Paper: arXiv 2610.25822

Authors: Levent Bulut

Abstract

Older observations can help a forecaster learn a predictive relationship while becoming misleading when that relationship changes. We study this tension using paired return histories that differ in older observations but share recent returns and the conditional distribution of the next return. Across five foundation-model forecasting procedures and three simulated processes, each evaluated on 1,024 independent replications, conflicting history reduces the probability of improving on zero in all 30 higher-signal comparisons. This loss of usefulness occurs despite successful signal recovery under stable conditions. A separate exploratory comparison on 56 assets finds that shortening inputs from 256 to 64 returns increases mean absolute error for every model by 1.77-4.58 basis points; neither length improves observed overall error over zero. The evidence distinguishes the ability to extract a signal from the ability to withstand conflicting history, while showing why context sensitivity alone does not justify discarding older data.

Complexity vs Empirical Score

  • Math Complexity: 6.0/10
  • Empirical Rigor: 8.0/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper presents a novel experimental design to isolate the impact of conflicting historical data on foundation model forecasts, using rigorous simulations and statistical evaluation. While the mathematical derivations are present, they are not overly complex, focusing more on the experimental setup and statistical analysis. The empirical rigor is high due to the extensive simulations and careful statistical methodology.

Research Flowchart

  flowchart TD
    A[Research Goal: Impact of Conflicting History on Foundation Model Return Forecasts] --> B{Methodology: Compare Forecasting Procedures under Conflicting History};
    B --> C{Data/Inputs: Paired Return Histories (differing older, shared recent returns) + Conditional Next Return Dist.};
    C --> D[Computational Processes: 5 Foundation Model Forecasting Procedures, 3 Simulated Processes, 1024 Replications Each];
    D --> E[Key Finding 1: Conflicting History Reduces Probability of Improving on Zero in 30/30 Higher-Signal Comparisons];
    E --> F[Key Finding 2: Signal Recovery Successful under Stable Conditions, but Loss of Usefulness with Conflicting History];
    F --> G[Key Finding 3: Shortening Input (256 to 64 returns) Increases MAE by 1.77-4.58 bps across models; Neither Length Improves Overall Error vs. Zero];
    G --> H[Outcome: Distinguishes Signal Extraction from Withstanding Conflicting History; Context Sensitivity Alone Doesn't Justify Discarding Older Data];