Sigma Deviation as Normalized Signal

σ Deviation as Normalized Signal

Raw numbers lie about magnitude. A move of 100 means one thing in NVDA and something completely different in gold. Standard deviations turn raw metrics into a unitless, comparable scale that lets the framework treat signals from different tickers, asset classes, and time horizons in the same language.

6.1.1Why Normalize at All

The single most pervasive analytical mistake in trading is comparing raw numbers across instruments without normalization. A ten-point move in SPY (currently $723.71) is roughly 1.4%. A ten-point move in NVDA is closer to 5%. A ten-point move in gold is about 0.4%. The same headline number maps to wildly different magnitudes of significance depending on the instrument’s natural scale and natural volatility. Sizing decisions made on raw numbers will be systematically wrong — oversized in low-vol names, undersized in high-vol names, and impossible to compare across.

Standard deviations solve this. A move of +1σ means the same thing in every instrument: one historical standard deviation away from the mean of the relevant baseline. The unit is not dollars, percent, or shares; it is “how unusual is this move relative to its own recent history.” That makes signals across instruments commensurable, which is the precondition for any cross-sectional ranking, any composite score, and any sizing framework that wants to treat opportunities as belonging to the same risk space.

The intuition map is direct. Roughly 68% of moves fall within +/-1σ of the mean, 95% within +/-2σ, and 99.7% within +/-3σ under a normal distribution. A +2σ move is in the 97.5th percentile of moves; a +3σ move is in the 99.85th percentile. When the framework flags a +2.4σ flow event, it is saying “this print is in the top 1% of similar prints over the relevant window.” That is information-dense in a way the raw print rarely is.

Raw numbers are vocabulary. Sigma is grammar. The grammar is what lets you compare anything to anything across the framework.

6.1.2Computing σ on Rolling Baselines

The standard deviation calculation requires a baseline (the mean) and a dispersion estimate (the standard deviation around the mean). Both should be computed on rolling windows, not on all-time history, because the right baseline for “is today unusual?” is the recent past, not the inception of the instrument. The AZTMM default is a 90-day rolling window for most signals, with shorter windows (30-day) for vol-sensitive signals and longer windows (250-day) for regime signals.

The mechanic is direct. For each day, compute the mean of the prior 90 days of the metric, compute the standard deviation of those same 90 days, and report today’s value as (today – mean) / stdev. The output is the σ deviation. A σ reading of +1.22 means today is 1.22 standard deviations above the 90-day mean. A reading of -2.1 means today is 2.1 standard deviations below the mean — a meaningfully negative outlier.

The rolling window adapts naturally to regime change. When realized volatility rises, the standard deviation in the denominator rises, and the σ reading on a given raw move falls — correctly noting that what looked like a +2σ move in a quiet regime is only a +1σ move in a volatile regime. When realized volatility falls, the same raw move becomes a higher σ reading. The framework adapts to the environment without requiring manual recalibration.

6.1.3Reading σ on Flow Data

The clearest application is options flow. Today’s call/put dollar ratio of 2.19× sounds high, but the question is whether it is high relative to the recent baseline. The 90-day mean of call/put ratios is 1.42×, with a standard deviation of 0.63. Today’s reading is therefore (2.19 – 1.42) / 0.63 = +1.22σ. That is meaningful but not extraordinary — in the top quintile of recent days but not in the top decile.

The interpretation depends on the regime. In a Bull regime, a +1.22σ call/put reading is consistent with the regime — bullish positioning is what the regime predicts — and is corroborating rather than informational. In a Crisis regime, the same +1.22σ call/put reading would be a meaningful divergence from the regime, suggesting either contrarian positioning or a leading indicator of regime transition. The same number reads differently because the baseline expectation has shifted.

The same logic applies to dark pool prints, gamma exposure, breadth indicators, and any other metric the framework tracks. The σ reading is not the answer; it is the input that makes the metric comparable across days, tickers, and regimes.

6.1.4Comparing Across Tickers and Asset Classes

The cross-sectional ranking use of σ is the workhorse application. The Daily Pulse and Weekly Pulse both rank tickers by signal strength using σ as the unifying metric. NVDA flow at +3.75σ ranks above SPY flow at +1.96σ, regardless of the dollar amounts. The reader does not have to remember the natural notional level of each ticker; the σ column does that work.

The same is true across asset classes. A +2.1σ move in 10Y yields, a +1.4σ move in DXY, and a +2.6σ move in HYG/LQD can all be displayed in the same multi-asset table and compared at a glance. Without normalization, the table would be unreadable; the natural units of yields (basis points), DXY (index points), and HYG/LQD (ratio decimals) are not commensurable. With σ normalization, the table becomes a single coherent ranking.

The cross-sectional discipline extends to portfolio construction. When ranking the day’s setups, the σ readings on each input (flow, dark pool, technicals, regime) can be summed or weighted into a composite score that is comparable across instruments. A composite of +5σ aggregated across four inputs on TSLA is a more compelling setup than a +3σ aggregate on AMD, regardless of the underlying tickers’ natural volatilities.

6.1.5Failure Modes & Fat Tails

The σ framework assumes a roughly normal distribution of the underlying metric. Markets are not normally distributed. They have fat tails — extreme events occur far more frequently than the normal distribution would predict. A 6σ event in equities under a normal distribution should occur roughly once every several billion trading days; in actual market history, 6σ events occur multiple times per decade. The model is wrong about the tail.

The implication is that σ is reliable for the body of the distribution (everything within +/-2.5σ) and unreliable for the extremes (beyond +/-3σ). For tail events, the better measure is percentile — “what fraction of the prior 250 days were as extreme as this?” For body events, σ is more useful because it preserves continuity and allows arithmetic. The discipline is to know which regime you are in: if a metric reads +4σ or worse, switch to percentile reasoning before sizing or interpreting.

Other failure modes are more subtle. σ assumes the variance is stable over the rolling window; if the variance itself is shifting (the regime is changing volatility), the σ reading lags and either understates the move (when vol is rising) or overstates it (when vol is falling). The fix is to overlay the σ reading with a separate vol-of-vol check: if realized vol over the last 10 days is meaningfully different from the 90-day baseline, treat all σ readings on the metric with caution.

6.1.6Common Mistakes

• Comparing raw numbers across tickers. A 10-point move in NVDA is not commensurable with a 10-point move in SPY.
• Reading σ without checking the baseline window. A 30-day σ and a 250-day σ on the same metric tell different stories.
• Trusting fat-tail σ readings. Beyond +/-3.5σ, switch to percentile.
• Ignoring vol-of-vol shifts. A σ reading is unreliable when the variance itself is changing.
• Treating the same σ the same across regimes. +1.22σ bullish flow reads differently in Bull vs Crisis regime.
• Aggregating σ readings without correlation adjustment. Summing four correlated +1σ signals is not a +4σ composite.