arXiv:2405.12234v1 Announce Kind: new
Summary: Machine Studying algorithms are infamous for offering level predictions however not prediction intervals. There are various purposes the place one requires confidence in predictions and prediction intervals. Stringing collectively, these intervals give rise to joint prediction areas with the specified significance stage. It’s a simple job to compute Joint Prediction areas (JPR) when the information is IID. Nonetheless, the duty turns into overly tough when JPR is required for time collection due to the dependence between the observations. This mission goals to implement Wolf and Wunderli’s methodology for developing JPRs and examine it with different strategies (e.g. NP heuristic, Joint Marginals). The strategy beneath examine is predicated on bootstrapping and is utilized to totally different datasets (Min Temp, Sunspots), utilizing totally different predictors (e.g. ARIMA and LSTM). One problem of making use of the tactic beneath examine is to derive prediction customary errors for fashions, it can’t be obtained analytically. A novel methodology to estimate prediction customary error for various predictors can also be devised. Lastly, the tactic is utilized to an artificial dataset to seek out empirical averages and empirical widths and the outcomes from the Wolf and Wunderli paper are consolidated. The experimental outcomes present a narrowing of width with robust predictors like neural nets, widening of width with rising forecast horizon H and reducing significance stage alpha, controlling the width with parameter okay in Ok-FWE, and lack of info utilizing Joint Marginals.
Supply hyperlink