Persistent Correlation-Eigenspace Reconfiguration SPY-to-Cash Risk Overlay

Strategy Card

Persistent Correlation-Eigenspace Reconfiguration SPY-to-Cash Risk Overlay — strategy card

Spec ID: spec-correlation-geometry-volatility-regime-overlay-us-stocks-1787304647 · Generated: 2026-08-21 09:47 UTC

Cluster: Risk Overlay · Sub Cluster: Correlation Eigenspace Spy-To-Cash Volatility Control

One-line description

A monthly risk overlay measures persistent rotation in the subdominant correlation eigenspace of a point-in-time large-cap stock universe. When REC improves completed out-of-sample SPY variance forecasts, positive REC readings reduce SPY exposure in favor of cash.

Why this trade exists

Large changes in the cross-sectional correlation structure may reveal latent market reorganization before that risk is fully reflected in index-level realized volatility. The signal focuses on modes two through four rather than the dominant market mode, seeking information about changing sector and stock relationships that conventional SPY volatility measures can miss.

The economic claim is deliberately conditional: REC is used only after its one-month variance forecasts achieve lower historical QLIKE than a baseline containing VIX and lagged SPY realized variance. This gate recognizes the paper-consistent null that correlation geometry may add no useful information for SPY or may fail to improve portfolio performance.

Algorithm

[code omitted from public view]

Parameters

Param	Value	Notes
Signal universe	Point-in-time top 500 US stocks	Ranked by capitalization; re-screened using aiquant_screening_table
Tradable asset	SPY	Long-only overlay; VIX is a nontradable control
Defensive asset	Cash	Configured return is 0%
Sampling	Monthly signals from daily data	Monthly prices are last available adjusted closes
Correlation lookback	12 months	Stocks with any missing return in a window are dropped
Eigenspace	Modes 2-4; K = 3	Leading market mode is excluded
REC persistence	3-month trailing mean	Raw REC is the mean squared sine of three principal angles
Portfolio standardization	Expanding; 36-month burn-in	Portfolio z-score clipped to [-2, 2]
Forecast standardization	Expanding; no fixed burn-in	No clipping; uses observations available through each forecast origin
Forecast horizons	1-9 months	Activation gate uses the one-month horizon
Forecast gate	At least 12 completed pairs	Activate only when augmented-model mean QLIKE is below baseline QLIKE
Active SPY weight	clip(1 - 0.5 × max(z_REC, 0), 0, 1)	Negative REC z-scores do not increase exposure above 100%
Inactive SPY weight	100%	Residual allocation is cash
Exposure bounds	0%-100% SPY	Executable position cap is 100%; no overlay leverage
Framework leverage ceiling	4×	Nonbinding for the specified 0%-100% SPY overlay
Rebalance	First available SPY close after month-end	New weight applies only after execution
Explicit overlay cost	1 bp per one-way notional	Charged on absolute changes in SPY weight
Platform costs	$0.004/share; $1 minimum; 1% maximum commission	Slippage is configured at 0 bps
Backtest window	2020-01-01 to 2024-07-01	Daily source bars with monthly signal construction
Reported trades	55	Trade accounting is platform-defined

Look-ahead audit

#	Concern	Status
1	Point-in-time universe membership	Specified: each window uses the capitalization-ranked universe applicable at its endpoint rather than a final-date constituent list.
2	Signal and execution timing	Specified: the month-end signal uses data through that close and executes at the first available subsequent SPY close; returns before execution are excluded.
3	Forecast-gate outcome availability	Specified: QLIKE uses only forecast outcomes fully completed by the evaluation month-end.
4	Standardization leakage	Specified: both portfolio and forecast regressors use expanding statistics available through the forecast origin; forecast-null simulations receive the same transformation.
5	Missing-data and price imputation	Specified: no return or execution-price imputation is permitted, and incomplete stocks are dropped within each estimation window.
6	Fundamental release timing	Not applicable: the strategy uses prices, capitalization screens, VIX, and realized variance rather than earnings or filing data.

Caveats / known limitations

• The fidelity audit retains an internal position-cap contradiction: executable position_sizing.max_position_size_pct is 1.0 (100%), while constant provenance assigns the same parameter 0.1 (10%). The implemented fixed-SPY and inactive weights require 100%, so a 10% interpretation would materially change the strategy; the paper's portfolio and timing results do not describe either version of this SPY overlay.
• The point-in-time top-500 universe and capitalization-weighted SPY variance target differ from the paper's coverage-filtered constituent panel and twelve-month equal-weighted constituent-volatility setting. This invalidates direct transfer of its REC distributions, spectral-placement frequencies, principal-angle summaries, forecast coefficients, simulated-null significance, and seven-to-eight-month forecast peak.
• The QLIKE activation gate and long-only clip(1 - 0.5 max(z_REC,0),0,1) exposure rule are prospective adaptations rather than paper rules. They invalidate comparison with the paper's short-variance Sharpe ratios, equal-volatility timing tests, capture rates, payoff shares, loss frequencies, and worst-month results.
• Forecast regressors use recursive standardization without the paper's attribution burn-in or clipping, while the portfolio signal uses a 36-month burn-in and two-standard-deviation clipping. The paper's reported AR coefficients, Newey-West statistics, simulated-null critical values, p-values, and regression sample therefore do not transfer.
• SPY execution, zero-return cash, and explicit turnover costs replace the paper's nontradable variance-harvest payoff. Its premium decomposition, VRP coupling, variance-harvest attribution, and crash-protection findings are not performance expectations for this strategy.
• The 2020-2024 window is short relative to the 12-month signal lookback, three-month smoothing, 36-month portfolio burn-in, and 12 completed-forecast activation requirement. The effective active-overlay sample can therefore be much smaller than the headline backtest interval.
• This output does not report the spectral diagnostics, horizon-level forecast tests, QLIKE comparison, conditional REC in tail months, or fixed-SPY and volatility-target benchmark returns needed to test the paper-derived predicted properties. The run therefore cannot establish the stated spectral placement, forecast peak, no-timing-alpha, or no-crash-protection properties.
• The specification contains both a one-basis-point overlay turnover charge and platform commission settings. Interpretation of net performance depends on whether both cost layers were applied consistently, which is not established by the truncated code.

Results

The backtest produced a positive 68.15% total return with a 0.73 Sharpe ratio and 0.91 Sortino ratio, but volatility remained 20.70% and maximum drawdown reached -36.47%. Beta versus SPY was 0.91, indicating that the portfolio retained substantial equity-market exposure. The 94.12% win rate and 52.58 profit factor are based on only 55 platform-defined trades and should not be interpreted as independent monthly observations. Because no fixed-SPY, volatility-target, QLIKE, or transaction-cost-drag results are supplied here, the run does not establish incremental forecasting value or superior risk-adjusted performance.

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	68.15%	matches the reported total return
Sharpe from these rows	0.73	stored 0.73
Volatility from these rows	20.70%	stored 20.70%
Max drawdown from these rows	-36.47%	stored -36.47%
CAGR from these rows	12.28%	stored 12.28%

Backtest metrics snapshot

Metric	Value
Total Return	68.15%
Sharpe	0.73
Sortino	0.91
Calmar	0.34
Max Drawdown	-36.47%
Volatility	20.70%
Beta vs SPY	0.91
Win Rate	94.12%
Profit Factor	52.58
Total Trades	55
Symbols	500 (A, AAPL, ABBV, ABNB, ABT, ACGL, ACN, ADBE, ADI, ADM, +490 more)