Ranking AI Investment Forecasts: Original Scoring, Dynamic Rescoring, and the Trade-Off Between Return-Focused and Risk-Adjusted Strategies

AI investment platforms face a two-clock problem: expensive multi-agent research changes slowly, while market prices move daily. We separate original forecast scoring from later rescoring and distinguish return-focused from volatility-adjusted investment objectives. An original-scoring study evaluates 48 fixed formulas across six stock cohorts, yielding 3,456 portfolios. A separate rescoring study tests 115 policies at 7-, 14- and 30-day checkpoints, yielding 8,280 portfolios. Each starts with USD100,000 and tracks actual distributions, corporate actions and foreign exchange. G10, a return-focused formula discounting ensemble disagreement, ties for the strongest Top 10 equal-weight return midrank. B02, a simple forecast excess-return-to-historical-volatility ratio, provides the strongest return-to-drawdown profile among 27 volatility-aware rules. For one-year BUY rescoring, a 75% original/25% updated G10 blend outperforms raw updated G10 in 15 of 18 primary cells, raising mean return from 12.80% to 19.54%. Stock-level attribution identifies retention of early risers as the leading mechanism. The advantage weakens under concentration controls across all six cohorts; raw updated G10 is retained for five-year rescoring on matched-cell consistency. A separate 203-stock research-reuse bridge finds a 1.23-percentage-point improvement from blending in unchanged-model short Top 10 windows, while broader portfolios and longer holding windows expose its limits. The findings support distinct ranking objectives, continuity-aware short-window reuse and fresh research when underlying theses need rebuilding. The experiments are retrospective, with overlapping and unequal holding windows. This record contains the full 28-page manuscript, LaTeX source, and research data and code for the two frozen scoring experiments and the separate 203-stock research-reuse bridge. The supporting archive includes all 3,456 original-scoring and 8,280 rescoring portfolio outcomes, the 48 and 115 fixed policy definitions, 2,135,172 derived daily portfolio-wealth observations, synthetic holding allocations, and the added mechanism, concentration, dividend and budget diagnostics. Python standard-library verification scripts reproduce the published calculations from those frozen research outputs. Author: Russlan Ramdowar, Future Edge Group FZE / iPulse AI; ORCID 0009-0004-9311-6957; [email protected]. Original manuscript and derived research outputs: CC BY 4.0. Focused research code: MIT. Licensed raw market observations are excluded. The approved commercial-interest and AI-assistance disclosures are retained in the manuscript. GitHub: https://github.com/TheFutureEdge/research-ranking-ai-investment-forecastsCompanion research-budget paper: https://arxiv.org/abs/2611.04348Companion bridge artifacts: https://doi.org/10.5281/zenodo.23941195 Version 1.0.0, 8 November 2026. Submitted to arXiv under cs.MA; the public arXiv identifier will be added after announcement.