What does an 80% band mean?

A price range, issued on a given day for a horizon of 7, 14 or 28 days, built so that the close on the target date should land inside it about 80% of the time. The band is a statement about the size of the move, not about which way price goes. This page grades every issued band against the close that actually happened and reports the realised rate next to the stated 80%.

Why are some cells UNDECIDABLE?

Because the honest sample size is the number of independent time windows, not the number of symbols. Hundreds of symbols graded in the same window move together, so they are one observation. A cell with fewer than 3 windows, or whose window-level interval is wider than ±0.05, cannot confirm or reject its stated coverage yet. We show it greyed rather than drop it; it gains a window every time another horizon-length period completes.

Do you forecast price direction?

No. Nothing on this site, and nothing on this page, says which way price will move. The bands are graded only on whether the terminal close stayed inside them. The one side-of-band fact we show is which edge a breach crossed (high or low), because that is a property of the band, not a prediction.

Accuracy · lab notebook

How often our forecast bands contain the close

Every band we publish is logged the day it is issued and graded on the close that actually happened. This page is the frozen monthly scoreboard of that record — misses, UNDECIDABLE cells and known data defects included — regenerated from the same table every month and never edited after the fact.

First frozen snapshot: Nov 1, 2026

The scoreboard, the headline row and every count below are written by the monthly snapshot job. Until it runs, this page shows the grading method, the verdict rule and the live breach list only.

1.What we publish, and what we don’t

For every active symbol in five markets, each UTC day, we publish three nested price bands — stated coverage 60% / 80% / 95% — at 7-day, 14-day and 28-day horizons. The band is sized from realised volatility (rv14) scaled by the horizon; the stated coverage is the promise: the close on the target date should land inside the N% band about N% of the time.

We do not forecast or grade price direction. A band says how far price is likely to travel, not which way. Nothing on this page scores which way price went; the only edge-of-band fact shown is which edge a breach crossed.

The numbers that will appear here come only from the frozen snapshot table. No figure is typed into the copy.

2.How a forecast is graded

Issue date → target date (issue + horizon calendar days) → terminal close (nearest daily close within ±3 days of the target; grading runs after the target day, so in practice the last close on or before it) → contained, or breached at the high edge, or breached at the low edge. A graded row is frozen and never re-graded. A day with fewer than seven daily returns issues no band rather than a fabricated one.

Worked example (a real recent row)

Issued Sep 2, 2026 for EA (Stocks), 28-day horizon. The 80% band was ±1.10% around the issue price.
Target date = issue date + 28 calendar days.
Terminal close used: Sep 30, 2026 (the nearest daily close within ±3 days of the target; in practice the last close on or before it).
Realised move -98.70% — outside the 80% band, crossing the low edge. The 95% band was also breached. That is the whole grade: inside or outside, and which edge.

3.The scoreboard — every market × horizon × band

Verdict rule — fixed before reading the numbers

The sample unit is an independent window: a Thursday issue batch from Jun 11, 2026 at stride = horizon, graded on its target date. Each window’s hit rate is one observation; the interval is a window-level t-interval (degrees of freedom = windows − 1). Then:

Fewer than 3 independent windows: UNDECIDABLE.
Window-interval upper bound below the stated coverage: UNDER-COVERS (the band is too narrow).
Window-interval lower bound above the stated coverage: OVER-COVERS (the band is too wide; conservative).
Interval contains the stated coverage and its half-width is at most ±0.05: HOLDS.
Interval contains the stated coverage but is wider than that: UNDECIDABLE. The data can neither confirm nor reject the stated coverage.

The scoreboard is populated by the first frozen snapshot (Nov 1, 2026).

4.Misses, explicitly

Trailing 30-day 80% breach rate

26.0%

129,626 graded rows realised since Sep 30, 2026 (live, not frozen; all markets pooled, so read it as a pulse, not a calibration).

The 20 largest 80%-band breaches of the trailing 30 days, ranked by how far past the edge the close landed (in half-widths).
Instrument	Issued	Horizon	80% half-width	Realised move	Edge	Overshoot	95% held
ForexUSD/IQD	Oct 5, 2026	7-day	±0.10%	+15.42%	high edge	149.56×	no
ForexUSD/IQD	Oct 2, 2026	7-day	±0.10%	+15.57%	high edge	148.20×	no
ForexUSD/IQD	Oct 4, 2026	7-day	±0.10%	+14.84%	high edge	144.89×	no
ForexUSD/IQD	Oct 1, 2026	14-day	±0.13%	+15.51%	high edge	122.52×	no
ForexUSD/IQD	Sep 30, 2026	14-day	±0.13%	+15.30%	high edge	120.93×	no
ForexUSD/IQD	Oct 3, 2026	14-day	±0.14%	+15.66%	high edge	109.68×	no
ForexUSD/IQD	Oct 5, 2026	14-day	±0.14%	+15.70%	high edge	109.68×	no
ForexUSD/IQD	Oct 2, 2026	14-day	±0.14%	+15.73%	high edge	107.78×	no
ForexUSD/IQD	Oct 4, 2026	14-day	±0.14%	+15.30%	high edge	107.58×	no
ForexUSD/IQD	Sep 25, 2026	14-day	±0.15%	+15.61%	high edge	105.56×	no
ForexUSD/IQD	Oct 6, 2026	7-day	±0.15%	+15.53%	high edge	102.45×	no
ForexUSD/IQD	Sep 29, 2026	14-day	±0.15%	+15.41%	high edge	101.51×	no
ForexUSD/IQD	Sep 28, 2026	14-day	±0.15%	+15.41%	high edge	100.93×	no
ForexUSD/IQD	Sep 27, 2026	14-day	±0.15%	+14.84%	high edge	96.95×	no
ForexUSD/IQD	Oct 3, 2026	7-day	±0.10%	+9.96%	high edge	96.51×	no
ForexUSD/IQD	Sep 30, 2026	28-day	±0.17%	+15.63%	high edge	88.89×	no
ForexUSD/IQD	Oct 1, 2026	28-day	±0.17%	+15.60%	high edge	88.71×	no
StocksEA	Sep 3, 2026	28-day	±1.10%	-98.72%	low edge	88.36×	no
StocksEA	Sep 4, 2026	28-day	±1.10%	-98.71%	low edge	88.35×	no
StocksEA	Sep 2, 2026	28-day	±1.10%	-98.70%	low edge	88.34×	no

Known data defects — shown with their counts, not filtered silently

July 2026 price-feed haltcount with first snapshot
Windows issued 2026-07-16 to 2026-07-26, or whose terminal close fell in the halt, are excluded because the issue price or the close was stale. Count = ladder windows dropped across the three horizons.
Forex band change (#640)count with first snapshot
Forex bands widened on 2026-07-24. Forex issued before that date is reported as its own historical series and never pooled with the band sold today. Count = historical-series cells shown separately.
Weekend and holiday copiescount with first snapshot
Stocks and indices issue every calendar day; weekend and holiday rows repeat the prior trading day's price, band and terminal close exactly. They are the same forecast counted twice and are removed. Count = rows removed from the counted windows.
Delisted before grading (survivorship)count with first snapshot
Matured bands whose symbol left the roster before its close could be fetched were never graded. Delisted names plausibly had outsized moves, so the realised rates may be flattered. Count = never-graded rows.
No close within tolerancecount with first snapshot
Matured bands still on the roster with no daily close within ±3 days of the target date. They stay ungraded rather than being graded against a stale print. Count = rows.

5.Independence of the sample

Hundreds of symbols graded in the same window are not hundreds of samples: in a given fortnight every stock rides the same market move. Daily issues at a 28-day horizon overlap each other about 28-fold too. So we count windows — issue batches spaced one horizon apart — and treat each as one observation. That makes the intervals wide, and the width is the truth.

A per-symbol interval that counts symbols as independent is not evidence and does not appear here. Pooling all markets into one headline is also not evidence: one market can sit under its stated coverage while another sits over it, and the pool lands on the stated figure by coincidence — which is why there is no fleet-wide row.

Horizon	Independent windows so far	Next window grades on
7-day	—	—
14-day	—	—
28-day	—	—

6.Changelog

2026-06-11#30 / #27Outcome logging begins. Every active symbol in five markets gets a 60/80/95 band at 7, 14 and 28 days each UTC day: price_at_issue × (1 ± rv14 × H^θ × m_band), form A, with a per-market θ exponent and a shared per-band multiplier. The logged edges are the edges the symbol pages draw.
2026-06-11#130rv14 is the sample standard deviation of the last 14 daily log-returns (close-to-close), in lockstep with the symbol-page envelope; fewer than 7 returns means no band is issued (rv_null), never a fabricated one.
2026-06-24T1Yang–Zhang 14-day volatility is logged as a SHADOW input for gap markets (stocks, indices, commodities). It never sizes the live band; its bands are reconstructed offline only.
2026-07-16OHLC haltPrice feed halt and freeze: issue prices frozen fleet-wide 2026-07-16 to 2026-07-26; crypto and forex targets 2026-07-16 to 2026-07-23 resolved 1–3 days early against pre-halt closes. Windows touching this range are excluded from every snapshot, and counted.
2026-07-24#640Forex 80% multiplier widened (m80 0.877 → 1.585). Forex issued before this date is a separate regime (“forex, before #640”) and is never pooled with the band sold today.
2026-09-27python-forecast #47Containment audit fixes the verdict method used on this page: non-overlapping Thursday windows from 2026-06-11 at stride = horizon, window-level t-interval, verdict rule fixed before reading the numbers.

Verdict changes between snapshots

None yet — a change is listed here the first time a cell’s verdict differs from the previous snapshot.

Reproducibility: outcomes table sn_rv_outcomes (written by cron/sn_rv_outcomes.php), snapshot table sn_accuracy_snapshots (written by cron/sn_accuracy_snapshot.php), method report analysis/containment_audit/containment_summary.md. HOLDS requires a window-interval half-width of at most ±0.05 and at least 3 windows.

7.Machine-readable

The same snapshot, as CSV. One row per market × horizon × band × snapshot date, every snapshot ever frozen; and the trailing breach list. Both are described in this page’s Dataset structured data.

Questions

What does an 80% band mean?: A price range, issued on a given day for a horizon of 7, 14 or 28 days, built so that the close on the target date should land inside it about 80% of the time. The band is a statement about the size of the move, not about which way price goes. This page grades every issued band against the close that actually happened and reports the realised rate next to the stated 80%.
Why are some cells UNDECIDABLE?: Because the honest sample size is the number of independent time windows, not the number of symbols. Hundreds of symbols graded in the same window move together, so they are one observation. A cell with fewer than 3 windows, or whose window-level interval is wider than ±0.05, cannot confirm or reject its stated coverage yet. We show it greyed rather than drop it; it gains a window every time another horizon-length period completes.
Do you forecast price direction?: No. Nothing on this site, and nothing on this page, says which way price will move. The bands are graded only on whether the terminal close stayed inside them. The one side-of-band fact we show is which edge a breach crossed (high or low), because that is a property of the band, not a prediction.