Operator brief · 174

Why is answered by attribution, or it is not answered.

The key idea

The problem with the second question

It is the only one of the two that can be answered badly and still sound answered.

Question one is protected by its own form. Location inside a percentile band is a fact with a number attached, and an operator either produces it or visibly does not. Question two has no such discipline built in. Asked why the difference exists, a review can return a fluent paragraph that identifies nothing, survives scrutiny because nobody can disprove it, and gets filed as an answer. The paragraph is usually true in the way weather reports are true. What it lacks is a contributor — a named branch, exit style, cost line, or deployment behaviour that accounts for a share of the gap and could be checked again next month.

The instruments

Four tables, four different decompositions of the same gap.

The attribution suite populates on a full-detail run and each table cuts the result along a different axis. Branch attribution decomposes contribution to return, drawdown, lock, and target hits — which branch earned its place, and which is carried. Variant attribution grades exit styles by usage, expectancy per trade, capture, tail contribution, and net value added, so a structural preference can be shown to pay or not. Failure attribution decomposes why paths died, sorting lock, stagnation, fee drag, low expectancy, overexposure, and correlation-cluster loss. Throttle efficiency measures conversion of deployed risk into return, with explicit penalties for both under- and over-deployment.

FigureWhich attribution table closes which version of the question

The gap looks like	Instrument	What it names
Return is short	Branch attribution	Which branch under-delivered against its weight
Winners feel smaller	Variant attribution	Which exit style traded expectancy for protection
Drawdown ran deep	Failure attribution	Lock, stagnation, drag, exposure, or clustering
Risk did not convert	Throttle efficiency	Under-deployment or over-deployment penalty

The second question is not one question. Each phrasing has an instrument that can answer it with a name and a share, and a phrasing with no instrument is a phrasing to be rewritten.

The precondition

Attribution requires a full-detail run, decided before the question was asked.

There is a practical trap here worth naming. The advanced attribution tables populate only at the highest output detail level; a standard run skips them and shows a note saying so. That means the ability to answer question two is committed to at configuration time, before the operator knows whether the month will produce a deviation worth explaining. A reviewer who reaches for branch attribution and finds a greyed placeholder has not hit a bug — they have hit a decision made a week earlier by someone optimising for run speed. The operational consequence is simple: benchmark runs intended to support live review are configured for full detail as a matter of routine.

The dependency between the questions

Attribution without location is analysis in search of a problem.

The tables are seductive enough to be run in the wrong order. Decompose a month with no established deviation and something will always look uneven — a branch under its weight, a variant with weak capture, a throttle reading slightly off model — because unevenness is what a finite sample looks like. Attribution has no built-in null hypothesis; it will faithfully apportion a gap that is not there. Question one exists partly to gate question two, which is why the order in the benchmark doctrine is fixed. Locate first. Only a placement that persistently sits outside the normal range has earned decomposition, and everything else is a pattern found by looking.

What a finished answer looks like

A contributor, a share, and something checkable next period.

The output of a properly run second question is not a narrative but a short set of claims with owners: this branch contributed disproportionately to drawdown relative to its weight, this exit style is showing negative net value at its current usage, fee drag accounts for a measurable slice of the shortfall. Each carries an implication for the next review, because each can be re-measured. That is the property adjectives lack. A month described as choppy generates nothing to check in thirty days; a month attributed to a specific branch's drawdown contribution generates a specific question, and a sequence of those questions is what distinguishes a diagnostic record from a diary.

The key idea

The second question is a demand for names, and the workbook can supply them.

Everything about the deviation taxonomy downstream — favourable alpha, variance, drag, excessive drawdown, governance failure — depends on knowing what actually differed, and the classification collapses into guesswork without it. The attribution tables exist precisely so that the second question has an answer of the same evidentiary quality as the first. Ask why and accept a summary, and the benchmark degrades into a monthly mood report with percentiles attached. Ask why and require a name, and the review produces something a future reviewer can argue with.

Connected inside MARS

Dynamic 7-Tier Benchmark →

Variant Attribution Module →

Monte Carlo Lab →

Every brief documents the same shipped system.

The complete MARS package — eleven workbooks, three TradingView indicators, the full manual library — $497.

Get MARS Pro — $497 Explore the Platform