← Dashboard 📈 Risk Indicators

HyperGraph Regime Gate

Structural Health Selection in Thematic Equity Networks: A Nonlinear Fiedler Eigenvalue Approach
128 US equity themes · 1,244 stocks · Jan 2021 – Mar 2026 · March 2026
1.245
C2 Sharpe
6.66%
Ann. Alpha
1.36
Info Ratio
4.28
Shuffle z-score
7.07
FM t-stat (60d)
$0.5-2B
Capacity
LIVE
Loading AI commentary…
🇺🇸 US Market
🇰🇷 KRX Market
One-Sentence Summary: Low-Fiedler themes (weakest internal stock synchronization) deliver Sharpe 1.25 because structurally healthy themes with distributed leadership avoid crowding-driven crashes — a tail effect confirmed by permutation tests (z = 4.28), stress-conditional analysis (+10.6 bps/day spread during stress), and theme-membership shuffles (z = 7.16).
📊 Empirical Evidence

Decile Sharpe Ratios (D1 = Lowest λ₂)

D1 consistently dominates — tail selection, not linear factor

Falsification Tests

Permutation z-scores — all significant at p<0.01

Tail-Dummy Fama-MacBeth t-statistics

Signal strengthens with horizon — structural, not noise

Transaction Cost Sensitivity

Signal survives institutional-grade execution (10-25 bps)

Rebalance Frequency

Weekly/monthly retain most signal — low-frequency structure

Liquidity Tier Distribution

225 US stocks — 64% large/mega cap

Conditional Spread: Stress vs Calm Periods

Entire D1-D10 advantage earned during market stress (+10.6 bps/day) — zero spread in calm (-0.1 bps/day)
📋 Key Results Summary
MetricValueInterpretation
C2 Sharpe1.245Bottom-20 themes by λ₂, equal weight
B&H Sharpe0.935All-theme equal weight benchmark
Alpha (ann.)6.66%After controlling for market beta
Market Beta0.956Near-full market exposure
Max Drawdown-22.6%vs B&H -23.8%
Turnover0.66 themes/day80% zero-change days
Fiedler Autocorr0.996Extremely persistent structural signal
Cross-sectional z4.28p<0.002, 500 permutations
Temporal z2.78p=0.006, date alignment matters
Membership z7.16Theme structure essential (30 trials)
FM t (20d)4.63Bottom-tail dummy, controlled
FM t (60d)7.07Significance grows with horizon
Mom. correlation0.005Orthogonal to momentum
Stress spread+10.6 bps/daySignal earns in stress periods
Calm spread-0.1 bps/dayZero spread in benign conditions
Capacity$500M-$2BWeekly rebal, 225 US large caps
Investor & Researcher Q&A
I. Signal Validity & Statistical Rigor
Q1 Your backtest is only 5 years. How can you trust a Sharpe of 1.25 on such a short sample?

Five years is the minimum we'd accept for a structural claim. We compensate with unusually extensive falsification:

  • Cross-sectional shuffle (500 trials): z = 4.28, p < 0.002
  • Temporal shuffle (500 trials): z = 2.78, p = 0.006
  • Membership shuffle (30 trials): z = 7.16, p = 0.000
  • Subperiod stability: C2 outperforms B&H in 7 of 10 half-year windows
  • Year-by-year: Positive Sharpe in every calendar year including 2022 (bear)

Out-of-sample validation on non-US markets (KRX, Japan) is the most important next step, and parallel HyperGraph systems already exist.

Q2 The Fama-MacBeth regression on rank(λ₂) produces t = -0.06. No signal?

No. The linear FM fails because the signal is genuinely nonlinear — a tail/cliff effect. With the correct tail-dummy FM specification:

Horizont-stat
5-day2.36
20-day4.63
60-day7.07

The t-statistic increases with horizon — hallmark of a structural, low-frequency signal. Many commercially successful screens (Altman Z-score, Piotroski F-score) share this property.

Q3 Decile monotonicity is weak (Spearman 0.12-0.36). Not a proper factor?

A proper linear factor, yes. But this is a tail selection effect. D1 consistently dominates, but D7-D9 also perform well. We are not claiming λ₂ is a linear pricing factor — it is a powerful threshold screen.

Q4 Could this be survivorship bias? Theme taxonomy applied retroactively.

All 128 themes have 100% data coverage from Jan 2021 onward. The F3 membership-shuffle test (z = 7.16) shows the specific grouping structure matters — random stock groupings produce Sharpe 0.88 ± 0.05, while real C2 at 1.245 is 7+ sigma above.

Q5 Multiple hypothesis corrections? 67 rescue experiments.

The 67 experiments were exploratory. The final paper tests a single, pre-specified rule (C2: bottom-20, equal weight). The cross-sectional shuffle is the definitive guard: 0 out of 500 random permutations achieved Sharpe 1.245. Empirical p < 0.002, no adjustment needed.

Q6 Is Fiedler eigenvalue just a proxy for volatility or correlation?

No. Fiedler-momentum rank correlation: 0.005. Fiedler-volatility rank correlation: -0.114. D1 themes have 20.4% vol — identical to D10's 21.4%. Lambda-2 captures an orthogonal structural dimension.

II. Economic Mechanism
Q7 What's the causal story? Why should low synchronization predict higher returns?

Proposed mechanism:

  1. Theme identity creates structured co-movement (F3: z = 7.16)
  2. λ₂ measures synchronization intensity
  3. Low sync = distributed leadership = less crowding
  4. Less crowding = more persistent returns

Key evidence: The entire low-vs-high λ₂ spread is earned during market stress (+10.6 bps/day), with zero spread in calm (-0.1 bps/day) — precisely the crowding/fragility prediction.

Q8 D10 doesn't have worse skew or drawdown. Contradicts crowding story?

Of six predictions from the synchronization-fragility hypothesis, 4/6 confirmed:

PredictionResult
D10 worse CVaRPASS (-2.89% vs -2.78%)
D1-D10 spread positive skewPASS (+0.854)
Rising λ₂ predicts worse returnsPASS (Sharpe 0.609)
High λ₂ crashes harder in stressPASS (+10.6 bps/day)
D10 worse skewFAIL
D10 deeper MaxDDFAIL

Recommended framing: structural health screen with crash-avoidance properties.

Q9 Is this just momentum in disguise?

Definitively not. Rank correlation with 20-day momentum: 0.005 (effectively zero). In the double-sort, the best cell is low-λ₂ / middle-momentum, not low-λ₂ / high-momentum.

Q10 Korean-defined themes working for US stocks?

The themes map to globally recognizable narratives: "Nuclear Energy," "Robotics," "Secondary Batteries," etc. The mapping includes 1,244 US equities covering all major large caps. The F3 test confirms groupings capture real structural information. Korean origin is irrelevant to the correlation structure.

III. Implementation & Capacity
Q11 What's the realistic fund capacity?

225 unique US stocks, 64% large/mega cap. Total daily volume: $116B.

RebalanceConservativeModerate
Daily$118M$500M
Weekly$589M$2B

Binding constraint: 2 micro-liquid stocks. Excluding them raises capacity materially. At weekly rebalance, realistic capacity is $500M–$2B.

Q12 Daily rebalancing impractical. Weekly/monthly loss?
FrequencySharpevs Daily
Daily1.245
Weekly1.115-0.130
Biweekly1.109-0.136
Monthly1.146-0.099

Monthly slightly outperforms biweekly — consistent with quarterly signal timescale. Cost difference between daily and weekly is minimal (80% zero-change days).

Q13 Transaction costs — does the signal survive?

At institutional-grade execution (10-25 bps), the signal remains robust. Breaks down at 100 bps (unrealistic for this universe).

Q14 How many positions? Operationally manageable?

~225 positions. On typical day, 0-1 themes rotate (~0-15 stock trades). 80% of days: zero changes. Mean streak: 29.8 days. Operationally trivial.

Q15 Long-only beta ~0.96. Bear market performance?

In persistent bear markets it loses money. However: positive for full 2022, best during regime transitions (2022 H2 Sharpe 1.458), drawdown recovery 2.8x faster (206d vs 581d).

IV. Robustness & Sensitivity
Q16 Persistence filters or inverse-λ₂ weighting improvements?

Both fail. Persistence filter: k=5 Sharpe 1.196. Inverse weighting: 1/λ₂ Sharpe 1.154, MaxDD worsens. Smoothed rank: monotonically degrades. The signal is already maximally persistent (autocorr 0.996) and binary. Simplicity is the correct implementation.

Q17 Selection depth k sensitivity?
kSharpeReturnMaxDD
101.38628.4%-21.7%
151.25525.3%-23.1%
201.24524.5%-22.6%
301.04920.5%-22.0%
400.98719.2%-22.3%

Monotonically decreasing — consistent with tail effect. k=20 balances signal strength and diversification.

Q18 Different graph constructions?

All adjacency variants produce Sharpe > 1.0. Positive-correlation-only baseline performs best (1.113 on quarterly grid). Directionally stable.

Q19 Correlation window sensitivity (60 days)?

Within-window shuffles have z = 0.48 — exact daily composition irrelevant. Quarterly-scale structural state is what matters. Robust to moderate window changes (40-90d).

Q20 Unnormalized vs normalized Laplacian?

Not tested. Unnormalized captures both synchronization strength and density. The normalized would remove density effect. Open empirical question.

V. Literature & Comparison
Q21 Relation to factor crowding literature?

Factor crowding measures crowding at the stock/factor level. We measure at the theme level. Our contribution is the measurement tool (Fiedler eigenvalue), not a new theory. The stress-conditional result is directly consistent with crowding premium frameworks.

Q22 Random matrix theory connection?

Conceptual rather than methodological. RMT uses full eigenvalue spectrum on full market matrix. We use a single eigenvalue (λ₂) on within-theme subgraphs.

Q23 IR of 1.36 seems very high. Comparable numbers?

High IR partly reflects low tracking error (4.88%). Most factors show 0.3-0.8 in-sample. We'd be comfortable with 50% decay live (IR ~0.7) — still commercially attractive.

VI. Practical Deployment
Q24 What data is needed for live implementation?

Only two inputs: (1) Static theme-to-ticker mapping (~128 themes), updated quarterly at most. (2) Daily close prices for ~1,244 stocks. Computation takes ~2 minutes. No external signals, no macro data, no NLP, no alt data.

Q25 Latency requirement? T+1 pricing OK?

Lag-0 vs lag-1 difference is negligible (1.251 vs 1.245). Even 20-day-old data produces Sharpe 1.143. Zero latency requirement. End-of-day pricing sufficient.

Q26 Product structure options?
  1. Thematic Selection Overlay: Alpha signal for thematic ETF allocators
  2. Enhanced Thematic ETF: Equal-weight 225 stocks, rebalanced weekly
  3. Systematic Thematic Fund: λ₂ as primary allocation signal
Q27 Live monitoring metrics?
MetricThresholdAction
Fiedler autocorrelation< 0.95Signal structure changing
Daily turnover> 3 themes/dayCheck data/structural break
Rolling 60d spreadNegative > 90dRegime review
Jaccard similarity< 0.85Signal unstable
Q28 Worst-case scenario?

Worst realized: 2022 H1 (Sharpe -0.873). Worst structural: market microstructure changes where crowding evolves faster than the 60-day window detects.

VII. Extensions & Future Work
Q29 Can you short the high-λ₂ themes?

Promising (spread Sharpe 0.609 for level signal, +10.6 bps/day stress advantage). But not backtested as long-short. High priority next step.

Q30 Does this work in other markets?

Parallel HyperGraph systems exist for KRX and Japan. Preliminary KRX results show similar structural health effect. Formal cross-market validation is the single most important extension.

Q31-34 Alternative taxonomies, ML, macro conditioning, λ₂ change signal?

Taxonomy: Requires 50+ groups with 5+ stocks each. GICS sectors (11) too few. GICS sub-industries (~160) might work.

ML: Evidence suggests no improvement. Composite scoring degrades to 0.961. Signal is a clean threshold.

Macro: C2 works best in high-return periods (Sharpe 3.306). Conditioning could help drawdowns but risks overfitting.

Change signal: Falling-λ₂ outperforms rising (spread Sharpe 0.609). Conceptually distinct from level-based C2.

VIII. Honest Limitations
Q35 What are you most worried about?
  1. 5-year sample: Unusual conditions (COVID, meme stocks, AI boom, rate hikes)
  2. Retroactive taxonomy: Can't test historically removed themes
  3. Single taxonomy: Dependent on Naver grouping quality
  4. No OOS test: No walk-forward or live paper trading yet
  5. Nonlinear = harder to hedge: Can't replicate with factor exposures
Q36 Why not just buy diversified stocks directly?

F3 membership-shuffle: random groupings produce Sharpe 0.88 ± 0.05. Real C2 at 1.245 is 7.16 standard deviations above. The theme structure is essential — it organizes correlations into semantically coherent subgraphs.

Q37 If widely known, arbitraged away?

Capacity: $500M-$2B. Quarterly timescale and low turnover resist HF arbitrage. Larger risk is signal degradation if many allocators use the same rule. Can absorb $1-2B before capacity binds.