This paper proposes a tripartite framework of design, evaluation, and postevaluation analysis for generating and interpreting economic forecasts. This framework's value is illustrated by re-examining mean square forecast errors from dynamic models and nonlinearity biases from empirical forecasts of US external trade. Previous studies have examined properties such as nonlinearity bias and the possible nonmonotonicity and nonexistence of mean square forecast errors in isolation from other aspects of the forecasting process, resulting in inefficient forecasting techniques and seemingly puzzling phenomena. The framework developed reveals how each such property follows from systematically integrating all aspects of the forecasting process.A framework for economic forecasting C229 forecast errors for certain estimation sample sizes and forecast horizons; the uncertainty arising from Monte Carlo simulation of mean square forecast errors when exact analytical results are unavailable; and a parallel uncertainty in estimating the bias from using deterministic forecasts from nonlinear models. Each case is of substantive economic and methodological interest, and each shows how the proposed framework clarifies and improves upon existing results and techniques.This paper is organized as follows. Section 2 develops the general framework, which partitions forecast activities into design, evaluation, and post-evaluation analysis. Design includes specification of the model for forecasting and selection of the forecast's characteristics of interest. Thus, design includes choice of the variables being forecast, the forecast horizon, the model's specification, and estimation method. Evaluation specifies how the forecasts are actually generated, and includes the choice of analytical or numerical techniques and the type of approximation used. Post-evaluation analysis includes presentation and summarization of the forecasts.Sections 3 and 4 illustrate the principles of design, evaluation, and post-evaluation analysis for some time-series models that have been studied previously. That choice emphasizes how existing results can be beneficially re-interpreted in the proposed framework. Section 3 presents analytical properties of the mean square forecast error (MSFE) for one-step and multi-step ahead forecasts from vector autoregressions, relying on approximations due to Schmidt (1974) and Baillie (1979b). While the general formula for the MSFE is useful for empirical applications within this class of models, many of its properties can be most easily understood for a special case, the univariate first-order autoregressive [AR(1)] process. As detailed below, Section 4 discusses three potential properties of the MSFE for the AR(1) model: nonmonotonicity, nonexistence, and uncertainty in its estimation by Monte Carlo simulation.Firstly, Section 4.1 re-interprets Hoque et al. 's (1988) exact numerical results on the MSFE for an AR(1) model without an intercept. Mean square forecast errors are commonly viewed as being monotonic in the forecast...