Title: Dynamic Objective Selection with Safeguards and LLM Oversight for Financial Decision-Making

Abstract:

In financial decision-making domains like portfolio allocation and stock recommendation, practitioners typically forecast future risk and returns to determine investor trades. The specific optimization objective chosen in this process is a critical determinant of actual performance. However, market environments are dynamic; a static objective often proves suboptimal as conditions shift. While pipelines that attempt to adapt via latent regime estimates can address this, they are frequently plagued by noise, delays, and the operational instability caused by high turnover from frequent switching.

To address these challenges, we introduce DOSS (Dynamic Objective Selection with Safeguards), a learning-based selector that identifies the most appropriate decision-relevant objective function at every time step. By analyzing interpretable statistical summaries of recent returns, DOSS selects from a limited pool of candidates—such as loss-averse, return-seeking, or risk-adjusted strategies—without relying on intermediate regime variables. This approach treats objective selection as a classification task, utilizing sequential updates within a rolling window to generate forward-looking predictions while preventing temporal leakage. Additionally, the model provides a confidence score for each recommendation.

To ensure stability during deployment, DOSS incorporates confidence-aware gating and a fail-safe mechanism. This system overrides low-confidence proposals with a conservative default and applies explicit controls to limit switching frequency, thereby reducing misselection and operational volatility. Furthermore, we embed governance into the framework by employing a Large Language Model (LLM) as an oversight agent rather than a generator of new objectives. The LLM is constrained to either approve a proposed objective or revert to a predefined safe default, with deterministic, rule-based constraints triggering overrides when necessary.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC