What is A Conformal Prediction Oracle?
Math
Conformal inference, or conformal prediction, is a way to overcome issues all together and produce a prediction interval that does not under or overcover. It relies heavily on the ranks of continuous random variables and their stochastic properties.
Introducing the conformal prediction oracle, in concept.
Ingredient (1):
Conformal prediction for time-series ensembles. Per a paper by Chen Xu and Yao Xie, one can use a recent window of leave-one-out-of-the-ensemble prediction residuals to empirically estimate a confidence interval. See [2010.09107] Conformal prediction for time series (arxiv.org) for this twist on the conformal prediction trick and theoretical results.
Ingredient (2):
A microprediction oracle. Skip the step where you personally create models, and simply publish the data instead (instructions: Publishing a stream | microprediction ). Furthermore, the ensemble will come to you, and it will probably be better than yours. It will also get better over time.
Ingredient (3):
Use the microprediction oracle to improve the implicit distributional estimate used in conformal prediction.
Kicker: A “z1-stream” for the ensemble residuals is automatic (see Collective Distributional Prediction (microprediction.com) for an explanation) and serves as ongoing assessment.
Moral. Just about any approach can be used in new ways in a world populated with microprediction oracles. Prediction is everywhere, including buried inside existing algorithms. It can be used recursively at a small marginal cost.
The book is here: Microprediction: Building an Open AI Network: Cotton, Peter: 9780262047326: Amazon.com: Books
Lastly, I hope it gets your wheels spinning about the possibilities for collective intelligence “in the small”.
What is A Conformal Prediction Oracle?
