The Bear Regime Signal™ identifies contemporaneous market regimes using a 25-factor universe spanning macroeconomic, market, and positioning data. The BRS factor set is partitioned into a Core-8 (yield curve, credit spreads, financial conditions, employment momentum, real consumption, equity volatility, breadth, and trend strength) and a Sub-17 (secondary macro, valuation, sentiment, and cross-asset positioning factors). All public features are sourced from FRED and a premium market data feed and processed monthly through a deterministic preprocessing pipeline that includes z-score standardization on rolling 24-month windows to handle factor drift across regimes.

The classifier is Cronus Regime Classifier (CRC) (sparse regularization, regularized), selected for two properties: sparsity, which produces interpretable factor weights and avoids overfitting in a 25-feature space with limited bear-regime observations; and robustness, which prevents any single factor from dominating the decision boundary. Training is conducted under strict expanding-window walk-forward protocol with monthly refitting — at month t, the model uses only data through month t−1, with no in-sample tuning on out-of-sample observations. This eliminates lookahead bias and produces honest out-of-sample performance metrics.

Operating-point selection follows an a priori criterion: minimum bear recall of 70%, then minimize false-positive rate. The resulting threshold (BRS composite ≤ 22) achieves 86% precision and a 2.5% false-positive rate across 304 out-of-sample months (2000–2026), with AUC 0.91. The signal is corroborated across alternative algorithms (L2 logistic, random forest, gradient boosting), all producing AUC in the 0.89–0.92 range, confirming the signal lies in the factor set rather than the algorithm choice. All eight bear cycles between 2000 and 2025 fall within the validation window; seven were classified within the documented 1–3 month coincident-to-near-coincident window.

The Bear Regime Signal™ is the only forward-looking model in the platform, predicting the probability of a 4-month-ahead drawdown event during bull regimes. MBS operates as two parallel binary classifiers: a 10%-correction model and a 5%-drawdown model. Each classifier is trained on the same 34-feature universe — a superset of the BRS/RRS factor space, augmented with technical momentum factors, valuation ratios (CAPE, ERP), VIX term-structure measures, and microstructure indicators that prior research identifies as leading drawdown signals.

The architecture is deterministic non-linear ensemble (subsample=1.0, =4, =0.08, configuration=100, scale_pos_weight=balanced, random_state=42), selected over linear alternatives because drawdown-precursor signals exhibit strong nonlinear interactions — particularly between volatility regime, valuation, and momentum — that linear models cannot capture cleanly. This is empirically confirmed: L1-Logistic regression corroborates the 10%-correction signal at AUC 0.88 but reaches only AUC 0.769 on the 5%-drawdown target, where nonlinear architecture is essential. The departure from the L1 backbone used in BRS/RRS is disclosed honestly as a methodological choice driven by the smaller-magnitude target.

The 4-month forward horizon was selected over alternatives (3, 5, and 6 months) through systematic out-of-sample comparison. Validation is walk-forward across 154 out-of-sample months (2013–2025), capturing every major drawdown event of the past decade — Q4 2018, the 2020 pandemic crash, the 2022 bear, and the 2025 Tariff Shock — with 1–2 months of lead time before market peaks. The 10%-correction model achieves 84.4% precision at 4.1% false-positive rate (AUC 0.916); the 5%-drawdown model achieves 83.3% precision at 6.5% false-positive rate (AUC 0.92). The shorter validation window relative to BRS/RRS reflects the data-availability constraint of the underlying drawdown definition, which requires consistent valuation, VIX, and credit data not uniformly available before 2013.

The Regime Recovery Signal™ classifies the current market regime across a seven-stage scale (Critical Bear → Strong Bull) and produces a high-conviction bull-recovery confirmation signal. The model shares the 25-factor universe with BRS but is trained against a different target: bull-regime classification rather than bear detection, requiring the classifier to identify sustained recovery rather than crisis onset.

The architecture is a stacked ensemble. An Cronus Regime Classifier (CRC) backbone is augmented with a walk-forward deterministic non-linear ensemble probability as a 26th feature, with the L1 stage learning the appropriate weighting between linear factor contributions and nonlinear interaction effects captured by the gradient-boosted model. This design preserves the interpretability of the linear backbone while permitting nonlinear pattern recognition in transition periods, where regime boundaries are inherently fuzzy. Walk-forward training is conducted under the same expanding-window protocol as BRS, with monthly refitting on observable history only.

The operating point (xgb_probability ≥ 0.75) balances precision (82%) with recall (39%), prioritizing precision over sensitivity. At this threshold the model achieves 82% precision and an 1.5% false-positive rate across 154 out-of-sample months (Jan 2013 – Oct 2025), with AUC 0.76. Operating-point and hyperparameter selection on same OOS as reported metrics — explicitly disclosed in spec file. The threshold reflects an explicit product-design choice: in commercial deployment, the cost of a false bull-recovery signal (premature re-engagement during a continuing bear) is asymmetrically higher than the cost of a missed signal. The conservative calibration produces a high-conviction confirmation signal — RRS stays silent during ambiguous recoveries and fires only when factor evidence supports a sustained bull regime. Cross-validation against BRS prevents bull confirmation from firing while bear detection is active, eliminating contradictory signal states.