RegimeFolio: Regime-Aware Decision Systems

• RegimeFolio is a regime-aware design pattern that conditions decisions on observable regime variables like budget ratios and market volatility.
• In Bayesian optimization, it uses metrics such as the Portable Regime Score (PRS) to switch between strategies like Greedy and UCB for performance gains.
• In finance, RegimeFolio frameworks adjust portfolio forecasting and risk controls via methods ranging from explicit thresholds to latent-state inference.

Searching arXiv for the cited RegimeFolio-related papers to ground the article in current records. arxiv_search.query({"search_query":"all:RegimeFolio OR ti:\"RegimeFolio\" OR ti:\"Regime-Conditioned Evaluation in Multi-Context Bayesian Optimization\" OR ti:\"FR-LUX\"","start":0,"max_results":10,"sortBy":"submittedDate","sortOrder":"descending"}) I’ll also search for related regime-aware portfolio papers that explicitly frame a “RegimeFolio” blueprint. arxiv_search.query({"search_query":"(all:\"regime-aware portfolio\" OR all:\"RegimeFolio\") AND (cat:q-fin.PM OR cat:cs.LG OR cat:stat.ML)","start":0,"max_results":15,"sortBy":"submittedDate","sortOrder":"descending"}) RegimeFolio denotes a class of regime-aware decision systems that condition selection, allocation, or adaptation on the current operating regime rather than optimizing a single regime-agnostic policy. In recent arXiv usage, the term appears in two distinct but structurally related settings: acquisition planning in multi-context Bayesian optimization, where budget ratio and prior quality determine whether exploitative or exploratory acquisition should be used, and portfolio management in nonstationary financial markets, where volatility, liquidity, latent HMM states, correlation structure, or change points determine forecasting models, risk controls, and allocation rules (Thomas, 6 May 2026, Zhang et al., 14 Sep 2025, Zhang, 3 Oct 2025, Boukardagha, 21 Feb 2026).

1. Conceptual scope

In the Bayesian optimization literature, RegimeFolio is formulated as a regime-aware portfolio manager that selects among acquisition strategies such as Greedy, UCB, Thompson sampling, and REIGN on a per-context basis. The central claim is that acquisition quality is not invariant across evaluation conditions: it depends on observable regime variables, especially the budget ratio $B/|A|$ and prior quality $\rho$, so unconditional leaderboards obscure the conditional effect that matters in deployment (Thomas, 6 May 2026).

In finance, the term is used more broadly for portfolio systems that detect, forecast, or adapt to market regimes before choosing allocations. The architectures differ substantially. Some are modular predict-then-optimize systems that combine explicit regime classification with regime-specific forecasting and mean-variance allocation; others are friction-aware RL systems, conformal risk-control overlays, continual-learning policy libraries, or explainable HMM-based allocators (Zhang et al., 14 Sep 2025, Zhang, 3 Oct 2025, Schmitt, 3 Feb 2026, Pan et al., 29 May 2026, Boukardagha, 21 Feb 2026).

This suggests that “RegimeFolio” functions less as a single standardized algorithm than as a design pattern: make regime structure explicit, use it to parameterize decision rules, and preserve conditional performance information that would otherwise be averaged away.

2. RegimeFolio in multi-context Bayesian optimization

The Bayesian-optimization version studies streaming multi-context Bayesian optimization with a finite discrete action set $A$, per-context budget $B$, and $K$ sequential contexts evaluated in replay against cached oracles. Four prior families are considered—no transfer, EMA transfer, structured, and oracle—and four planners are evaluated: Greedy, UCB, Thompson sampling, and REIGN. Performance is reported through Terminal Hit@k and Cumulative Discovery AUC, with the core comparison centered on Greedy versus UCB (Thomas, 6 May 2026).

Its main organizing variable is the Portable Regime Score,

$\mathrm{PRS}(B,|A|,\rho)=\frac{B}{|A|}(1-\rho),$

where $\rho$ is the observable rank correlation between prior means and true action values at context start, or between prior means and observed outcomes over queried actions. Higher PRS corresponds to larger budget ratio and/or weaker prior, so exploration is more likely to pay. In deployment, $\hat\rho$ is estimated from $K_0\ge 3$ pilot contexts using Spearman rank correlation, and the practical decision rule is to use Greedy when $\widehat{\mathrm{PRS}}<\theta$ with $\rho$0, otherwise UCB or the online regime-adaptive planner. Online adaptation replaces $\rho$1 with $\rho$2 and uses one-way switching from UCB to Greedy as the prior sharpens.

The theoretical core is the No-Free-Leaderboard proposition. Let $\rho$3 denote the conditional average treatment effect of acquisition $\rho$4 over $\rho$5 under regime $\rho$6, and let $\rho$7 average this over a benchmark mixture. Whenever CATE changes sign across regimes, the reported ATE becomes a function of the benchmark mixture rather than a stable property of the algorithms. The paper formalizes this by showing that ties, wins, and losses for either acquisition are all achievable through benchmark reweighting.

Empirically, the paper audits 40 transfer-BO papers from NeurIPS, ICML, ICLR, AISTATS, UAI, TMLR, JMLR, and AutoML-Conf (2022–2025) and reports that 98% do not vary $\rho$8 as a controlled axis, 80% report only a single metric family, and 76% fix $\rho$9 among multi-context papers. On GDSC2, changing only the budget reverses the ranking: at $A$0, Greedy exceeds UCB by $A$1 Hit@1, while at $A$2, UCB exceeds Greedy by $A$3. Across 79 conditions spanning Buchwald–Hartwig chemistry, GDSC2 drug-response biology, and HPO-B, a hierarchical model gives $A$4 with 95% CI $A$5, $A$6, and ICC $A$7; 19% of conditions lie in an equivalence zone where $A$8 Hit@1. The constructive algorithmic instantiation, RegimePlanner, wins all 16 HPO-B search spaces at $A$9, exceeds the matched $B$0 per-context oracle on GDSC2 by 18% on Hit@1, and attains pre-registered prediction accuracy of $B$1, rising to $B$2 within EMA prior families (Thomas, 6 May 2026).

3. Regime representation in portfolio management

Financial RegimeFolio systems differ primarily in how they define the regime variable. One line uses explicit observable thresholds. The sector-specialized "RegimeFolio: A Regime Aware ML System for Sectoral Portfolio Optimization in Dynamic Markets" partitions market states into Low, Medium, and High volatility using daily VIX and rolling 252-day terciles $B$3 and $B$4, then trains regime- and sector-specific Random Forest or Gradient Boosting forecasters and couples them to a dynamic mean-variance optimizer with Ledoit–Wolf shrinkage (Zhang et al., 14 Sep 2025).

A second line uses multivariate volatility–liquidity states. FR-LUX defines $B$5 using rolling quantile thresholds on volatility $B$6 and illiquidity $B$7, and conditions both policy $B$8 and value $B$9 on that discrete state. The regime label is therefore an explicit control variable inside the MDP rather than a purely descriptive annotation (Zhang, 3 Oct 2025).

A third line uses latent-state inference. "Explainable Regime Aware Investing" fits a strictly causal Gaussian HMM, selects model order by out-of-sample predictive log-likelihood, and preserves regime identity by matching estimated Gaussian components to persistent templates via the closed-form squared 2-Wasserstein distance between Gaussians. The resulting regime probabilities are used directly in transaction-cost-aware mean-variance allocation (Boukardagha, 21 Feb 2026). A simpler HMM-based tactical allocator over SPY, TLT, and GLD estimates a three-state Gaussian HMM selected by BIC, with states interpreted as low-volatility, transitional, and high-volatility; the estimated transition matrix exhibits high persistence, and state-conditional means support regime-conditioned allocation rules (Verma et al., 27 May 2026).

A fourth line uses change-point or task segmentation. ReCAP’s adaptive regime detection module applies a CUSUM-style detector

$K$0

with $K$1 and threshold $K$2, to regime-sensitive market-level features such as VIX, turbulence, Bollinger bands, and RSI-30. Detected variable-length segments become continual-learning tasks whose policy vectors are stored, merged, or discarded in a policy library (Pan et al., 29 May 2026).

A fifth line treats regimes as asset-specific rather than market-wide. The statistical jump model plus XGBoost framework identifies bullish and bearish states separately for each asset using return-derived features and macro covariates, while the KMRF family extends KAMA+MSR detection to ex ante prediction of Bullish, Bearish, and Other states for subsequent portfolio construction or MPC control (Shu et al., 2024, Pomorski et al., 2023).

4. Decision layers and optimization mechanisms

Once a regime has been inferred, RegimeFolio implementations diverge in the mechanism that maps regimes to actions. Some systems select among discrete strategies; others deform the objective, the policy class, or the feasible set.

Variant	Regime signal	Decision layer
Acquisition RegimeFolio	$K$3, $K$4, PRS	Greedy/UCB/Thompson/REIGN selection
FR-LUX	LL/LH/HL/HH	Regime-conditioned RL with trade-space trust region
Sectoral RegimeFolio	VIX Low/Medium/High	Regime- and sector-specific forecasts plus mean-variance QP
Wasserstein HMM system	HMM template probabilities	Transaction-cost-aware mean-variance optimization
ReCAP	ARD-detected variable-length regimes	Policy library plus regime-gate composition

In the BO setting, the decision layer is a portfolio over acquisition functions. PRS orders the explore–exploit choice, the threshold $K$5 induces switching, and the expected advantage on Hit@1 can be approximated as $K$6 with $K$7, subject to domain-specific calibration (Thomas, 6 May 2026).

In friction-aware portfolio optimization, FR-LUX embeds a microstructure-consistent cost model

$K$8

directly into the reward, constrains policy updates with a KL trust region, and adds an explicit trade-space penalty $K$9. The framework provides formal results on optimal stationary policies under convex frictions, monotonic improvement under a trust region, long-run turnover bounds, proportional-cost inaction bands, positive value advantage for regime-conditioned policies, and robustness to cost misspecification (Zhang, 3 Oct 2025).

The sectoral ML variant uses a more classical architecture. Regime- and sector-specific ensemble learners forecast one-step-ahead asset returns, regime-specific covariance matrices are regularized by Ledoit–Wolf shrinkage, and allocations solve a long-only mean-variance problem with budget, position-cap, and turnover constraints. Transaction costs are applied at 10 bps per trade, and turnover is constrained through $\mathrm{PRS}(B,|A|,\rho)=\frac{B}{|A|}(1-\rho),$0 (Zhang et al., 14 Sep 2025).

The explainable HMM system keeps the optimization layer convex but changes the inference layer. Regime probabilities are aggregated into mixture moments,

$\mathrm{PRS}(B,|A|,\rho)=\frac{B}{|A|}(1-\rho),$1

and then used in a transaction-cost-aware mean-variance program with either $\mathrm{PRS}(B,|A|,\rho)=\frac{B}{|A|}(1-\rho),$2 or quadratic trading penalties (Boukardagha, 21 Feb 2026).

Additional RegimeFolio variants extend the control layer in other directions. Regime-weighted conformal risk control wraps any conditional quantile forecaster with exponentially time-decayed and regime-similarity weights to produce daily VaR bounds that target a desired exceedance rate under drift (Schmitt, 3 Feb 2026). DeePM replaces explicit state labels with regime-robust structure: strictly lagged cross-sectional attention resolves ragged filtration, a macroeconomic graph prior regularizes cross-asset dependence, and a SoftMin over window Sharpes acts as a differentiable EVaR-style worst-window penalty (Wood et al., 9 Jan 2026). ReCAP freezes a base policy and a library of regime-specific policy vectors, then learns a regime-gate module that composes

$\mathrm{PRS}(B,|A|,\rho)=\frac{B}{|A|}(1-\rho),$3

with $\mathrm{PRS}(B,|A|,\rho)=\frac{B}{|A|}(1-\rho),$4, updating only the gate and current regime vector in continual trading (Pan et al., 29 May 2026). Other systems use explicit RL environments with regime-aware rewards, capital resets, and shock injections (Raj, 17 Sep 2025), or combine HMM regime inference with tabular policy iteration over a small discrete action set (Verma et al., 27 May 2026). Under label scarcity, a teacher–student pipeline uses a CVaR optimizer as label generator, t-copula synthetic data augmentation, and semi-supervised sandwich training for Bayesian and deterministic student networks (Chattopadhyay, 4 Apr 2026).

5. Empirical evidence

The empirical case for RegimeFolio is strongest where regime variables flip rankings or materially alter risk-adjusted performance. In multi-context Bayesian optimization, the evidence is direct: the same benchmark can reverse the ordering of Greedy and UCB when only budget changes, and the paper’s audit argues that transfer-BO leaderboards are often reporting ATEs over hidden regime mixtures. RegimePlanner operationalizes the regime-aware view and wins all 16 HPO-B search spaces at $\mathrm{PRS}(B,|A|,\rho)=\frac{B}{|A|}(1-\rho),$5, while the pre-registration exercise shows that PRS has substantial predictive content before full comparisons are run (Thomas, 6 May 2026).

In financial portfolio management, the best-documented gains come from systems that align inference, execution costs, and optimization with regime structure. FR-LUX evaluates a 4-by-5 grid of regimes and cost levels with scenario-level aggregation and reports the top average after-cost Sharpe across all 20 scenarios, strictly positive pairwise improvements over vanilla PPO, mean–variance with caps, and risk parity after Romano–Wolf stepdown, a flatter Sharpe–cost slope, and superior risk–return efficiency for a given turnover budget (Zhang, 3 Oct 2025). The sector-specialized RegimeFolio over 34 U.S. large-cap equities from 2020 to 2024 reports 137% cumulative return, annualized return of approximately 18.9%, Sharpe 1.17, max drawdown of $\mathrm{PRS}(B,|A|,\rho)=\frac{B}{|A|}(1-\rho),$6, and a 12% lower maximum drawdown than the S&P 500 benchmark; it also reports 15–20% lower MAE than regime-agnostic ML baselines and ablations showing deterioration when either regime segmentation or sector specialization is removed (Zhang et al., 14 Sep 2025).

The explainable Wasserstein-HMM system reports out-of-sample Sharpe 2.18 versus 1.59 for equal-weight and 1.18 for SPX buy-and-hold, with maximum drawdown of $\mathrm{PRS}(B,|A|,\rho)=\frac{B}{|A|}(1-\rho),$7 versus $\mathrm{PRS}(B,|A|,\rho)=\frac{B}{|A|}(1-\rho),$8 for SPX. During the early-2025 selloff labeled “Liberation Day,” the system shifted away from SPX toward USD, bonds, and gold, and materially reduced peak-to-trough loss (Boukardagha, 21 Feb 2026). ReCAP, evaluated on DOW30, NAS100, SP500, NIKKEI30, and COMMODITY_ETF, reports cumulative returns and Sharpe ratios of 96.76% and 1.00 on DOW30, 164.89% and 1.14 on NAS100, 145.02% and 1.14 on SP500, 133.29% and 1.00 on NIKKEI30, and 97.80% and 0.95 on COMMODITY_ETF, and outperforms retrain, finetune, experience replay, EWC, and CoR continual-learning baselines on average performance (Pan et al., 29 May 2026).

More specialized variants reinforce the same pattern. The HMM-plus-RL allocator over SPY, TLT, and GLD reports out-of-sample cumulative return 131.4%, annualized return 14.3%, volatility 17.3%, Sharpe 0.83, and max drawdown $\mathrm{PRS}(B,|A|,\rho)=\frac{B}{|A|}(1-\rho),$9 for the RL policy, exceeding passive SPY on risk-adjusted terms and lowering drawdowns while remaining interpretable through discrete regime-dependent actions (Verma et al., 27 May 2026). DeePM, trained on 50 diversified futures and FX contracts from 2010 to 2025, reports net Sharpe 0.93 versus 0.50 for Passive Equal Risk and 0.45 for TSMOM, with maximum drawdown $\rho$0 and materially better performance than the Momentum Transformer baseline (Wood et al., 9 Jan 2026). Under label scarcity, the Bayesian sandwich student BNN-S attains Sharpe $\rho$1 on C2A and $\rho$2 on D2A, while maintaining approximately 11.9% weekly turnover versus approximately 23.6% for DNN-S, which the paper interprets as implicit turnover regularization from Bayesian averaging (Chattopadhyay, 4 Apr 2026). On the risk-control side, regime-weighted conformal calibration improves regime-conditional stability of one-sided VaR in some drifting settings, while time-weighted conformal calibration is identified as a strong default when the base quantile forecaster is already flexible (Schmitt, 3 Feb 2026).

6. Limitations, reporting standards, and outlook

A recurring limitation is that regime variables are useful only when they are both observable and stably estimated. In the BO setting, PRS depends on observation noise $\rho$3, is recommended mainly as a within-domain ordering, and requires discretization or density reformulation in continuous domains. The same paper also notes that $\rho$4 is predictive but not exactly sufficient under heterogeneous priors with arm-specific noise, and that high-coverage edge cases or low $\rho$5 can collapse differences between acquisitions (Thomas, 6 May 2026).

In finance, regime definitions themselves are often the weak link. VIX terciles are interpretable but may be sample-sensitive and can miss latent structure; the sectoral RegimeFolio also uses a fixed 34-stock U.S. large-cap universe and a transaction-cost model limited to fixed 10 bps per trade, excluding explicit price impact and slippage beyond that assumption (Zhang et al., 14 Sep 2025). Weighted conformal coverage relies on weighted exchangeability or smooth drift assumptions that the paper explicitly describes as idealized, and aggressive localization can shrink effective sample size enough to destabilize quantile estimates (Schmitt, 3 Feb 2026). Gaussian HMMs can under-represent heavy tails and crisis jumps, while discrete regime boundaries can become ambiguous during gradual transitions (Verma et al., 27 May 2026). ReCAP’s ARD detector can struggle with subtle or gradual transitions, and its similarity-based library maintenance may merge distinct policies that happen to be parametrically close (Pan et al., 29 May 2026). DeePM’s fixed macroeconomic graph prior improves regularization but hard-codes topology, and its strict lag in cross-sectional attention intentionally sacrifices information freshness for robustness under asynchronous closes (Wood et al., 9 Jan 2026).

The most consistent methodological recommendation across the literature is explicit regime-conditioned reporting. In BO, the practical protocol is to report prior condition, $\rho$6, metric family, and context count $\rho$7, and to present CATE conditioned on $\rho$8 rather than only ATE rankings (Thomas, 6 May 2026). In friction-aware portfolio optimization, the corresponding standard is scenario-level inference with bootstrap confidence intervals, multiple-testing correction, and cost calibration from observable liquidity proxies, so that claims remain tied to after-cost, implementable performance rather than frictionless abstractions (Zhang, 3 Oct 2025).

RegimeFolio therefore marks a shift in emphasis from unconditional optimization to conditional decision-making. Whether the object being allocated is an acquisition function, a portfolio weight vector, a risk buffer, or a continual-learning policy vector, the underlying thesis is the same: strategy quality is regime-dependent, and a useful system must expose, estimate, and exploit that dependence rather than average it away.