model-averaging

A statistical technique that combines predictions from multiple candidate models by computing a weighted average, rather than selecting a single best model. Given m candidate models with predictions f_i(x) and non-negative weights w_i summing to 1, the averaged prediction at input x is Σ_{i=1}^{m} w_i · f_i(x). The weights are typically derived from a model selection criterion such as AIC: w_i = exp(−Δ_i / 2) / Σ_j exp(−Δ_j / 2), where Δ_i is the difference between model i's AIC and the minimum AIC across all candidates. Model averaging reduces selection uncertainty and often produces more accurate predictions than any individual model, especially when no single model is clearly superior. Computed by evaluating each candidate model, deriving weights from a criterion, and forming the weighted sum. Established in statistical practice through Burnham and Anderson (2002) and earlier Bayesian model averaging (Raftery et al., 1997), it persists through statistical textbooks, software packages (R's MuMIn, AICcmodavg), and research practice across ecology, econometrics, climate science, and machine learning. [formal: MA | substrate: mind | horizon: a moment | explicit: yes | epoch: 2.11]