Live vs Benchmark · Classifying Deviation

Alpha, variance, drag, or failure.

Classes

Alpha conditions

Named after

PERSISTENCE

How MARS uses this

Classifying a deviation as alpha, variance, drag, or failure starts with the drawdown axis. Returns above the model with drawdown outside these bands is over-risk wearing an alpha costume; returns below the model with drawdown well inside them is often drag, not decay. The distribution is the classifier's first witness.

How it benefits you

You get a pre-agreed definition of normal pain, decided calmly before the pain started. That single artifact prevents the two classic errors - panic de-risking inside ordinary variance, and denial when damage is genuinely abnormal. You always know which situation you are in.

The classifier's first check: where live drawdown sits against the simulated distribution decides which deviation story is even plausible.

The waiting rule

A deviation earns its label only by surviving the next window.

Single-review gaps stay unlabelled by rule. A deviation receives its classification after it persists through a further evidence window, because most gaps close on their own and premature labels trigger premature interventions. The queue of items being watched but not yet named is a designed feature of the process rather than evidence of indecision, and its length is itself worth watching: a queue that never empties usually means the review window is too short for the metric being judged.

Drawdown decides first

The risk axis is the classifier's first witness.

Where live drawdown sits against its own distribution determines which stories are even plausible before any other reading is considered. Returns above the model with drawdown outside the bands is over-risk wearing an alpha costume, and no amount of supporting metric strength converts it. Returns below the model with drawdown well inside the bands is more often drag than decay, which points the review at friction and capture rather than at the edge itself.

The five conditions

Alpha is an intersection, not a threshold.

True outperformance under this standard requires benchmark outperformance, controlled drawdown, positive expectancy, efficient risk conversion, clean profit quality, and sustainable acceleration. The definition is demanding on purpose. A looser definition would classify a fortunate quarter as edge, and the operator would then scale a system whose apparent advantage disappears the moment the sequence stops cooperating.

Any one missing condition changes the class, not the confidence.
The classification is recorded in the review, so the sequence of classes becomes readable.
Sustained alpha, and only sustained alpha, opens the conversation about expansion; gate authority still rules it.

Model mismatch

Sometimes the gap is in the ruler.

One of the closing verdicts is model mismatch, and it exists because the benchmark can be wrong about the account rather than the account being wrong about itself. Live branch mix drifting from the modelled weights, a changed trade rhythm, or a different cost environment all produce gaps that no execution repair will close. The QA checklist requires confirming the live system is still operating under the benchmark's assumptions before any comparison is trusted, precisely so that this verdict is reached deliberately rather than as a last resort.

Reference

The classifier: what a gap reads as, given its company

Live pattern	Reads as	Because
Above model, drawdown inside bands, quality metrics confirming	Alpha	All five conditions present at once
Above model, drawdown worse than adverse bands	Over-risk	The return was purchased outside the modelled risk envelope
Above model, heavy override use	Governance leakage	The result was produced outside the authority the model assumes
Above model, weak expectancy	Variance	The edge did not produce the return, so the return will not persist
Above model, poor risk conversion	Wasted risk	Deployment is inefficient even where outcomes are good
Below model, drawdown normal, expectancy intact	Normal underperformance	Inside variance; check execution, fees, and branch mix
Below model, drawdown elevated	Weak and risky	Diagnostic review, and reduced aggression if the stack supports caution
Below model, expectancy decaying across windows	Structural failure	The gap is the edge, not the sequence

Doctrine

Connected inside MARS

This module doesn't work alone.

Connected — 01

Live vs Benchmark (full page)

Open system page →

Connected — 02

The Comparison Stack

Open system page →

Go deeper

Operator briefs on this territory.

Deep dive — 01

Noise, drift, or break: classifying the gap before touching anything.

Variance heals, drift trends, breaks step-change. Misclassify the gap and every response is wrong.

Read the full brief →

Deep dive — 02

Five conditions, all of them at once, or it is called something else.

Beat the model on four of five and you have not produced alpha. You have produced a diagnosis.

Read the full brief →

Deep dive — 03

Deviations wait for their names, and the waiting is the process.

Most gaps close on their own. A label applied before that happens creates a response the evidence never asked for.

Read the full brief →

Every module ships in the complete MARS package.

One price. Eleven workbooks, three TradingView indicators, and the full manual library — $497.

Get MARS Pro — $497 Explore the Platform