Abstract:Information criteria provide a cogent approach for identifying models that provide an optimal balance between the competing objectives of goodness-of-fit and parsimony. Models that better conform to a dataset are often more complex, yet such models are plagued by greater variability in estimation and prediction. Conversely, overly simplistic models reduce variability at the cost of increases in bias. Asymptotically efficient criteria are those that, for large samples, select the fitted candidate model whose predictors minimize the mean squared prediction error, optimizing between prediction bias and variability. In the context of prediction, asymptotically efficient criteria are thus a preferred tool for model selection, with the Akaike information criterion (AIC) being among the most widely used. However, asymptotic efficiency relies upon the assumption of a panel of validation data generated independently from, but identically to, the set of training data. We argue that assuming identically distributed training and validation data is misaligned with the premise of prediction and often violated in practice. This is most apparent in a regression context, where assuming training/validation data homogeneity requires identical panels of regressors. We therefore develop a new class of predictive information criteria (PIC) that do not assume training/validation data homogeneity and are shown to generalize AIC to the more practically relevant setting of training/validation data heterogeneity. The analytic properties and predictive performance of these new criteria are explored within the traditional regression framework. We consider both simulated and real-data settings. Software for implementing these methods is provided in the R package, picR , available through CRAN.

Information criteria for multistep ahead predictions

Identification of k-step-ahead prediction error model and MPC control

Nonparametric Multistep-Ahead Prediction in Time Series Analysis

Posterior Averaging Information Criterion

Asymptotically minimax Bayesian predictive densities for multinomial models

An analytic example of latent information prior

On the overestimation of widely applicable Bayesian information criterion

Gibbs-Based Information Criteria and the Over-Parameterized Regime

A new class of information criteria for improved prediction in the presence of training/validation data heterogeneity

Kullback-Leibler Divergence and Akaike Information Criterion in General Hidden Markov Models

Nearest-Neighbours Estimators for Conditional Mutual Information

Equivariant online predictions of non-stationary time series

Prediction Interval for Autoregressive Time Series Via Oracally Efficient Estimation of Multi‐Step‐Ahead Innovation Distribution Function

The Focused Information Criterion for Stochastic Model Selection Problems Using $M$-Estimators

Human tracking with variable prediction steps based on Kullback-Leibler divergence

A Bayesian Information Criterion for Portfolio Selection

Prediction from compression for models with infinite memory, with applications to hidden Markov and renewal processes

Inferential Moments of Uncertain Multivariable Systems

A new uncertainty processing method for trajectory prediction

A general framework for multi-step ahead adaptive conformal heteroscedastic time series forecasting

Caveats for information bottleneck in deterministic scenarios