When the Horizon Changes the Verdict: A Paired Study of 14,000 Multi-Agent Investment Reports

Abstract

Investment ratings are often presented as if they were timeless properties of assets. In practice, they are conditional on a forecast horizon. This study measures how often the same production research system reaches a different verdict when the horizon changes from one year to five years. The data contain 14,400 individual configured-agent investment reports contributing to 1,247 matched one-year/five-year consensus pairs across 377 assets and four production batches. Each pair holds the asset, batch, contributing task set, underlying forecast path, and scoring procedure constant; the paired contrast changes only the point on the forecast path used for the horizon summary. The exact five-level verdict changes in 56.5% of pairs, while the broader negative/neutral/positive direction changes in 50.0%. Among exact changes, 93.0% move upward at five years. The average implied annual-return estimate rises by 6.55 percentage points, configured-agent dispersion falls, and directional consistency increases. A prespecified trace of the largest ordinal flip shows the mechanism in concrete form. Results are descriptive of recorded production outputs and do not test realized investment performance. Horizon is a first-class input to an investment thesis: a consensus label without its horizon can conceal economically meaningful disagreement. The record includes the working paper, accessible manuscript source, aggregate results, auditable extraction SQL, production-lineage notes, a reference audit, publication figures, and sanitized case-study summaries. Underlying report text, licensed inputs, proprietary prompts, task identifiers, and row-level production data are not publicly released. This material is research about an investment-research system and is not investment advice.