Validation, Methodology & Research
Market signals earned through research, built on innovation & validated on statistical science.
RegimeSignal™ maps the S&P 500's full market cycle in advance. This is the research protocol behind it — walk-forward, expanding window, monthly refit, robustness testing, and continuous monitoring. What decides what becomes a signal.
01
Walk-Forward Testing
Every signal is evaluated under strict walk-forward protocol — no peeking, no in-sample tuning of out-of-sample windows.
02
Expanding-Window Validation
Models are validated across an expanding history so each call is judged against everything previously knowable.
03
Historical Regime Analysis
Behaviour benchmarked across full bull, weakening, correction and recovery cycles since the late-1990s.
04
Monthly Refit Methodology
A controlled, calendared refit cadence — frequent enough to adapt, slow enough to prevent overfitting to recent noise.
05
Signal Robustness Testing
Each signal is stress-tested against parameter perturbation, sub-period sampling and feature ablation before shipping.
06
Drawdown Analysis
Peak-to-trough behaviour studied around every historical trigger — what the call would have caught, missed, and signalled too early.
07
Recovery Cycle Analysis
Regime Recovery Signal (RRS) confirmation behaviour mapped across recoveries — characterizing both true confirmations and quiet false-starts.
08
Multi-Regime Performance Review
Precision, hit-rate and forward-window distributions reported per regime, not just as a single blended number.
09
Research Discipline Framework
Hypothesis → protocol → walk-forward → robustness → review. Every change passes the same gates, in the same order.
10
Adaptive Validation Procedures
When the data-generating process changes, validation windows and refit cadence adapt — explicitly, not silently.
11
Continuous Model Monitoring
Live signals are monitored for pre-trigger drift, conformal-uncertainty widening and out-of-distribution conditions.
12
Structural Research Architecture
A versioned research stack where every signal, refit and decision is logged and reproducible end-to-end.
Factor Architecture · Reference
25–35 factor predictive structure
How the model is built · BRS shown
Proprietary methodology
Portions of the RegimeSignal™ methodology — including the specific algorithm implementations, model configurations, and factor weightings — are proprietary trade secrets, and the descriptions here are intentionally generalized to protect intellectual property that is highly sought-after in the marketplace. This is by design, not omission: the methods have been independently validated and reproduced bit-for-bit under audit, and full technical specifications are available to qualified reviewers under NDA.
Each RegimeSignal™ signal operates on its own purpose-built factor universe, ranging from 25 to 35 factors depending on the signal (BRS = 25, MBS Tier 1 / Tier 2 = 34, RRS = 35). The example shown here is the Bear Regime Signal (BRS) universe: 25 predictive factors organized as 8 Core Macro Drivers (70% weight) and 17 Market Drivers (30% weight).
A proprietary sparse-regularization technique selects the signal-carrying factors and drives the rest to zero — producing a sparse, transparent architecture. Weights are refit monthly via expanding-window walk-forward validation.
Signal quality is independently corroborated by three alternative learning methods on the same data — all landing in a tight AUC 0.89–0.92 band across 304 out-of-sample months (Dec 2000 – Apr 2026). Proprietary factor constructions are not disclosed in full; methodology is reproducible from source data.
8 Core Macro Drivers · 70% Weight
- • Inflation (18% — highest weight)
- • Corporate Earnings (18%)
- • Federal Reserve Policy
- • Valuation (Forward P/E)
- • Consumer Sentiment
- • Economy (GDP/LEI)
- • Government Policy
- • Liquidity & Financial Conditions
17 Market Drivers · 30% Weight
- • Credit Spreads (HY OAS)
- • Market Breadth
- • VIX Term Structure
- • Yield Curve (10y2y)
- • NFCI Leverage
- • Equity Trend
- • Equity Momentum
- + 10 more market & technical factors
Architecture · Narrative
Twenty-Five Factors. Eighty-Plus Live Streams. Proprietary Weightings.
Most regime models claim a factor count. RegimeSignal™ shows the architecture underneath.
The four walk-forward validated signals — BRS, MBS Tier 1, MBS Tier 2, and the Regime Recovery Signal™ — each operate on their own purpose-built factor universe, ranging from approximately 25 to 35 factors per signal, with set composition calibrated to each signal's prediction horizon and target regime event. Those factors are not raw observations. They are the engineered output of 80+ live economic and market data streams flowing continuously from 14 institutional and proprietary sources: FRED macro data, premium market feeds, real-time equity, credit, and volatility prices, energy curve providers, government statistical agencies, geopolitical event databases, and a live three-AI council. Each factor is constructed, normalized, and z-score standardized through a deterministic preprocessing pipeline — no curve fitting, no lookahead, no judgment overlays applied to the quantitative core.
The factor architecture is transparent and audit-ready. The data sources are documented. The methodology is reproducible. The weights — the proprietary sparse-regularized coefficients that determine how each signal's factors combine into a single score — are the institutional edge. Developed over three decades. Refined across eight bear cycles (walk-forward OOS validation covers 2000–2025; the 1990 cycle sits in pre-OOS training history). Validated walk-forward at AUC 0.91. And not for disclosure.
Proprietary Algorithm Architecture
Inside the engine — a proprietary algorithm architecture
RegimeSignal™ runs a multi-engine quantitative architecture: four deployed production signals, a deterministic macro-guardrail floor, and three independent corroborating learners. We name our engines — their internals are proprietary.
BRS Engine
Cronus Regime Classifier (CRC)
A sparse, self-pruning regularized linear classifier that drives low-signal factors to zero for a sparse, interpretable model. Refit monthly, walk-forward. Walk-forward AUC 0.91 across 304 out-of-sample months; caught 8 of 8 financially-driven bear cycles.
BRS Layer 2
8-Factor Macro Guardrail Floor
A deterministic rules layer on Federal Reserve data that must independently corroborate before a Bear classification can be issued. Not a machine-learning model — a structural safety floor.
Market Break & Recovery Engines
Cronus Algorithmic Analytics Engine (CAE-1, CAE-2, CAE-R)
Deterministic non-linear ensemble classifiers on 34–35 lagged factors. CAE-1 predicts −5% pullbacks, CAE-2 predicts −10% corrections, CAE-R predicts +10% recoveries — each fully reproducible under audit.
Corroboration
Independent Alternative Learners (IAL)
Three independent learning methods run on the identical walk-forward data. All land in a tight AUC band (0.89–0.92), confirming the signal lives in the factor set, not the algorithm — an internal robustness check, not a deployed signal.
Proprietary methodology
The specific algorithm implementations, model configurations, and factor weightings are proprietary trade secrets and are intentionally described here in generalized terms to protect intellectual property that is highly sought-after in the marketplace. The methods have been independently validated and reproduced bit-for-bit under audit; full technical specifications are available to qualified reviewers under NDA.
Independently validated and audited → PhD Validation · PhD Audit
Quantitative Layer · Algorithm Catalog
The full deployed architecture — every algorithm, disclosed
The RegimeSignal™ platform deploys a multi-algorithm quantitative architecture spanning 4 walk-forward validated production signals, 3 corroborating algorithms (cross-checks against the BRS deployed signal), and 2 within-state directional engines (Bull Velocity, Bear Velocity). None of these layers are AI — they are deterministic, rules-based mathematical processes built into the model architecture. The AI interprets outputs; these layers compute them.
Every metric below is bit-exact reproducible from locked training data and locked spec files. Reproduction scripts published in the PhD Evaluator Package match locked predictions to within 1e-6 probability difference. Aggregate ~84% avg precision · ~4% avg FPR across the 4 production signals.
Deployed production signals · 4 walk-forward validated
Cronus Regime Classifier (CRC) — BRS Layer 1 (deployed)
Deployed signal generator for BRS bear-regime classification. A sparse, self-pruning regularized linear classifier that automatically drives low-signal factor weights to zero, producing a sparse and interpretable model — fit monthly on historical data only, no lookahead. Output is the BRS composite score on a 0–100 scale. Walk-forward AUC 0.91 across 304 OOS months (Dec 2000 – Apr 2026). Caught 8 of 8 financially-driven bear cycles.
8-Factor Macro Guardrail Floor — BRS Layer 2 (deployed)
Deterministic structural-conditions check that must independently corroborate before Layer 1 can issue a Bear classification. Each guardrail is a rules-based threshold on FRED economic data: GDP YoY, INDPRO YoY, FFR 3-mo change, UMCSENT level, CPI YoY, WTI YoY, HY OAS, Fwd P/E. At zero breaches the system is floored at Neutral; at one breach maximum classification is "Marginally Bearish"; only at ≥2 simultaneous breaches is the full bearish range unlocked. Bear Alert trigger: BRS_composite < 22 AND guardrail_breach_count ≥ 2. This two-layer structure is the architectural feature distinguishing BRS from a single-layer ML classifier.
CAE-1 — MBS T1 (5% Pullback Predictor, deployed)
Deterministic non-linear ensemble classifier targeting P(any month within next 4 months is part of a 5% peak-to-trough drawdown in S&P 500). Hyperparameters locked in spec file; deterministic stochastic subsampling makes the model fully reproducible under audit. 34 features (price, volatility, credit, macro, sentiment, valuation, positioning) all lagged 1 month before training. Triggers Neutral state at locked probability threshold. Locked metrics: precision 83% · FPR 6.5% · recall 76% · AUC 0.88 · 154 OOS months (Jan 2013 – Oct 2025).
CAE-2 — MBS T2 (10% Correction Predictor, deployed)
Deterministic non-linear ensemble classifier — identical architecture, hyperparameters, and feature set to MBS T1. Only the target variable definition and decision threshold differ. T2 targets P(any month within next 4 months is part of a 10% peak-to-trough correction). Same 34 features, same walk-forward protocol, same OOS window. Triggers Underperform state at locked probability threshold. Locked metrics: precision 84% · FPR 4.1% · recall 82% · AUC 0.92 · 154 OOS months. Caught all 4 corrections in OOS (Q4 2018, COVID, 2022 Bear, 2025 Tariff).
CAE-R — RRS (10% Outperform Predictor, deployed)
Deterministic non-linear ensemble classifier targeting P(any month within next 4 months reaching +10% peak-to-trough surge in S&P 500 from current level). Hyperparameters in the same family as MBS T1/T2; locked in spec file. 35 features distinct from the MBS feature set — additional momentum and volatility-regime indicators more relevant to upside detection. Triggers Outperform state at locked probability threshold. Locked metrics: precision 82% · FPR 1.5% · recall 39% · AUC 0.76 · 154 OOS months.Disclosed limitations: Recall (39%) is materially lower than the bear-side signals — design-intentional precision-first calibration. Operating point and hyperparameter selection were performed on the same OOS window the published metrics report on; explicitly disclosed in spec file and one of the explicit reasons for commissioning the independent PhD review.
Corroborating algorithms · cross-checks against BRS deployed signal
The deployed BRS signal is independently corroborated by three alternative algorithms run on identical walk-forward OOS data. All four algorithms produce AUCs in a tight 0.89–0.92 band — confirming the signal lies in the 25-factor set, not in the algorithm choice. The CRC method was chosen as deployed for sparsity, interpretability, and reproducibility under audit.
IAL Method 1 — corroborating
Alternative regularized linear method run on identical 25-factor data and identical walk-forward protocol. Produces walk-forward AUC in the same 0.89–0.92 range — tight corroboration of the deployed signal. Not deployed; the sparse CRC method was chosen for automatic factor selection and interpretability.
IAL Method 2 — corroborating
Independent nonlinear tree ensemble run on the same 25-factor set and same walk-forward protocol. Produces AUC in the 0.89–0.92 band — confirming the signal survives algorithm substitution into nonlinear space. Higher FP rate at the selected operating point versus the deployed CRC demonstrates the cleaner calibration achievable with sparse linear methods on the BRS factor set.
IAL Method 3 — corroborating
Boosted ensemble method run on the same 25-factor set. Walk-forward AUC tracks CRC deployed results in the 0.89–0.92 band. Confirms the signal is in the factor set — not specific to linear vs. nonlinear architectures, not specific to bagging vs. boosting. Cross-algorithm corroboration band is tight enough that the signal cannot be attributed to algorithm-specific overfitting.
Within-state directional engines · not formal signals
Both engines below are deployed as directional context only. Neither triggers a state transition. They are excluded from the 4-signal aggregate precision headline.
Bear Velocity — bear-to-recovery directional gauge
Bear-to-recovery directional gauge. 21 LOO-CV samples (small, COVID-dominated), locked threshold. Not walk-forward validated. Used as confirmatory directional context that often precedes RRS by weeks during recovery transitions — never as a fire signal. Powers the Bear Velocity directional gauge.
Bull Health Tracker (BHT) — 0–100 score
Descriptive 0–100 score paired with Bull Velocity, calibrated against 33 years of bull history (310 clean bull months, 1993–2026). Quartiles Q1 60–67.5, Q2 67.5–75, Q3 75–82.5, Q4 82.5–90. Provides the descriptive lens to Bull Velocity's quantified lens — same evidence, two views.
Structural conclusion
Across the 4 deployed production signals, average precision ~84% · average FPR ~4% (simple mean: BRS 2.5%, MBS T1 6.5%, MBS T2 4.1%, RRS 1.5%) · average forward window ~4 months. BRS is the always-on state classifier; MBS T1, MBS T2, and RRS fire as directional state triggers (Neutral, Underperform, Outperform respectively). Each signal validated separately on its own out-of-sample window. Cross-algorithm corroboration confirms the BRS signal sits in a tight 0.89–0.92 AUC band across three independent alternative learners on identical data — the signal is in the factor set, not the algorithm choice.