Abstract
This paper develops methods to analyze and quantify the dynamic effects of path conditioning in conditional forecasts from reduced-form (VAR)vector autoregressive models. Building on Kalman filtering and the observation-weight extraction method of Koopman and Harvey (2003), we characterize how imposed future paths of observables influence forecast outcomes. We establish the analytical relationship between observation weights and unscaled generalized impulse response functions (GIRFs): under single-period conditioning, the weights coincide with unscaled GIRFs, while multi-period path conditioning generates a richer, non-local structure of influence. We use these weights to construct measures of overall and marginal variable importance that disentangle model-implied sensitivity from realized sample dynamics. An empirical illustration shows how the proposed framework enhances the interpretability of scenario forecasts by attributing forecast revisions to specific variables and horizons. The approach is model-consistent, structurally agnostic, and well suited for policy-oriented scenario
analysis.
Keywords: Vector Autoregression; Conditional Forecast; Filter Weights; Impulse Response
Analysis.