Beyond Forecasting: Recasting Volatility Control as a Routing Problem

The paper proposes VolRouter, a modular framework that reformulates volatility control as persistent, state-conditioned routing over a library of estimator–controller policies. The authors identify policy selection as an explicit layer between risk estimation and portfolio execution, shifting the objective from finding one universally best rule to selecting the appropriate control policy for the current state.

The key insight is that "the same increase in estimated volatility can call for very different responses across market states. At crash onset, rapid de-risking may be appropriate. During a noisy reversal, a smoother adjustment may be preferable. In a low-volatility regime, preserving exposure may matter more."

Routing Formulation: The framework separates three decisions: state inference, switch review, and pair selection. Formally:

• State inference: zt = Rtheta(Itstate)

• Switch review: gt = Gtheta(zt, Itroute, kt−1) ∈ hold, switch

• Pair selection: kt = Stheta(zt, Itroute, Ct kt−1) when switch is justified

The decision chain is: It → st → gt → kt → wt. Pair selection is invoked only when the switch-review layer decides that the current pair should not be retained.

Policy Library: The library spans realized- and EWMA-based estimators, GARCH/HAR-style models, machine-learning and market-specific risk estimators, together with scaling, regime-aware, trend, drawdown, tail-risk, and multi-asset portfolio controllers. The full single-asset master configuration contains 209 named estimator–controller pairs before eligibility filters.

Router Instantiations: The Router can be instantiated by rules, learnable selectors, contextual bandits, or LLMs. For the LLM Router, the language model is used only to infer the current market state, determine whether the active pair should be retained, and, when necessary, select a replacement from an explicitly supplied candidate set. The LLM does not predict returns or directly generate portfolio weights.

Timing Convention: Routing decisions use only completed historical transitions. For each candidate, the Router receives trailing diagnostics over Ht = t − H,..., t − 1, H = 63. No routing diagnostic at date t contains rt+1 or later realized portfolio outcomes.

Four out-of-sample settings were evaluated:

• S&P 500: 10% volatility target, 5 bp transaction cost, 252-day estimation window

• Multi-Asset: 10% target, 5 bp cost, 126-day window, cross-asset panel

• Bitcoin: 35% target, 8 bp cost, 90-day window, 365-day annualization

• USDT: 2% target, 2 bp cost, 90-day window, 365-day annualization

Baselines include RV + Naive Scaling, EWMA Targeting, Regime-Aware Fixed, Mixture-of-Experts, and Contextual Bandit.

S&P 500: "VolRouter achieves the highest Sharpe ratio in three of the four settings. On S&P 500, it increases Sharpe from 0.952 for RV + Naive Scaling to 1.222 while reducing maximum drawdown from 15.10% to 12.58% and daily CVaR from 1.76% to 1.32%. Importantly, annualized return is slightly lower (10.84% versus 11.11%), showing that the Sharpe improvement is not obtained by simply increasing return or exposure."

Multi-Asset: RV + Naive Scaling produces the highest raw annualized return (16.35%), but VolRouter achieves the highest Sharpe (1.540 versus 1.498) together with substantially lower CVaR (1.18% versus 1.56%).

Bitcoin: VolRouter obtains both the highest annualized return (30.81%) and highest Sharpe ratio (1.123), while reducing maximum drawdown from 51.69% under RV + Naive Scaling to 39.39%. Regime-Aware Fixed and Mixture-of-Experts achieve smaller drawdowns but their realized volatilities fall to only 6.73% and 7.51% against the 35% Bitcoin target, indicating substantial under-exposure.

USDT: USDT provides a boundary case in which simpler state-aware selectors remain competitive. Contextual Bandit (9.566) and Regime-Aware Fixed (9.140) achieve higher Sharpe ratios than VolRouter (8.379), though all state-aware methods strongly outperform simple volatility scaling.

The component ablation shows that the largest losses within the state-context group arise when the Router loses previous-action information or recent trend/return context. Removing relative pair comparison produces a substantially larger loss than removing a single reference signal. Replacing dynamic routing with a fixed pair causes one of the largest performance degradations.

The authors conclude: The gain does not come simply from exposing the system to a larger policy library. It comes from combining informative state context, relative candidate judgement, and persistent but adaptive pair selection.

• Target volatility shows the strongest variation: Sharpe reaches 1.55 at the 5% target and declines to approximately 0.9–1.0 at the more aggressive 15–25% targets.

• Transaction costs produce relatively compact variation, indicating results are not explained by a single favorable friction assumption.

• Switching sensitivity is more consequential and non-monotonic: both persistent and aggressive routing configurations can materially change performance.

• Candidate pool size: Performance broadly improves as the candidate pool becomes richer, although the relationship is not monotonic at every pool size.

"Backbone rankings vary across both markets and evaluation dimensions. A model that is stronger on S&P 500 or Bitcoin is not necessarily strongest on downside-risk metrics in USDT or Multi-Asset. There is therefore no uniformly dominant Router backbone across all control environments. The authors note that useful routing behavior appears across multiple backbones. The empirical effect therefore reflects the routing formulation rather than depending exclusively on one particular language model."

1. Risk allocation, not return prediction: The main benefit of VolRouter is not uniformly higher raw return, but better risk deployment across market conditions.

2. Control heterogeneity matters: Routing is most useful when different market states require genuinely different control policies.

3. Decision decomposition is essential: The benefit therefore comes from structured, persistent adaptation rather than simple winner selection.

4. More freedom doesn't guarantee improvement: Neither more frequent switching nor a larger policy library guarantees better control. The results favor selective adaptation that balances policy diversity with switching discipline.

The paper concludes: "These findings support a broader view of volatility management: beyond estimating risk, a system must decide which control policy should act, and when. Making this decision layer explicit provides a modular direction for adaptive and interpretable risk-control systems."

Improvements to AI Systems:

1. State-Conditioned Policy Routing Module: Add a routing layer to any AI control system that selects among multiple candidate policies based on inferred environmental state, rather than relying on a single fixed policy. The router uses three sub-modules: (a) a state encoder that maps observations to a compact latent state, (b) a switch-review gate that decides whether to retain or change the active policy, and (c) a pair selector that chooses a replacement only when switching is justified. This prevents unnecessary policy churn while enabling adaptive responses to regime changes.

2. Persistent Memory of Active Policy: The system maintains a memory of the currently active policy and conditions routing decisions on this history. This prevents oscillation between policies and ensures that switching occurs only when the state change is significant enough to warrant it. The improved system can avoid the performance degradation observed when the router loses previous-action information.

3. Relative Candidate Comparison with Contextual Diagnostics: Instead of scoring policies in isolation, the system evaluates each candidate policy relative to the currently active one, using trailing diagnostics (e.g., 63-step historical windows) that include recent trend, return, and volatility context. This relative judgment improves selection accuracy over absolute scoring, as demonstrated by the ablation showing larger losses when relative comparison is removed.

4. Decomposed Decision Pipeline: The system separates state inference, switch review, and policy selection into distinct, interpretable stages. This modularity allows each stage to be optimized independently and makes the system's reasoning auditable. The improved system can explain why it switched policies (state change) and which policy it chose and why (relative candidate comparison).

5. Risk-Aware Objective Integration: The system optimizes for risk-adjusted returns (e.g., Sharpe ratio) and downside-risk metrics (CVaR, max drawdown) rather than raw return. It can achieve higher Sharpe ratios without increasing exposure, as shown by the S&P 500 result where return slightly decreased but risk-adjusted performance improved substantially. The system learns to deploy capital more efficiently across market conditions.

6. Backbone-Agnostic Router: The routing formulation works across different underlying models (rules, learnable selectors, contextual bandits, LLMs). For LLM-based routers, the language model is constrained to only infer state and select policies from an explicit candidate set—it never predicts returns or generates weights directly. This makes the system robust to backbone choice and prevents hallucinated risk decisions.

7. Switching Discipline with Adaptive Frequency: The system includes a tunable switching penalty that balances policy diversity against persistence. It can be configured to be more or less aggressive based on the environment, with the insight that neither extreme (constant switching or fixed policy) is optimal. The improved system can self-calibrate switching frequency based on observed state volatility.

8. Multi-Asset and Cross-Market Generalization: The system can operate on cross-asset panels (e.g., multi-asset portfolios) and adapt to different volatility targets (2% to 35%) and transaction cost regimes. It learns to route differently across markets—e.g., more aggressive de-risking in Bitcoin, smoother adjustments in USDT—without manual reconfiguration.

What the Improved AI System Can Do:

• Dynamically switch between risk-control policies (e.g., scaling, regime-aware, tail-risk) in response to market state changes, without overreacting to noise.

• Achieve higher Sharpe ratios and lower drawdowns/CVaR than any single fixed policy, even when the fixed policy has higher raw returns.

• Operate transparently: each routing decision is explainable as a state inference, a switch/no-switch decision, and a candidate selection with relative justification.

• Adapt to new markets or assets by simply extending the policy library and re-training the router, without redesigning the core logic.

• Handle boundary cases (e.g., stablecoins with low volatility) where simpler state-aware methods may suffice, while still outperforming naive scaling.

• Avoid the pitfalls of over-frequent switching or over-large policy libraries by using a structured, persistent routing mechanism that balances exploration and exploitation.