Calibrated Ridge-Logistic VIX Stress Defensive SPY Allocation

Strategy Card

Calibrated Ridge-Logistic VIX Stress Defensive SPY Allocation — strategy card

Spec ID: spec-calibrated-volatility-regime-defensive-allocation-us-stocks-1789597980 · Generated: 2026-09-16 23:29 UTC

Cluster: Defensive Allocation · Sub Cluster: Calibrated Vix Stress Defensive Allocation

One-line description

A class-weighted ridge-logistic model estimates the probability of near-term VIX stress and continuously scales a long-only SPY allocation, leaving the balance in cash. Volatility targeting, drawdown control, and exposure smoothing further reduce risk.

Why this trade exists

Volatility stress tends to cluster, while equity drawdowns, realized volatility, liquidity, volume, and VIX state variables can contain information about elevated near-term stress risk. Investors may adjust slowly or remain structurally fully invested, allowing a systematic overlay to reduce equity exposure when estimated stress risk rises.

The strategy seeks to exchange some equity-market upside for lower volatility and drawdown. Temperature calibration converts classifier scores into probabilities, while continuous sizing avoids relying solely on a brittle binary threshold. This benefit is conditional: false alarms can leave the portfolio underinvested during rallies, and sudden shocks can occur before the predictors react.

Algorithm

[code omitted from public view]

Parameters

Param	Value	Notes
Rule type	CALIBRATED_VIX_STRESS_DEFENSIVE_ALLOCATION	Class-weighted ridge-logistic feature-subset model with temperature calibration and continuous defensive sizing.
Tradeable universe	SPY&US&ETF, CASH	SPY is the sole risky asset.
Signal universe	SPY&US&ETF, VIX&US&INDEX	VIX is signal-only and cannot be held.
Bar / rebalance	1d / close	Daily MOC execution.
Backtest window	2020-01-01..2024-07-01	New out-of-sample extension, not the paper sample.
Development split	60% / 20% / 20%	Chronological training, validation, and internal untouched test before 2020-01-01.
Stress target	H=5, W=252, q=0.80	Intended event is a future five-day VIX maximum crossing the causal rolling 80th-percentile threshold.
Candidate predictors	28	Causal SPY/VIX price, volatility, drawdown, volume, liquidity, momentum, and correlation features.
Active predictors	4-12	Features are active when z_j > 0.5.
Ridge penalty	10^-5 to 10^2	Decoded as lambda = 10^(-5+7u_lambda).
Classification threshold	0.10-0.85	Decoded as tau = 0.10+0.75u_tau.
Calibration temperature	0.50-3.00	Decoded as T = 0.50+2.50u_T.
Minimum model exposure	0.00-0.60	e_min = 0.60u_e; this can conflict with the platform cap.
Probability exponent	0.50-4.00	gamma = 0.50+3.50u_gamma.
Volatility target	8%-30% annualized	sigma_star = 0.08+0.22u_sigma; uses 20-day SPY volatility.
Smoothing	0.00-0.95	kappa = 0.95u_kappa.
Drawdown trigger	3%-20%	Based on 60-day SPY drawdown.
Initial model exposure	100%	Operational initialization before the first valid probability; the realized position remains subject to the 50% platform cap.
Maximum SPY position	50%	Platform equal-weight cap arising from the two resolved universe keys.
Maximum leverage	4.0	Platform constraint; the allocation policy itself is long-only and capped below 1x in SPY.
Rebalance band	0.5%	Default minimum exposure move before emitting a trade.
Optimizer budget	Population 24; 40 iterations; 984 evaluations/run	Thirty independent procedure runs and thirty equal-budget random-search runs.
Commissions	$0.0040 per share, minimum $1.00 per order, capped at 1.00% of trade value	Charged per fill by the results module; all metrics are net of them
Slippage	0 bps — not applied	Not an omission — MOC (market-on-close) fills at the auction print the backtest uses
Costs not modelled	short borrow fees / rebate, margin financing on leverage, market impact, exchange/regulatory/clearing pass-through fees, taxes	Excluded deliberately, not unknown
Execution	daily bars, MOC (market-on-close)	Exposure selected at date t applies to the next close-to-close return interval.
Bootstrap / alpha	10,000 / 0.05	Evaluation specification; expected calibration error uses 10 bins.

Look-ahead audit

#	Concern	Status
1	Predictor timing	✓ All SPY and VIX predictors are defined from observations dated no later than signal date t.
2	Return alignment	✓ Date-t exposure is applied only to the subsequent SPY close-to-close return, not the return used to form the signal.
3	Close execution	✓ Rebalances execute MOC at the official close used by the platform; VIX remains non-tradeable.
4	Training leakage	✓ Standardization is fitted on training data and frozen; extension tuning is prohibited.
5	Forward-label boundaries	⚠ The intended five-day labels are purged at boundaries, but the operational target fields are incomplete and the stated purge conflicts with the paper's reported effective sample count.
6	Applied costs	✓ commissions $0.0040/share (min $1.00/order); slippage 0 bps — not applied

Caveats / known limitations

• The decoded minimum SPY exposure can reach 60%, while the platform permits at most 50%; the 100% initial exposure is also infeasible. This makes part of the policy search space impossible and invalidates direct use of the paper's financial, turnover, and mean-exposure results.
• The operational target block omits the stated H=5, W=252, q=0.80 response definition, and the operational decoding of ridge penalty, threshold, and temperature is not fully specified. In addition, purging five labels before each of two internal boundaries would reduce 1,042 observations to 1,032, conflicting with the paper's reported 625/208/209 split; the fitted classifier and paper holdout samples therefore are not fully reproducible from the specification.
• SPY volume change and volume volatility use dollar volume rather than the paper's stated volume input. The corrected-significance claims also lack the required Friedman/Iman-Davenport, paired Wilcoxon, and Holm procedures, while the reported VIX ablation lacks its 22-feature re-optimization and 110-evaluation protocol; those feature, significance, and ablation findings do not describe this run directly.
• The exact 28-feature dictionary and probability-to-exposure equation are reconstructed, and unprinted quadratic penalty coefficients are excluded. Consequently, the paper's feature-selection frequencies, optimized predictive results, Pareto sets, scalar-loss rankings, policy returns, Sharpe ratios, drawdowns, turnover, and mean exposure are not transferable to this implementation.
• The implementation uses inverse-frequency class weights and training-only frozen standardization, neither of which is fully specified by the paper. Its fitted coefficients, selected ridge penalties and subsets, calibrated probabilities, thresholds, predictive results, and allocation results therefore are not directly comparable with the paper tables.
• The 2020-2024 extension includes market regimes absent from the paper sample, and the 50% SPY cap removes the paper's full 0%-100% implementation range. Paper holdout figures and buy-and-hold or fixed-80% comparisons cannot be treated as benchmarks for these results.
• No benchmark return series or predictive/calibration output accompanies this run, so the card cannot establish whether de-risking beat SPY, improved drawdown relative to fixed allocations, reproduced the reported ROC-AUC or balanced accuracy, or retained sensitivity across VIX regimes.
• Market impact, exchange/regulatory/clearing pass-through fees, and taxes are excluded. Borrow and leverage financing are economically immaterial to the stated long-only, cash-funded realized book.

Results

From 2020-01-01 through 2024-07-01, the backtest returned 22.53% with a 0.89 Sharpe ratio, 1.20 Sortino ratio, 5.63% volatility, and an -8.64% maximum drawdown. Beta versus SPY was 0.25; 511 trades produced an 80.92% win rate and 9.23 profit factor. These figures are net of the platform-applied commissions. Without matched benchmark, exposure, predictive, or calibration results, they show a profitable low-beta defensive path but do not establish superiority to SPY or reproduce the paper's model-level claims.

Run diagnostics

Recomputed here from the run's stored daily returns, because every annualized figure in the table above is derived from the LENGTH of that array (years = n / 252). Only total return and max drawdown do not depend on the row count, so when the recomputed and stored values differ by a common factor, those two are the numbers to trust.

Check	From the run's own returns	Note
Daily-return rows	1,131	1.00x the 1,134 trading days in 2020-01-01..2024-07-01
Distinct dates	1,131	one row per date
Date span	2020-01-02 .. 2024-07-01
Sum of daily returns	22.53%	matches the reported total return
Sharpe from these rows	0.89	stored 0.89
Volatility from these rows	5.63%	stored 5.63%
Max drawdown from these rows	-8.64%	stored -8.64%
CAGR from these rows	4.63%	stored 4.63%

Backtest metrics snapshot

Metric	Value
Total Return	22.53%
Sharpe	0.89
Sortino	1.20
Calmar	0.54
Max Drawdown	-8.64%
Volatility	5.63%
Beta vs SPY	0.25
Win Rate	80.92%
Profit Factor	9.23
Total Trades	511
Symbols	1 (SPY)