Forecast Combination with Temporal Validation and Context Memory

A forecast combination must decide how much historical expert performance remains relevant at a new origin. We study a fixed context-weighted rule and two variants that select their parameters using completed historical target blocks. ContextMemoryCV selects context shrinkage and temperature; ValidationMix also considers regularized convex stacking. Each inner fit excludes outcomes not yet observed at its validation origin. The original fifteen-case experiment remains negative for fixed ContextMemory, with mean MASE 1.4504 versus 1.4489 for its recency-only ablation. In a separate external evaluation, 294 of 359 M4 Weekly series satisfy the fixed history criterion, and 1,422 of 1,428 M3 Monthly series remain after exact deduplication. ValidationMix has mean series MASE 0.2855 versus 0.3109 for RecencyMean on M4 Weekly, and 0.8993 versus 0.9085 on M3 Monthly. AutoTheta gives 0.3076 and 0.8472, respectively: improvement over the recency control does not establish superiority on both collections. Interval scores also retain unfavorable outcomes. Full forecast paths, scales, weights, validation decisions and source hashes support independent recomputation of 53,148 external forecasts. These results concern the configured algorithms and admitted series; they neither establish priority for context weighting nor prove population-wide superiority or calibrated uncertainty. Source code, admitted inputs and computational records: https://github.com/fsantibanezleal/CAOS_RES_ChronoScope . Research atlas: https://chronoscope.fasl-work.com . Not peer reviewed.