Rolling EV: the tab that catches decay while the average still smiles.

The inertia problem

A long average is an anchor — that's its virtue and its blindness.

Cumulative EV earns its authority from sample size: hundreds of trades, noise averaged away, the most statistically secure number in the workbook. But the same mass that makes it secure makes it slow. A system whose true edge stepped down three months ago still shows a healthy all-time EV, because three months of mediocrity is a small minority of the total sample — the anchor drags the average toward history. The troubleshooting table names the resulting signature as its own diagnostic: rolling EV deteriorates while all-time EV is strong means recent regime shift or edge decay, and routes the investigation to exactly this tab. The cumulative number isn't wrong. It's answering 'what has this edge been?' when the operating question is 'what is it now?'

The instrument

Weekly EV, rolling windows, sigma, drift, and z — a regime tracker, not a scoreboard.

The Rolling EV Engine converts the selected lens into a week-by-week view with several readings layered on it: selected weekly EV (raw, noisy, honest), the rolling window average (noise reduced, trend exposed), rolling sigma (how reliable the weekly numbers are), and drift and z-readings that express where the current rolling state sits relative to the long-run baseline. Together they classify the lens into a regime state — improving, stable, deteriorating, diverging — which is a categorically different output than a score. A score invites comparison against hope. A regime state invites exactly one question: has the state changed, and has it persisted?

Reading it honestly

Rolling numbers are fast because they're small — treat them accordingly.

The rolling view's sensitivity is bought with sample size, which means it inherits every hazard the statistical battery exists to manage: a rolling window is a permanently small sample, its weekly points are individually meaningless, and its dips are usually variance. The discipline is the same two-sided read the deviation doctrine prescribes everywhere. A rolling dip inside normal sigma with a stable all-time baseline is weather — noted, not acted on. A rolling trend that persists across windows, pushes the z-reading out, and begins dragging the marginal and EWMA reads with it is the early signature of drift — and the correct routing is diagnostic, into Conditional Context to isolate where the decay lives, not straight to conclusions about the edge. The rolling engine raises questions early. It answers none of them alone.

• One weak rolling window is a data point. A directional sequence of windows is a signal. The distinction is everything.
• Rolling deterioration localized to one session or ATR state is a context finding, not an edge obituary — slice before concluding.
• The rolling read is the second pillar of the edge-change threshold: necessary evidence, never sufficient alone.

The pairing doctrine

Cumulative and rolling are a two-instrument system.

The two views are designed to be read together because each covers the other's blindness. Cumulative EV without rolling is an archive — deeply true about the past, silent about the present. Rolling EV without cumulative is a mood ring — exquisitely current, statistically flimsy. Read as a pair they produce the four states that matter: both healthy (operate normally), both weak (the edge question is live and the battery has probably already said so), cumulative strong with rolling weak (the decay signature — investigate context and mechanics now, while the damage is weeks old instead of quarters), and cumulative weak with rolling strong (recovery or a young improvement — encouraging, unproven, and no basis for scaling yet). Every state has a next action, and none of them is panic.

The key idea

The rolling engine buys time, which is the scarcest diagnostic resource.

Edge decay discovered at the cumulative level is discovered late — after quarters of degraded deployment, with the benchmark comparison already flashing and capital already spent learning the lesson. The rolling engine's entire contribution is moving that discovery earlier: weeks instead of quarters, while the response can still be a context repair or a mechanical fix instead of a system overhaul. It pays for its noisiness many times over, provided the operator honors the deal — early questions in exchange for patient answers.