Context
The current implementation reconstructs the seasonal state [s_2, ..., s_{S-1}] from s1_obs_pre history for post-period propagation (sampler.rs L574-586). This is an approximation because process noise η accumulates in the smoother draws.
Problem
When sigma2_seasonal is large, the reconstruction from past s1_obs values diverges from the actual smoother state. This could affect CI parity in high-noise seasonal scenarios.
Proposed fix
Modify local_level_seasonal_smoother() to return the full state vector at t = pre_end - 1 (S-dimensional) alongside levels and s1_obs. Use this exact state as the starting point for post-period propagation.
Impact
Eliminates approximation error in post-period seasonal predictions. May improve R parity for edge cases.
Estimated effort
Medium
Context
The current implementation reconstructs the seasonal state
[s_2, ..., s_{S-1}]froms1_obs_prehistory for post-period propagation (sampler.rs L574-586). This is an approximation because process noise η accumulates in the smoother draws.Problem
When
sigma2_seasonalis large, the reconstruction from pasts1_obsvalues diverges from the actual smoother state. This could affect CI parity in high-noise seasonal scenarios.Proposed fix
Modify
local_level_seasonal_smoother()to return the full state vector att = pre_end - 1(S-dimensional) alongsidelevelsands1_obs. Use this exact state as the starting point for post-period propagation.Impact
Eliminates approximation error in post-period seasonal predictions. May improve R parity for edge cases.
Estimated effort
Medium