Research

Regime Detection

Prior art, methods, and data sources for building a daily multi-asset regime classifier and tactical asset allocation engine

This document catalogs the institutional frameworks, quantitative methods, and data sources for regime detection across equities, fixed income, FX, commodities, crypto, and DeFi yield. The goal is not to prescribe a single model, but to map the design space so builders can pick methods suited to their governance constraints, data budget, and asset universe. Written to be model-agnostic. Date of brief: April 2026.

Scope & Audience

Cross-asset regime detection covering equities, fixed income, FX, commodities, crypto, and DeFi yield, with explicit attention to data and indicators that exist as of 2026. Intended audience: analysts or AI agents designing a daily multi-asset regime classifier and tactical asset allocation engine.


1. What "Regime Detection" Actually Means

The term covers three distinct problems that are often conflated. Any build should pick one as the primary target before selecting a method.

Regime Classification

Assign each point in time to one of N discrete states (typically 2–4: risk-on / neutral / risk-off, or expansion / slowdown / contraction / recovery). Output is a label or probability vector. This is the right framing for a top-down TAA tilting engine because allocations are themselves discrete shifts.

Vol/Correlation Regime

Identify shifts in the volatility and correlation structure of asset returns rather than direction. Useful for risk budgeting, less directly useful for directional allocation.

Return Prediction

Forecast next-period return for each asset. Different problem, much harder, generally low signal-to-noise at horizons under 30 days. Should not be the primary signal in a TAA system; can be a soft tilt overlay if backtests support it.

Institutional consensus (Goldman, BlackRock, Bridgewater, AQR): regime classification is more tractable and more explainable than return prediction, and a defensible TAA framework runs allocations off classified regimes rather than off forecasted returns.


2. Institutional Prior Art

The named programs below are the reference points for "what good looks like." They share three properties: composite indicators built from many independent inputs, explicit decomposition (you can always say which sub-indicator drove the signal), and a published methodology that survives external scrutiny.

2.1

Goldman Sachs Financial Conditions Index (GS FCI)

A weighted composite of five components: short-term real interest rate, long-term real interest rate, trade-weighted dollar, credit spreads (corporate), and equity prices. Weights derived from the estimated impact of each component on GDP growth one year ahead.

Published in real time, decomposable into "what is driving conditions today." The methodology has been public for ~25 years and is a model for a transparent composite.

Limitation: focused on US economic conditions, not directly a market-timing signal.

2.2

Chicago Fed National Financial Conditions Index (NFCI)

Composite of 105 financial market indicators across risk, credit, and leverage subindexes. Released weekly. Constructed using dynamic factor analysis. The Adjusted NFCI removes the part explained by current macro conditions, isolating "financial conditions independent of the cycle."

A useful free public benchmark — historical data freely available on FRED. An obvious validation reference for any composite built privately.

2.3

St. Louis Fed Financial Stress Index (STLFSI)

18 weekly inputs (rates, spreads, vol). Updates weekly, freely available. Simpler than NFCI but a useful sanity check.

2.4

BlackRock Macro GPS / Inflation GPS

Directional indicators for global growth and inflation, intended to lead consensus economist forecasts by ~3–6 months. Constructed from "big data" inputs (web search trends, satellite, shipping, etc.) plus traditional macro.

Output is a directional surprise indicator, not a regime label. The methodology is partially proprietary; the publicly explained version uses Bayesian shrinkage to combine many noisy leading indicators into a smoothed signal.

Relevant principle: leading indicators should be aggregated, not picked.

2.5

Bridgewater All Weather / Risk Parity / "Regime Cube"

Allocates by economic regime rather than market regime. The four-quadrant framework crosses growth (rising / falling) with inflation (rising / falling), producing four regimes, each with an asset class that historically outperforms in it.

Allocation balances risk contribution across regimes so the portfolio is robust to which regime materializes. This is structurally different from a tactical tilting engine — it does not try to predict the regime, it tries to be neutral to it.

Worth understanding as an alternative philosophy: instead of "detect risk-off and tilt to safe assets," the answer is "always be balanced across regimes." For a system with strong governance constraints, the All Weather framing is sometimes cleaner because it requires no prediction.

2.6

AQR Style Premia and Regime Overlays

AQR runs systematic style portfolios (value, momentum, carry, defensive) and overlays regime conditioning on them. Their published research argues that regime detection works best for style timing (tilting between value and momentum, e.g.) rather than direct asset class timing.

Methodology is generally tree-based ML with explicit feature attribution. Useful precedent for the explainability tier.

2.7

Two Sigma / Renaissance / D.E. Shaw

Largely opaque. Published research suggests heavy use of: alternative data, ensemble ML (gradient boosted trees), short-horizon mean reversion signals, and order book microstructure. These shops do not publish regime methodologies.

The relevant lesson: black-box ML beats explainable models on raw predictive power but is not appropriate where governance, communication, or auditability matter. A token-governed protocol cannot run on Renaissance-style methods because the model itself must be defensible to voters.

2.8

JP Morgan / Marko Kolanovic-style Positioning Indicators

Composite positioning indices built from CTA exposure estimates, dealer gamma positioning, options skew, and futures COT data. Treated by buyside as contrarian indicators — extreme positioning typically precedes mean reversion.

Less directly applicable to crypto where positioning data is fragmented but the underlying logic (extreme positioning = late-cycle = risk-off lean) carries over to perp funding rates.

2.9

Conference Board Leading Economic Index (LEI)

The original "leading indicators" composite, ten components covering manufacturing, building permits, equities, credit, consumer expectations. LEI has a documented track record of leading recessions by 6–12 months. Free, monthly. Usually included as one input in a broader macro composite.

2.10

Ned Davis Research Five-Factor Recession Probability Model

Logistic regression on five inputs (yield curve, credit spreads, ISM new orders, building permits, ECRI weekly leading index). Outputs probability of recession in next 6 months. Explainable, well-documented, free in summary form.

A common reference for "how to combine 5–10 leading indicators into a single regime probability."


3. Quantitative Methods Catalogue

Methods organized by complexity tier. Each tier trades off explainability against potential predictive accuracy. The most appropriate method depends on governance constraints, available training data, and required cadence.

Tier 1: Fully Explainable

Composite z-score indicator

N indicators, each z-scored over a rolling 3- or 5-year window, signed so positive means risk-off pressure, weighted (equal or by historical predictive accuracy), summed into a composite. Output mapped to discrete regimes by percentile thresholds.

This is the methodology behind Goldman FCI, NFCI, and most central bank financial conditions indices.

  • Strengths: transparent, decomposable, robust to overfitting, easy to validate
  • Weaknesses: linear, doesn't capture interactions, weights need periodic review

Logistic regression on regime labels

If historical regimes are labeled (NBER recession dates, NDR-style binary), fit a logistic regression with leading indicators as features. Outputs probability of being in regime. Explainable via coefficient inspection. Standard in macro. Limitation: requires labeled regime history, which doesn't really exist for crypto.

Rule-based threshold systems

If yield curve inverts AND HY spreads > X AND VIX term inverted, flip to risk-off. Brittle in isolation but useful as confirmation logic on top of a composite.

Tier 2: Conditionally Explainable

Hidden Markov Models (HMM)

Hamilton (1989) pioneered the regime-switching framework. Fit a 2- or 3-state HMM on a panel of indicators; output is the probability of being in each regime at each time. Each regime has its own mean and covariance for the indicators.

Output is interpretable (you can describe what each regime "looks like" in terms of indicator values) but the state assignment itself is statistical, not rule-based. Standard implementations: hmmlearn (Python), regime-switching packages in R.

  • Strength: captures non-linear regime dynamics; well-suited to detecting structural breaks
  • Weakness: number of regimes is a hyperparameter, sensitive to feature selection, probabilities can be unstable on small windows

Markov-Switching VAR (MS-VAR)

Extension that allows the dynamics between indicators to change by regime. More flexible, more parameters, harder to fit reliably with limited data.

Gaussian Mixture Models (GMM)

Cluster historical data points by indicator signature; each cluster is a "regime." Less common than HMM for time series because GMM doesn't model state persistence — adjacent points are treated independently. Useful for unsupervised regime discovery before fixing a number for HMM.

Tree-based ML with SHAP attribution

Gradient boosted trees (XGBoost, LightGBM) on a feature panel predicting regime probability or forward return. SHAP values give per-prediction feature attribution, providing post-hoc explainability. AQR-style overlay logic. Strength: handles non-linearities and interactions. Weakness: needs careful walk-forward validation; easy to overfit on noisy financial data.

Change point detection

Methods like CUSUM, PELT, or Bayesian online change point detection identify structural breaks in time series. Useful for detecting that a regime has shifted, less useful for classifying which regime is current. Often paired with HMM (CUSUM as alarm, HMM for classification).

Tier 3: Low-Explainability, High-Flexibility

LSTM / Transformer sequence models

Deep learning on sequential indicator data. Can capture long-range dependencies. Generally requires more training data than financial markets provide. Highly opaque. Used by quant funds with deep data infrastructure; not appropriate for governance-defensible systems.

Reinforcement learning for allocation

Treat allocation as a sequential decision problem; train an agent to maximize risk-adjusted return. Theoretically appealing, practically fragile (reward hacking, distribution shift, sample inefficiency in low-frequency financial data). Active research area; not production-ready for institutional TAA.

Topological data analysis, multi-fractal methods

Various exotic approaches. Generally not used in production at scale.

Method Selection Guide

ConstraintRecommended Tier
Governance vote / regulatory review requiredTier 1
Need to publish reasoning for each callTier 1
Backtest sample > 20 years, stable structureTier 1 + Tier 2
Backtest sample < 10 years, regime shifts likelyTier 2 (HMM)
Pure performance, no governance constraintTier 2 + selective Tier 3
< 5 years of clean dataTier 1 only

Common institutional pattern: Tier 1 as primary signal, Tier 2 as validation, Tier 3 only if backtest demonstrates statistically significant lift after costs and over a sufficient out-of-sample window.


4. Indicator Universe

Organized by category. Each indicator is tagged Leading (L), Coincident (C), or Lagging (Lg) relative to risk asset prices. Leading indicators belong in regime detection; lagging indicators are for monitoring and confirmation only.

4.1 Macro / Interest Rate

IndicatorL/C/LgSourceNotes
US 10Y real yield (TIPS)LFRED (DFII10)Single most important risk-asset driver post-2022
2s10s yield curve slopeLFRED (T10Y2Y)Recession proxy; inversion-to-recession lag is variable
3m10y curveLFRED (T10Y3M)Fed-preferred recession indicator
5y5y forward inflation breakevenLFRED (T5YIFR)Forward inflation expectations
Federal funds rateCFRED (FEDFUNDS)Policy stance, not predictive
CPI YoYLgFRED (CPIAUCSL)Backward-looking
PCE core YoYLgFRED (PCEPILFE)Fed's preferred inflation gauge
Real GDP YoYLgFRED (A191RL1Q225SBEA)Quarterly, lagged release
ISM Manufacturing PMI new ordersLISM (paid) or FRED proxyStrong recession lead
Initial jobless claims (4w MA)LFRED (IC4WSA)Labor market lead

4.2 Credit

IndicatorL/C/LgSourceNotes
HY OAS (BofA HY index)LFRED (BAMLH0A0HYM2)Highest-information single macro indicator; widens before equities crack
IG OASLFRED (BAMLC0A0CM)Less sensitive than HY
TED spread / SOFR-OISCBloomberg / institutionalBank stress indicator (post-LIBOR transition)
CDX HY series spreadLMarkit / BloombergCredit derivative pricing of distress

4.3 Equity Microstructure

IndicatorL/C/LgSourceNotes
SPX vs 200d MACYahoo / PolygonTrend regime, binary
VIX levelCCBOE / FRED (VIXCLS)Coincident fear gauge
VIX term structure (front/6M)LCBOEInversion = acute fear, high-conviction risk-off marker
SPX % above 200d (breadth)CvariousConfirmation signal
Put/call ratioC/LCBOESentiment, contrarian at extremes
Realized volatility (60d)Ccomputed from pricesVol regime
SKEW indexCCBOETail risk pricing

4.4 FX / Commodities

IndicatorL/C/LgSourceNotes
DXY (dollar index)LICE / YahooStrong dollar = risk-off pressure on EM and crypto
Gold/copper ratioLcomputedSafe-haven vs cyclical demand
Brent / WTI level + changeCICE / NYMEXInflation input + cyclical demand
EM FX indexCvariousRisk-on/off in EM currencies

4.5 Crypto-Native

IndicatorL/C/LgSourceNotes
BTC dominanceCCoinGeckoRising = risk-off within crypto
ETH/BTC ratioLCoinGecko / exchangesRising = risk-on, leads alt rotation
Stablecoin total float changeLDefiLlama, on-chainFresh capital signal; faster than addresses
Aggregate perp funding ratesLCoinglass / VeloExtreme positive = late-cycle euphoria; sustained negative = capitulation
Open interest in BTC/ETH perpsCCoinglass / VeloLeverage in the system
BTC MVRV z-scoreC/LGlassnodeLong-cycle positioning
BTC realized capLgGlassnode / CoinMetricsCost basis of holders
Spot BTC ETF net flowsLissuer disclosures (BlackRock IBIT, etc.)New in 2024+; clean institutional signal
Spot ETH ETF net flowsLissuer disclosuresNew in 2024+
ETH staking ratio / queueCbeaconcha.in / NansenSlow-moving positioning
L1/L2 transaction countsCvarious block explorersActivity, not directional
L1/L2 fees revenueCToken TerminalFundamental valuation input
DeFi TVL by chainCDefiLlamaCapital deployed
DEX/CEX volume ratioCvariousCrypto-native vs centralized appetite

4.6 2026-Specific Additions

IndicatorL/C/LgSourceNotes
Agent-economy activity (e.g. Virtuals aGDP)LVirtuals dashboardNew asset-class-specific signal
Agent token sector indexCconstructedRobot Money's most relevant micro-cap proxy
Stablecoin yield spreads (Aave vs Morpho vs Compound)CDefiLlama yields APIBucket A optimization input
Restaking TVL (EigenLayer, Symbiotic)CDefiLlamaRisk-seeking yield measure

4.7 Sentiment and Positioning (Use with Caution)

IndicatorL/C/LgSourceNotes
AAII bull/bear surveyCAAIIRetail sentiment; contrarian at extremes
CFTC COT positioningLCFTCInstitutional positioning, weekly
Crypto Fear & Greed IndexCalternative.meCrypto-specific composite
Farcaster activityCNeynar APINative attention measure for agent-token sector

5. Data Infrastructure for 2026

Free / Public

  • FRED (St. Louis Fed): comprehensive macro, rates, credit. Highest-quality free source.
  • Yahoo Finance / yfinance: equities, ETFs, indices. Reliable but rate-limited.
  • DefiLlama: DeFi TVL, yields, protocol revenue, stablecoin float. Best free crypto fundamentals source as of 2026.
  • CoinGecko: prices, market caps, volumes, basic on-chain. Free tier sufficient for most use.
  • Etherscan / Basescan / Arbiscan: raw on-chain data, free.
  • ETF issuer disclosures: spot BTC and ETH ETF flows published daily by BlackRock, Fidelity, Bitwise, etc.
  • CFTC: futures positioning, COT reports, free weekly.
  • Treasury.gov: yield curve daily, issuance.

Paid, Low Cost (under $500/month)

  • Coinglass or Velo: derivatives data — funding rates, perp OI, basis, options IV. Essential for any serious crypto regime model.
  • Token Terminal: protocol fundamentals (revenue, P/F, P/S). Useful for Bucket C fundamentals.
  • CoinMetrics community/pro: on-chain metrics, network value ratios.

Paid, Institutional

  • Glassnode or Nansen: deep on-chain analytics, holder behavior, exchange flows. ~$1k+/month for serious tier.
  • Kaiko: institutional-grade crypto pricing and order book data.
  • Bloomberg Terminal or Refinitiv Eikon: full tradfi coverage. ~$24k/year per seat.
  • Dune Analytics (paid tier): custom on-chain SQL queries at scale.
  • Allium: multi-chain on-chain data warehouse.

Recommendation by Stage

  • v0 build (proof of concept): FRED + Yahoo + DefiLlama + CoinGecko + Coinglass. Sub-$200/month.
  • v1 production: add Token Terminal, expanded Coinglass tier. ~$500/month.
  • v2 institutional: add Glassnode and Kaiko. ~$2-3k/month.

6. Allocation Methods

Once a regime is detected, weights need to translate into portfolio positions. Brief survey of methods used by institutional TAA:

Mean-Variance Optimization (Markowitz)

The textbook method. Maximize expected return for given variance, or minimize variance for given return. Inputs: expected return vector and covariance matrix. Generally inappropriate for TAA because outputs are notoriously sensitive to expected return inputs (which are unstable), produces corner solutions, and is hard to defend. Used in academic illustrations more than in production.

Minimum Variance

Drop the expected return input; just minimize portfolio variance subject to weight constraints. More robust than full MVO. Conservative — produces low-vol but also low-return portfolios. Good baseline for a defensive sleeve.

Risk Parity (Equal Risk Contribution)

Weight assets so each contributes equally to portfolio risk. Bridgewater All Weather is the canonical implementation. Robust to expected return uncertainty. Standard reference benchmark for cross-asset allocation.

Hierarchical Risk Parity (HRP)

Lopez de Prado (2016). Cluster assets first by correlation distance, then allocate hierarchically using inverse-variance within and between clusters. Doesn't require expected returns, handles noisy correlation matrices well, robust to small universe sizes. Best-in-class for token-level allocation within sleeves where correlations are unstable.

Black-Litterman

Bayesian framework combining a market-equilibrium prior with analyst views. Produces tilted portfolios that are stable under uncertainty. Requires a market-cap baseline (reasonable for tradfi, awkward for crypto micro-caps). Useful for layering regime-conditional views onto a strategic baseline.

Strategic Asset Allocation + Tactical Bands (Rule-based)

Define a SAA (neutral portfolio). Define per-asset bands (min and max weights). Compute tactical weight = SAA + (max tilt) × f(regime z-score), where f is typically tanh to bound tilts. Re-normalize. This is what most institutional GTAA shops actually do. Transparent, defensible, easy to backtest, easy to communicate to governance.

Recommended Pattern for Regime-Driven TAA

  • Top layer: SAA + bounded tactical tilts via tanh of regime composite. Rule-based, defensible.
  • Within-sleeve: HRP for token-level allocation in equity/crypto sleeves.
  • Within stables: linear program for yield maximization with concentration and liquidity constraints.

7. Explainability Tiers and Governance Fit

Defensibility varies by method. A token-governed protocol or institutional fund needs to explain decisions at three depths:

To Governance / Voters

"Why are we more defensive this month?" Answer must be one sentence, citing 2–3 specific indicators. Only Tier 1 methods support this natively.

To Analysts / Due Diligence

"What's the model's hit rate, and how does it perform in each regime?" Walk-forward backtest with regime-conditional performance. All tiers support this if backtested properly.

To Regulators / Auditors

"What inputs drive the model and how are they computed?" Full data pipeline traceability. Tier 1 best, Tier 2 acceptable with documentation, Tier 3 problematic.

Regulatory note: The SEC has not specifically opined on AI-driven asset management for tokens, but the emerging EU AI Act treats financial decision systems as high-risk, requiring documentation and explainability. Tier 1 + Tier 2 architectures will age better against regulatory drift than Tier 3.


8. Catalogued Examples of Prior Practitioner Approaches

Documenting specific implementations as data points in the survey, neutrally and without ranking. Each is one approach among many, included to illustrate the range of choices made by different practitioners.

Goldman Sachs FCI

Tier 1 composite, 5 macro inputs, weighted by GDP impact. Real-time. Public methodology since ~2000. Strengths: simplicity, transparency, GDP-grounded weights. Limitations: no asset-class allocation output, limited crypto applicability.

Chicago Fed NFCI

Tier 1 + dynamic factor analysis (mild Tier 2). 105 inputs. Weekly. Free, public. Often used as a benchmark for private composites.

Bridgewater All Weather Framework

Risk parity allocation across regime-tagged assets. Does not predict the regime; allocates to be neutral to it. Philosophy alternative to tactical tilting.

NDR Five-Factor Recession Model

Tier 1 logistic regression on 5 macro leading indicators. Outputs recession probability. Documented hit rate and false-positive rate.

Consensys "Aggregate Risk" Framework (2023, Sokolin)

Four-quadrant qualitative composite (Cycle / Global Macro / Crypto Macro / Web3 Fundamentals), each scored 1–5 by analyst, summed monthly. Tier 0 (subjective). Strengths: communicable, decomposable, captures Web3-specific category that institutional frameworks lack. Limitations: monthly cadence, subjective scoring drifts with consensus, lagging by construction. Documented as one example of a qualitative composite.

Consensys ETH Factor Valuation Model (2022)

Multi-method ensemble for single-asset price target (not regime classification). Methods: factor regression, Monte Carlo, DCF, multiples. Factor regression used 10 inputs blending crypto-native (active addresses, MVRV, network fees, supply in smart contracts, exchange outflow, ETH inflation, ETH staked) with tradfi (SPX, US 2Y forward rate). Strengths: ensemble framing, explicit cross-asset blend, range output rather than point estimate. Limitations: targets price not regime, did not distinguish leading from lagging indicators, single-asset focus, struggles with structural breaks.

Glassnode / Coinmetrics On-Chain Composite Indices

Various composites combining MVRV, NVT, SOPR, realized cap, etc. Tier 1. Crypto-native, no tradfi inputs. Useful as crypto-only sub-composite.

Two Sigma / Renaissance / D.E. Shaw Style

Tier 3 ensemble ML on alternative data. Strong predictive performance, opaque, inappropriate for governance-driven systems.


9. Selection Criteria

To pick among methods, evaluate against the following constraints in order:

  1. Governance / explainability requirement: if voters or regulators must understand each decision, restrict to Tier 1.
  2. Available training data: HMM and ML methods need 10+ years of stable data per regime; crypto only has ~7–10 years usable depending on cutoff. Prefer Tier 1 if sample is thin.
  3. Cadence requirement: real-time → composite z-score; daily → composite or HMM; weekly → any; monthly → any.
  4. Operator override needed: rule-based and composite methods support clean overrides; ML methods don't.
  5. Asset universe complexity: single asset → factor regression viable; cross-asset → composite indicator + per-asset tilts.
  6. Compute / engineering budget: Tier 1 runs in seconds on a laptop; Tier 3 may need GPU and pipeline infrastructure.

10. Recommended Starting Architecture

For a daily multi-asset TAA engine with governance constraints (the Robot Money use case being one example):

Layer 0 — Strategic Asset Allocation

Define neutral portfolio across sleeves (e.g., stables / large-cap crypto / equities / commodities). Set tactical bands per asset. This is policy, not modeled.

Layer 1 — Composite z-score regime classifier (Tier 1)

Two sub-composites:

  • Macro composite: 5–7 leading indicators (real yields, HY OAS, yield curve, DXY, PMI new orders)
  • Crypto composite: 4–6 crypto-native (stablecoin float change, ETH/BTC, funding rates, spot ETF flows)

Each indicator z-scored over 3y rolling window, signed, weighted (equal as starting point), summed. Map to discrete regimes by percentile. Combine sub-composites into asset-specific tilts.

Layer 2 — HMM validation (Tier 2)

Fit 3-state HMM on the same indicator panel. Use as a sanity check, not a primary signal, until 6+ months of operation establish where it agrees and disagrees with Layer 1.

Layer 3 — Manual override channel

Operator can override the systematic call with a documented, dated, public reason. Captures qualitative information the data hasn't priced yet.

Allocation Output Formula

target_weight = SAA + tactical_band × tanh(regime_zscore)

Normalized to sum to 1.

Backtest discipline: walk-forward, with realistic transaction costs. Benchmark against 60/40, all-stables, fixed SAA (no tilts), and naive risk parity. The fixed-SAA benchmark is the critical one — it tests whether the regime model itself adds value, separate from the asset mix.

Output cadence: indicators and regime probabilities computed daily; allocation changes recommended weekly (avoid daily churn at low TVL); strategic weights reviewed monthly.


11. Open Questions and Known Limitations

Crypto Sample Size Remains Thin

All cross-asset regime models for crypto are working with ~7–10 years of data and only 1–2 full cycles. HMM and ML approaches are statistically underpowered. This is structural and won't resolve for several more cycles.

Macro/Crypto Correlation Regime is Non-Stationary

The 2020–2022 high-correlation period (crypto as risk asset) was followed by 2023–2024 partial decoupling (BTC as inflation hedge / debasement trade). Models calibrated on one period mis-specify the other.

Spot ETF Flows as a Signal are Post-2024 Only

No historical analog. Cannot be backtested against prior cycles.

Agent-Economy Indicators are Post-2025 Only

Virtuals aGDP, agent sector indices — will need 2–3 years of history to be statistically usable.

Fed Policy Regime Shift Post-2022

From ZIRP to active rate management changes the relevance of pre-2022 macro relationships for the crypto-macro link. Limit cross-asset crypto backtests to ~2018+ (post-ICO normalization, post-CME futures).


12. References

Foundational Academic

  • • Hamilton, J. D. (1989). "A New Approach to the Economic Analysis of Nonstationary Time Series." Econometrica. — HMM regime switching origin.
  • • Stock, J. H., & Watson, M. W. (2002). "Macroeconomic Forecasting Using Diffusion Indexes." — dynamic factor models.
  • • Lopez de Prado, M. (2016). "Building Diversified Portfolios that Outperform Out-of-Sample." — HRP origin.
  • • Chen, N. F., Roll, R., & Ross, S. A. (1986). "Economic Forces and the Stock Market." — multi-factor model template.

Institutional Methodology

  • • Hatzius, J. et al. (2010). "Financial Conditions Indexes: A Fresh Look after the Financial Crisis." — Goldman FCI methodology.
  • • Brave, S. & Butters, R. A. (2011). "Monitoring Financial Stability: A Financial Conditions Index Approach." — Chicago Fed NFCI methodology.
  • • Dalio, R. (2012). "An In-Depth Look at Deleveragings." — Bridgewater regime framework.
  • • BlackRock Investment Institute. "Macro Insights" series. — GPS methodology overview.

Crypto-Native

  • • Burniske, C. & Tatar, J. (2017). Cryptoassets. — early framework for crypto valuation.
  • • Coinmetrics "State of the Network" reports. — on-chain factor methodology.
  • • Glassnode Insights. — MVRV, SOPR, realized cap definitions and applications.

Data Sources (Active Links)

  • • FRED: https://fred.stlouisfed.org
  • • DefiLlama: https://defillama.com
  • • Coinglass: https://www.coinglass.com
  • • Token Terminal: https://tokenterminal.com
  • • CFTC COT: https://www.cftc.gov/MarketReports/CommitmentsofTraders

Appendix A: Glossary

CVaR: Conditional Value at Risk; expected loss in the worst α% of scenarios.
EWMA: Exponentially Weighted Moving Average.
GTAA: Global Tactical Asset Allocation.
HMM: Hidden Markov Model.
HRP: Hierarchical Risk Parity.
Ledoit-Wolf shrinkage: Method to stabilize covariance matrix estimation.
MVRV: Market Value to Realized Value ratio (crypto on-chain valuation metric).
NFCI: National Financial Conditions Index (Chicago Fed).
OAS: Option-Adjusted Spread (credit).
SAA: Strategic Asset Allocation (neutral baseline weights).
TAA: Tactical Asset Allocation (deviations from SAA based on conditions).
z-score: Standardized value: (x − rolling mean) / rolling standard deviation.

End of exhibit. This document is a methodology survey, not a product specification. Use as reference material when scoping a regime detection system; pick methods based on the constraints documented in Section 9 and the recommended starting architecture in Section 10.

Date of brief: April 2026 · Document type: Reference exhibit / methodology survey