Risk Rules Don't Escape Post-Hoc Fitting Either — A 189-Month Test of Correlation and Volatility Caps

In Article 9 we performed an autopsy on how a return signal (the theme sleeve) was manufactured by hindsight selection. This time it is a risk rule. Rules that reduce losses draw less suspicion than return signals — which makes them more dangerous.

How the Rule Was Born

In July 2026, the US momentum strategy we were running posted -22.5% in a single month (Article 10). Right afterward, this rule was proposed: pick stocks in ranking order, but skip any stock whose 126-day daily-return correlation with an already-selected stock is 0.5 or higher, and skip any stock whose annualized volatility exceeds 100%. Reconstructing that July, -18.5% shrank to -8.0%.

The problem is obvious. The thresholds 0.5 and 100% were set by looking at July. Working well in July is not a test — it is a definition.

Test Method

• Using the production pipeline's ranking function as-is, monthly rankings were reconstructed for 189 months from 2010-12 to 2026-08 (current S&P 500 constituents, survivorship bias present — identical across all arms)
• Top 10 stocks equal-weighted with monthly rebalancing vs. 10 stocks after applying the caps
• 8 threshold variants (correlation 0.4/0.5/0.6, volatility on/off, 0.8/1.2) are all reported
• Key window: the period before the rule was created = a genuine out-of-sample

Results — The 158 Months Before the Tuning (2010-12 to 2024-01)

• It did not block the tail. Over 16 years, the worst month went -19.8 → -19.4 and MDD -29 → -28. The large drawdowns of 2011, 2015–16, 2018 and 2020 were beta crashes in which the whole market fell, and cutting cross-stock correlation leaves market exposure untouched. A correlation cap only blocks concentrated crashes like July, and there are only a handful in 16 years.
• Risk-adjusted, it is neutral. IR +0.09 sits inside the 1.36–1.48 spread across the 8 variants. Even in the in-sample window where the rule was created (2024-02 to 2026-06), IR went 1.60 → 1.31 — actually worse. That one July month was the whole story.
• Half the compounding. Continually replacing top-ranked stocks with lower-ranked ones dilutes momentum every month. It also missed the August rebound (cap -6% vs. no cap +2%).
• The volatility cap (100%) almost never triggered and was therefore inert. The entire effect came from the correlation cap.

What This Result Says

• Risk rules get post-hoc fitted too. The framing "it reduces losses" loosens the testing standard
• You must first ask which kind of drawdown a rule blocks. This rule is for concentrated crashes only, and most of our losses were beta crashes
• For the same purpose (tail defense), a rule that fires only in extreme regimes has a better cost structure than one that pays a cost at all times — we kept the already-deployed market-overheating guard and did not adopt this rule

Limits

• Because it is based on current constituents, the absolute figures are biased upward. Only cross-arm comparisons are valid
• For the 2010–2021 window, financial data is unavailable, so the quality feature is treated as neutral and the ranking is momentum-only (same convention as the production engine)