predictionrevisited 1.0.0

Prediction Revisited module for observation-based prediction

Prediction Revisited

Version 1.0.0

This python module implements the core statistical concepts from the book "Prediction Revisited: The Importance of Observation" by Megan Czasonis, Mark Kritzman and David Turkington. In short, these methods form predictions by interpreting data as 'experiences' rather than variables, which builds from a rigorous base of information theory and facilitates intuitive interpretation of predictions as averages of prior outcomes. This procedure works on the principle of relevance, which is determined using the Mahalanobis distance.

The baseline predictions from a relevance-weighted average match those of ordinary least squares linear regression. However, this view allows us to gain additional insight from the importance of each observation, and it allows us to extend linear regression to Partial Sample Regression, which relies on a subset of the most relevant data and is therefore locally conditional to the circumstances of every prediction.

Relevance-weighted prediction also allows us to evaluate the fit of every individual prediction, which is distinct from the R-squared of a model. In fact, for a full-sample regression the R-squared is equal to the information- weighted average fit of every prediction trial. We also define a measure of precision and asymmetry, as well as co-occurrence which is a single- observation component of correlation.

"Prediction Revisited" also uses pairwise comparisons between observations to elucidate new intuition. While theoretically elegant, pairwise comparisons are often computationally inefficient, so this code module offers the flexibility to perform most computations more efficiently with shortcut=True.

Most of the module is built on top of numpy for matrix manipulation, but we also include ancillary features and a demo that ingests Pandas DataFrames. Note that for clarity, we keep results as 2-dimensional arrays in most cases.

For more information, see: www.predictionrevisited.com

Dependencies

numpy (for all matrix calculations) pandas (for ancillary features involving data formatted in DataFrames)

License

MIT (see LICENSE.md)

"""