Speaker
Description
Traditional methods of fitting data involves hypothesis testing, with "discovery" declared if the null hypothesis is rejected. This approach often assumes that the data is generated by the model, under-estimates systematic errors, and leads to overfitting. Common methods for overcoming the last issue, such as LASSO, AIC, and BIC do not perform well. In addition, the entire methodology relies on the dubious prospect of finding the global minimum of a complex multidimensional objective function. I propose to address these issues by reframing discovery as "predictiveness" -- namely does a postulated effect (eg, a new particle) assist in predicting new measurements. The method obviates many of the traditional problems and leads to more robust results. The implementation of the scheme and applications to simple problems will be presented.