Point-in-Time Triadic Network-Stress Defensive Overlay

Strategy Card

Point-in-Time Triadic Network-Stress Defensive Overlay — strategy card

Spec ID: spec-triadic-network-stress-defensive-overlay-us-large-cap-etfs-1786536770 · Generated: 2026-08-12 12:57 UTC

Cluster: Risk Overlay · Sub Cluster: Network-Stress Defensive Allocation Overlay

One-line description

A point-in-time correlation-network indicator shifts a large-cap equity portfolio into a 75% MBB defensive allocation when systemic co-movement reaches an expanding-history extreme. Hysteresis and asymmetric signal memory reduce rapid state reversals.

Why this trade exists

Diversification can fail abruptly when normally distinct assets begin moving together. The Triadic Stress Index combines weighted clustering, network density, degree heterogeneity, and the inverse Laplacian spectral gap to identify unusually interconnected market states in which equity drawdown risk may be elevated.

The asymmetric memory filter responds quickly to rising stress but decays slowly, while separate 90th-percentile entry and 75th-percentile exit thresholds create a hysteresis band. The economic hypothesis is that temporarily replacing most equity exposure with Treasury and agency mortgage-backed bonds during these states can reduce portfolio risk. This allocation benefit is an extension of the source research rather than a result established by that paper.

Algorithm

[code omitted from public view]

Parameters

Param	Value	Notes
Network window	20 trading days	Daily adjusted close-to-close log returns; no imputation
Network size	42 assets	Held constant when replacement candidates are available
Recompute frequency	Every 3 trading days	Signal state is carried forward between calculations
Adjacency	A = abs(rho), diagonal 0	Pearson correlations
Clustering	Tr(A^3)/(n(n-1)(n-2))	Complete-graph normalization specified by this implementation
Density	Mean off-diagonal adjacency	sum(A_ij)/(n(n-1)), for i != j
Degree coexistence	Population variance of weighted degrees	Coex = Var(k)
Spectral term	gap = lambda_2(diag(k)-A)	Invalid/nonpositive gaps do not change state
Raw TSI	(C*D/M)*Coex, M=1/gap	Equivalent to C*D*gap*Coex under the stated definitions
Memory coefficients	Rise 0.60; decline 0.08	Faster response to increasing stress and slower decay
Risk-off entry	Expanding percentile >= 90%	Uses only point-in-time recomputation history
Risk-on exit	Expanding percentile < 75%	Creates a 15-point hysteresis band
Execution	Close, one-trading-day lag	Signal uses closing data and the state change is scheduled for the following trading close
Risk-on allocation	100% equities / 0% MBB	Equal-weight eligible screened assets
Risk-off allocation	25% equities / 75% MBB	MBB is a Treasury and agency MBS ETF
Position limits	10% per equity; 75% MBB	Defensive asset is exempt from the general equity cap
Universe	PIT US stocks and ETFs, capitalization-ranked	Yearly selection; ADRs excluded; positive dollar volume required
Backtest window	2016-01-01 to 2026-08-11	Daily bars; warm-up history begins in 2000 when available
Primary turnover cost model	10 bps per one-way turnover	Sensitivity cases: 0, 5, 10, and 25 bps
Platform trading costs	$0.004/share; $1 minimum; 1% maximum commission	Slippage configured at 0 bps
Annualization / risk-free rate	252 / 0%	Used for reported risk ratios
Maximum leverage	4.0	Framework limit; target allocations themselves sum to 100%

Look-ahead audit

#	Concern	Status
1	Universe membership and ranking	Pass by specification: yearly screening is point-in-time and uses the contemporaneous capitalization and liquidity fields.
2	Return-window construction	Pass: each signal uses only observed adjusted closes through its calculation date; missing returns are not imputed.
3	Expanding percentile	Pass: thresholds use only filtered TSI observations available through the current recomputation, not a full-sample percentile.
4	Close-derived signal execution	Pass: the strategy measures the signal at the close and executes the state change at a later close after its explicit one-trading-day lag.
5	Missing execution prices	Pass by policy: the affected rebalance is skipped rather than filled with an interpolated or carried-forward price.
6	Annual universe transitions and data survivorship	Requires data verification: the PIT flag is specified, but database snapshots, delisted-name coverage, and historical classifications must be independently confirmed.

Caveats / known limitations

• The open paper-fidelity finding states that the Absorption Ratio diagnostic omits the same asymmetric filter required for a like-for-like comparison. Mixing filtered TSI 0.447 with raw Absorption Ratio 0.174 yields 0.273, whereas the filtered point estimates 0.447 and 0.134 imply 0.313; consequently, the paper's filtered benchmark comparison does not describe this diagnostic implementation.
• The audit records two unresolved, undescribed deviations. One invalidates direct use of the paper's raw/filtered TSI magnitudes, F1 results, unexplained-alarm rates, and effective-rank/Absorption-Ratio comparisons; the other invalidates its crypto and other cross-market findings, including Terra/Luna and MATIC-USD attribution, as descriptions of this strategy.
• The traded expanding-percentile hysteresis gate differs from the paper's complete-series top-decile alarm test, and the yearly screened US stock/ETF network differs from its named market panels. The paper's F1, precision, recall, exact alarm-budget, banking, AI, OFR, sovereign-debt, sector, crypto, FX, commodity, and 42-asset robustness results therefore do not transfer to the traded episodes.
• The source does not provide the interior values of its 36-combination alpha grid or the dates belonging to its crisis-label sets. This implementation cannot reproduce the reported alpha-grid stability, label-dependent F1 statistics, bootstrap intervals, p-values, or Holm-Bonferroni results.
• The clustering denominator, weighted-density convention, combinatorial Laplacian, and population degree variance are fixed implementation choices where the supplied source is incomplete. These choices prevent direct comparison of TSI levels, percentiles, attribution results, crisis z-scores, correlations, and alarm overlap with the paper unless independently verified against its code.
• The 25% equity/75% MBB allocation, 75th-percentile exit, execution lag, turnover convention, and cost assumptions are portfolio extensions rather than paper-tested rules. MBB also carries duration, mortgage-spread, and convexity risk and is not equivalent to cash or a pure Treasury holding.
• The run reports portfolio performance but not TSI alarm F1, stress-index lead/lag correlations, diagnostic overlap, defensive-state duration, turnover, downside capture, or an always-invested benchmark result. It therefore cannot test the paper's predicted signal-detection properties or determine whether the overlay improved return, drawdown, or volatility versus its specified benchmark.
• The universe mixes individual companies with broad and style ETFs such as SPY, IVV, VOO, QQQ, and VUG. Their overlapping holdings can mechanically raise correlations, duplicate exposure, and make both the network statistic and the nominally equal-weight equity portfolio less diversified than the asset count suggests.

Results

From 2016-01-01 through 2026-08-11, the backtest returned 229.67% with a 0.87 Sharpe, 1.06 Sortino, and 0.26 Calmar. Realized volatility was 25.00%, maximum drawdown reached -46.66%, and beta versus SPY was 0.67, indicating reduced market sensitivity but substantial residual tail risk. The run recorded 1,304 trades, a 73.04% win rate, and a 4.88 profit factor. Because no always-invested benchmark or gross-versus-net decomposition is supplied, these figures do not establish that the defensive overlay improved performance relative to continuous equity exposure.

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	2,666	1.00x the 2,674 trading days in 2016-01-01..2026-08-11
Distinct dates	2,666	one row per date
Date span	2016-01-04 .. 2026-08-11
Sum of daily returns	229.67%	matches the reported total return
Sharpe from these rows	0.87	stored 0.87
Volatility from these rows	25.00%	stored 25.00%
Max drawdown from these rows	-46.66%	stored -46.66%
CAGR from these rows	11.94%	stored 11.94%

Backtest metrics snapshot

Metric	Value
Total Return	229.67%
Sharpe	0.87
Sortino	1.06
Calmar	0.26
Max Drawdown	-46.66%
Volatility	25.00%
Beta vs SPY	0.67
Win Rate	73.04%
Profit Factor	4.88
Total Trades	1304
Symbols	42 (AAPL, ABBV, AMD, AMZN, AVGO, BAC, BND, BRK-B, CAT, COST, +32 more)