-stationary data.
This project is about exploring and extending ideas in conformal prediction for time series data with focus on challenges within financial time series. Abelee will provide relevant and real data for the project.