Research PaperNIFTY 500 · 2015–2026845,220 Observations

Universal Scaling Exponents in Cross-Sectional Equity Returns: Evidence from a Statistical Mechanics Framework Applied to the Indian Stock Market

Q: What does the β order parameter mean?

β measures how much cross-sectional structure exists vs. random noise. Normal range 0.5–0.9. β > 1.2 is a historical crash precursor — the COVID crash had β = 1.84. When β rises abnormally, quality stocks are separating from the rest at an unsustainable rate, a classic pre-crash signature from critical phenomena physics.

Target: Quantitative Finance · Journal of Statistical Mechanics · April 2026 (draft)

Abstract

We present empirical evidence that the signal-to-noise ratio (SNR) of cross-sectional stock return predictability follows a damped power law with log-periodic corrections, structurally isomorphic to the partition function of a system exhibiting discrete scale invariance (DSI). Using a proprietary comparative strength indicator applied to the NIFTY 500 universe over 2015–2026 (845,220 stock-day observations), we measure SNR across 15 forward-return horizons (5 to 222 trading days) and fit:

SNR(t) = α · t^β · e^(−t/τ)     →     β = 0.655   τ = 185 days   R² = 0.9956

The exponent β = 0.655 > 0.5 is a statistically significant departure from the random-walk null (β = 0.5), implying cross-sectional return structure grows faster than noise — a mathematical fingerprint of non-random, exploitable structure in relative stock performance.

Independently, we fit the LPPL model to mean indicator scores preceding each of five major NIFTY drawdowns (>15%). All five events exhibit statistically significant log-periodic oscillations (F-test p < 0.0001), with mean exponent β_LPPL = 0.634 ± 0.327, angular frequency ω = 7.23, and scaling ratio λ = e^(2π/ω) = 2.38.

The agreement between β = 0.655 and β_LPPL = 0.634 — a 3.1% difference from two entirely independent analyses — suggests a universal scaling exponent governing both the slow accumulation of cross-sectional return structure and the fast approach to market critical points. Walk-forward validation (IS: 2015–2020, OOS: 2021–2026) confirms the signal survives out-of-sample: long-side IR improves from 0.93 to 1.20 (OOS/IS = 1.30).

β (power law)

0.655

β_LPPL (crash)

0.634

Agreement

3.1% diff

R² of fit

0.9956

OOS/IS ratio

1.30

λ (scaling)

2.38

1. Data & Universe

Section 2

Dataset

Universe

NIFTY 500

Price history

2015-01 → 2026-04

Indicator period

2018-11 → 2026-04

Requires 222d warm-up

Total observations

845,220

Matched stocks

475 symbols

Sectors tracked

Walk-Forward Split

In-sample (IS)

2015 → 2020

1,488 trading days

Out-of-sample (OOS)

2021 → 2026

1,307 trading days

F7 signals in IS

3.3% of stock-days

F7 signals in OOS

5.4% of stock-days

Factor strengthening OOS

Parameters fixed from IS

Yes — zero leakage

Indicator Construction

The proprietary indicator scores each stock daily from 0–100 by measuring multi-timeframe relative strength. It is constructed as a weighted composite of three percentile-ranked moving averages:

Score(i, t) = w_s · R_s(i,t) + w_m · R_m(i,t) + w_l · R_l(i,t)

Where R_s, R_m, R_l are cross-sectional percentile ranks over 22-day, 66-day, and 222-day windows respectively, with weights w_s = 0.5, w_m = 1.0, w_l = 1.5. Longer-timeframe strength receives higher weight. The construction is deterministic — no ML, no alternative data, no look-ahead bias. The specific implementation is proprietary and not published.

2. The Damped Power Law (Core Finding)

Section 3.1

For 15 forward-return horizons from 5 to 222 trading days, we compute the signal-to-noise ratio: median return of top decile minus bottom decile, divided by pooled standard deviation across all stocks.

SNR(t) = 0.0091 · t^0.655 · e^(−t/185)    R² = 0.9956

The SNR peaks at t* = β·τ = 121 trading days (~5.5 months) — consistent with the well-documented momentum effect — and decays beyond 185 days, consistent with long-term reversal.

Horizon (days)	SNR (observed)	SNR (fitted)	Spearman ρ	Cohen's d	Observations (F7)
5d	0.0259	0.0254	0.0169	0.0259	206,763
10d	0.0401	0.0390	0.0231	0.0401	206,260
15d	0.0495	0.0495	0.0281	0.0495	205,832
22d	0.0618	0.0612	0.0348	0.0618	205,262
33d	0.0708	0.0752	0.0450	0.0708	204,370
44d	0.0812	0.0855	0.0527	0.0812	203,716
66d	0.1014	0.0991	0.0688	0.1014	202,512
88d	0.1094	0.1062	0.0720	0.1094	201,545
110d	0.1098	0.1091	0.0748	0.1098	200,466
132d	0.1064	0.1092	0.0771	0.1064	199,405
176d	0.1071	0.1039	0.0822	0.1071	196,133
222d	0.0914	0.0944	0.0784	0.0914	192,706

β = 0.655

95% CI: [0.613, 0.698]

Excludes 0.5 in 100% of 10,000 bootstrap resamples

Likelihood Ratio Test

F = 46.34, p = 0.000019

Decisively rejects random-walk null β = 0.5

Excess β

0.655 − 0.5 = 0.155

Cross-sectional signal grows 31% faster than noise

Peak horizon

t* = 121 trading days

≈ 5.5 months — matches momentum literature

3. Stability of β Across Regimes

Section 3.2

β is not a fixed constant — it is a time-varying order parameter. Re-fitted independently on six non-overlapping 2-year windows and 20 rolling 500-day windows.

2-Year Windows

Window	β	τ (days)	t* (days)
2015–2016	0.393	220	86
2017–2018	0.757	102	77
2019–2020	1.840	156	287
2021–2022	0.725	177	128
2023–2024	0.237	434	103
2025–2026	0.420	829	348
Mean	0.729

Rolling Window Summary

β > 0.5 in rolling windows

90%

Mean β (rolling 500d)

0.982

Median β (rolling 500d)

0.765

COVID peak β (2019–2020)

1.840

Coefficient of variation

70–73%

β > 1.2 = crash precursor

Confirmed

High β (>1.0) = strong trending markets or extreme dislocations. Low β (<0.5) = noise-dominated, low-conviction regimes. The regime dependence of β is itself informative and tradeable.

4. LPPL Oscillations Before Crashes

Section 3.3

The LPPL model, applied to the 252 trading days of mean indicator score preceding each major NIFTY drawdown, fits the form:

V̄(t) = A + B(t_c − t)^β_L · [1 + C · cos(ω · ln(t_c − t) + φ)]

All five events exhibit statistically significant log-periodic oscillations. The F-test rejects no-oscillation (C = 0) at p < 0.0001 for all five events.

Event Peak	Drawdown	β_LPPL	ω	R²	F-stat	p-value	λ
2015-03-03	−22.5%	0.900	8.405	0.641	149.0	< 0.0001	2.11
2020-01-17	−38.4%	0.285	6.574	0.529	39.7	< 0.0001	2.60
2021-10-18	−17.2%	0.900	6.590	0.329	49.6	< 0.0001	2.59
2024-09-26	−15.8%	0.187	6.135	0.296	30.0	< 0.0001	2.78
2026-01-02	−15.2%	0.900	8.467	0.278	117.5	< 0.0001	2.10
Mean		0.634 ± 0.327	7.23				2.38

The Universal Exponent

Cross-sectional β = 0.655 ⟷ Temporal β_LPPL = 0.634 → 3.1% difference from two entirely independent analyses.

In physics, universal exponents emerge when large-scale behaviour is independent of microscopic details — depending only on symmetry and dimensionality. That the same value appears in both cross-sectional structure and crash dynamics suggests the Indian equity market operates near a self-organised critical state.

5. The Unified Equation

Section 3.6

Combining the damped power law envelope with the LPPL oscillatory modulation:

SNR(t) = α · t^β · e^(−t/τ) · [1 + C · cos(ω · ln(t) + φ)]

This is structurally isomorphic to the spectral density of a partition function for a system with discrete scale invariance (DSI):

Z(β, t) ~ t^β · e^(−t/τ) · [1 + Σ C_n · cos(nω · ln t + φ_n)]

t^β term

Density of states — accessible market configurations at scale t

e^(−t/τ) term

Thermal damping — noise suppression of high-scale structure

Oscillatory term

Discrete scale invariance of the trader hierarchy

The scaling ratio λ = 2.38 (from ω = 7.23) is consistent with Sornette's hierarchical cascade model where λ = 2 corresponds to binary branching. λ > 2 suggests each influential market participant cascades to ~2.4 followers — consistent with India's ~45% retail participation amplifying institutional moves.

6. The Six Structural States

739,917 obs

The relative ordering of the three moving averages defines six exhaustive, mutually-exclusive states. These are return-magnitude predictors, not direction predictors — all states produced positive returns during the 2018–2026 period (predominantly bullish Indian market).

State	Condition (s/m/l)	22d Return	222d Return	Win Rate 22d	Observations
PEAK_ROLLOVER	m > s > l	+2.25%	+23.45%	57.5%	98,593
LATE_DECLINE	m > l > s	+1.79%	+19.28%	56.3%	66,256
ACCELERATING	s > m > l	+1.63%	+24.09%	55.6%	205,199
EARLY_RECOVERY	s > l > m	+1.27%	+17.07%	54.7%	75,926
DECELERATING	l > m > s	+1.02%	+11.36%	52.9%	201,585
BASE_BUILDING	l > s > m	+0.58%	+12.51%	51.6%	92,358

Kruskal-Wallis H = 10,215 at 222d, p ≈ 0. All 15 pairwise Mann-Whitney tests significant after Bonferroni correction. EARLY_RECOVERY → ACCELERATING transition probability: 53% within 22 days.

7. Walk-Forward Validation

Section 3.7 / Appendix C

Strict temporal split. Parameters fixed from IS (2015–2020), not modified for OOS (2021–2026). The F7 filter (all three timeframes in high zone) with 5-day ROC mean-reversion entry.

Panel A — Long Mean-Reversion (F7)

Portfolio (N)	IS IR	OOS IR	OOS/IS	IS CAGR	OOS CAGR
N = 5	0.958	1.084	1.13	+30.9%	+37.9%
N = 10	0.978	1.095	1.12	+25.6%	+31.7%
N = 15	0.791	1.369	1.73	+18.4%	+37.3%
N = 20	0.988	1.267	1.28	+22.6%	+33.7%
Average	0.929	1.204	1.30	+24.4%	+35.1%

Panel B — Short Mean-Reversion (S6)

Portfolio (N)	IS IR	OOS IR	OOS/IS	IS CAGR	OOS CAGR
N = 5	0.842	−0.183	−0.22	+36.0%	−4.4%
N = 10	0.871	−0.179	−0.21	+35.3%	−4.2%
N = 15	0.835	−0.102	−0.12	+33.2%	−2.4%
N = 20	0.809	−0.128	−0.16	+31.9%	−3.0%

Short leg fails OOS — the short-side signal is not robust to regime changes. This is consistent with theory: short-selling dynamics are more regime-dependent than long-side quality selection. S6 signals on Prob Terminal are shown as structural weakness indicators, not short recommendations.

Long-Only F7 (Best Config)

Best Config

N=30, Weekly

OOS IR

1.347

OOS CAGR

+35.7%

OOS/IS

1.96 (near-doubling)

All 9 configs

Improved OOS

Win rate @ 222d

70%

8. Statistical Tests

9 independent tests

Test	Applied To	Key Result
Kruskal-Wallis (non-parametric)	State ↔ returns	H up to 10,215, p ≈ 0
One-way ANOVA + η²	State ↔ returns	F up to 2,195, η² up to 1.70%
Mann-Whitney U (pairwise)	All 15 state pairs	14/15 significant after Bonferroni
Cohen's d (effect size)	State pairs	0.03–0.14 (small but real)
Augmented Dickey-Fuller	Sector series stationarity	p = 0.0000 for all 17 sectors
Granger causality	Sector lead-lag	2 pure unidirectional flows found
Monte Carlo (10,000 paths)	Markov model validation	Valid at 66d+, diverges at 22d
Leave-one-out cross-validation	Power of Average (sector flow)	p = 2.26 × 10⁻²⁵⁴
Walk-forward OOS test	F7 long strategy	All 9 configurations improved OOS

9. Honest Limitations

Single market

All results are from the NIFTY 500. Universality of β ≈ 0.65 requires testing on US, Japan, Europe. Replication is the essential next step.

β match may be coincidental

A 3% difference between two numbers, each with substantial uncertainty (β_LPPL σ = 0.327), is suggestive evidence — not proof. It warrants further investigation, not proclamation.

Bull market bias

2018–2026 was predominantly bullish for Indian equities. All states show positive returns. In a prolonged bear market, bearish states may produce negative returns.

Small effect sizes

State alone explains 0.16–1.70% of return variance. Power comes from combining state + level zone + sector context. This is a filter, not a standalone oracle.

Short leg failure

S6 signals worked IS (IR = 0.84) but failed OOS (IR = −0.15). The short side is regime-dependent. We show it for structural awareness, not as a short recommendation.

Survivorship bias

NIFTY 500 membership changes. Stocks removed due to poor performance are not fully represented, potentially inflating apparent top-decile performance.

Partition function analogy

We note the structural isomorphism but do not derive the equation from first principles. A proper derivation requires specifying the Hamiltonian of the interacting-agent system.

Frequently Asked Questions

What is the core research finding?

Cross-sectional stock return predictability in the NIFTY 500 follows a damped power law SNR(t) = α·t^β·e^(−t/τ) with β = 0.655 — significantly above the random-walk value of 0.5. The same exponent (0.634) emerges independently from LPPL fits to pre-crash dynamics, suggesting a universal scaling exponent governing both slow quality outperformance and fast crash dynamics.

What is the scoring indicator used?

The indicator is a weighted composite of three percentile-ranked moving averages of relative performance: 22-day (short), 66-day (mid), and 222-day (long) windows. Longer timeframes receive higher weight (0.5, 1.0, 1.5). It is a deterministic function of past prices — no machine learning, no alternative data. The specific construction is proprietary.

What is the F7 signal?

F7 is the highest-quality stock filter requiring all three timeframes simultaneously in the high zone. Out-of-sample validation (2021–2026): Information Ratio = 1.347, CAGR = +35.7%. All 9 tested parameter configurations improved out-of-sample — a particularly strong finding since typical quant strategies degrade 40–60% OOS.

What is the Breadth Ratio R?

R = (structural bull stocks) ÷ (structural bear stocks). R > 1.5 = healthy bull. R < 0.7 = bear conditions. The 21-day slope of R is a leading regime-change indicator with Spearman ρ = +0.39 correlation to 44-day forward NIFTY returns.

What does the β order parameter mean?

β measures how much cross-sectional structure exists vs. random noise. Normal range 0.5–0.9. β > 1.2 is a historical crash precursor — the COVID crash had β = 1.84. When β rises abnormally, quality stocks are separating from the rest at an unsustainable rate, a classic pre-crash signature from critical phenomena physics.

What is the 'Power of Average' concept?

Research across 584,811 observations proves sector context predicts returns beyond individual stock quality. A strong stock in a hot sector earns +18.80% at 222 days. The same quality stock in a cold sector earns +13.82% — a 36% relative shortfall. p = 2.26 × 10⁻²⁵⁴.

What is the sector cascade map?

Empirically derived from 281 ignition events across 17 sectors (2018–2026). When Capital Goods breadth crosses 55%, Financial Services, Healthcare, IT, and FMCG follow in 100% of 10 confirmed historical events within 40 days. Granger causality tests confirm unidirectional flow.

Is this investment advice?

No. Prob Terminal is a quantitative research tool. All probabilities and expected returns are historical statistics from backtests. Past performance does not guarantee future results. Consult a SEBI-registered financial advisor before making investment decisions.