Taxonomy of Prediction
Abstract
A prediction makes a claim about a system's future given knowledge of its past. A retrodiction makes a claim about its past given knowledge of its future. We introduce the ambidextrous hidden Markov chain that does both optimally -- the bidirectional machine whose state structure makes explicit all statistical correlations in a stochastic process. We introduce an informational taxonomy to profile these correlations via a suite of multivariate information measures. While prior results laid out the different kinds of information contained in isolated measurements, in addition to being limited to single measurements the associated informations were challenging to calculate explicitly. Overcoming these via bidirectional machine states, we expand that analysis to information embedded across sequential measurements. The result highlights fourteen new interpretable and calculable information measures that fully characterize a process' informational structure. Additionally, we introduce a labeling and indexing scheme that systematizes information-theoretic analyses of highly complex multivariate systems. Operationalizing this, we provide algorithms to directly calculate all of these quantities in closed form for finitely-modeled processes.