Skip to main content
Back

DOSS Lets an LLM Override AI Trading Decisions When Markets Turn Uncertain

DOSS uses a learned classifier and LLM oversight to dynamically pick the right portfolio objective as market conditions shift, beating static strategies without the noise of regime detection.

Paper, Full Access
  • DOSS is a learned classifier that dynamically picks the best portfolio optimization objective (return-seeking, loss-averse, or risk-adjusted) at each time step.
  • It skips noisy regime detection entirely, going straight from statistical market summaries to objective selection.
  • A confidence-aware gating mechanism overrides low-confidence picks to a safe conservative default, reducing costly misselections.
  • An LLM acts as an oversight layer, restricted to accept or override proposals -- it cannot generate new objectives, keeping behavior auditable.
  • DOSS-GateDown outperforms static single-objective strategies on HPA (Hit Probability Average) and approaches the Oracle upper bound in backtests.
  • No code released; accepted to the ICLR 2026 Financial AI Workshop as a research prototype.

Every portfolio optimization system has to answer a deceptively simple question: what are you actually optimizing for? Maximize returns? Minimize downside risk? Balance both? The answer that works in a bull market can destroy capital in a downturn, and vice versa. Most systems pick one objective and stick with it. A new paper accepted at the ICLR 2026 Workshop on Advances in Financial AI proposes a smarter approach: let a model learn when to switch objectives, and put an LLM in charge of final approval.

The system is called DOSS (Dynamic Objective Selection with Safeguards). Rather than committing to a single optimization target, DOSS treats objective selection itself as a classification problem, choosing at each time step from a small menu of candidates -- return-seeking, loss-averse, and risk-adjusted -- based on statistical summaries of recent market returns.

The problem with picking one objective and sticking with it

Financial markets cycle through regimes: trending bull runs, volatile crashes, sideways chop. Traditional rule-based or static optimization approaches often fail to capture the nonlinear interactions among investor behavior, market volatility, and evolving financial objectives. The standard fix is a regime-switching pipeline -- a model that first detects which market regime you are in (bull, bear, stable), then picks the appropriate strategy. Regime-switching poses both problems and opportunities for portfolio managers: if a switch in the behaviour of the markets is not quickly detected, it can be a source of loss since previous trading positions may be inappropriate in the new regime.

The deeper issue is that regime detection is an intermediate, unobservable variable. You never actually know which regime you are in -- you infer it from noisy signals, and those inferences arrive late. DOSS sidesteps this entirely by skipping regime estimation and going straight to the decision: which objective should I use right now?