FreeOpen sourcePine Script v6
Did price fit the model available beforehand, or did the model move to fit price? Separate the current regression, the preceding model and scheduled fixed-origin paths to study slope, refitting and subsequent observations.
An ordinary least-squares center and residual envelope are combined with update attribution and scheduled fixed-model research. The study separates removing an old observation from adding a new one, and center movement from width changes. Scheduled origins are not selected retrospectively by outcomes.
Solid cyan/pink boundaries enclose the current center marked OLS. ±1w means one selected half-width, not necessarily one standard deviation. Dashed EXT is a geometric continuation excluded from statistics and alerts. Amber marks the preceding model; thin F-numbered channels retain their origin parameters.
Fits N consecutive Source values against equally spaced bar positions in price units, not log prices or elapsed calendar time. A residual is Source minus fitted value at the same window position. Default N is 120 and half-width is twice the square root of residual SSE divided by 118.
Standard error uses multiplier × sqrt(SSE/(N−2)); mean absolute residual uses multiplier × mean absolute residual; percentile uses a nearest-rank absolute residual. The percentile ignores the multiplier. Mean absolute residual is not median absolute deviation. Tick and numerical floors can dominate nearly exact-line data.
Evaluate the previous model, common N−1-observation model and new model at the same current coordinate. Drop = common minus previous; Add = new minus common. Their sum is total center movement. The stipulated order removes the oldest value before adding the newest; displayed units are prior half-widths, not percentages.
Margin equals half-width minus absolute residual. Center contribution is old absolute residual minus new absolute residual; width contribution is new width minus old width. Divide both by prior width and their sum equals normalized margin change. This is geometric attribution, not economic causality.
The same confirmed observations are measured before/after refitting. Coverage change equals (included count minus excluded count) divided by paired count, or 100 times this for percentage points. After includes each observation in its own fit and is not an out-of-sample forecast.
Extend the earlier model by one or H bars while retaining origin width. Skill = 100 × (1 − model squared-error sum / unchanged-origin-Source squared-error sum). Positive is smaller error than the matched baseline; negative is worse and zero baseline error is unavailable. It is not a return or win rate.
Keep origin center, slope and width unchanged while evaluating the next L confirmed Source observations. Default L24 plus gap24 schedules origins 48 bars apart. The origin itself is not an outcome, and an exit neither resets the model nor deletes the path.
Use standard time-based candles and a price-valued source. Begin with close, length 120, twice the residual standard error and Confirmed close updates. Read Model first, then compare the observation against the model available beforehand.
Open the public Linear Regression Channel Fit Audit page and add it to TradingView. No SG Group registration is required.
Overlapping lines, bands, backgrounds or tables can be duplicate drawings of the same kind of information. Never treat visual overlap by itself as stronger evidence or signal agreement. Current fit, preceding-model reference, successive one-step traces, F-numbered paths and EXT have different roles. Other channels can obscure those roles or make related lines look like independent evidence.
Move through four panel pages while keeping fitting samples, endpoint tests and completed-path samples distinct. Tooltips provide further context.
Slope is fitted change over 20 bars divided by current half-width. R² measures in-window fit; constant data are unavailable and the five-part meter is coarse. In/Next uses different samples. Prior/Now use different denominators, so subtracting them is not a raw-price change. A full-sample check mark is not a performance pass.
Residual persistence is the adjacent-residual product sum divided by SSE, not a significance test. Older/recent half-window slopes share the full-window width; odd N assigns the extra observation to the older half.
CENTER means center movement alone suffices; WIDTH means width alone; JOINT needs both; EITHER permits either alone. These compare new-center/old-width and old-center/new-width geometries. Their counts sum to inclusion count, without proving a price reversal.
Compare one-step and H-step results with their own identical-sample baselines. The two horizons may have different eligible counts. H-step output includes absolute error, baseline error, error in origin-width units and upper/lower misses. Overlapping endpoint tests are dependent.
Frozen-path End count equals All plus Back over the same valid completed paths. An exit at age6 followed by an inside endpoint at age24 is BACK, not ALL. With close as Source, containment tests closes rather than every intrabar wick.
F/A compares origin fitted-center extrapolation with origin Source plus the same slope, against a common unchanged-Source baseline. Only the starting level changes. Scores pool squared errors, not path percentages. Roll/Fix compares updating-only and fixed-only containment on matched observations, but updating has intermediate information and possibly different widths.
Set the interval and check the actual selected close. Later candles remain visible, so research is not blind replay. Unfinished paths at cutoff remain open. Later timestamps may place annotations, but later Source prices do not enter the selected model or statistics.
Numeric history can differ from the visible panel. Record M/R/H/F and page ID. Cause codes: 0 none, 1 center, 2 width, 3 joint, 4 either. Event bits: 1 outside, 2 inside again, 4 refit inclusion, 8 exclusion, 16 coverage gap, 32 split disagreement, 64 path start, 128 completion, 256 first exit. Coexisting bits add; glyph filtering does not remove them.
Fitted center / selected half-width boundaries / geometric extension excluded from research.
Outside transition / refit inclusion / inside-again transition. Placement is not a trade price.
Center alone / width alone / both required / either sufficient.
Fixed-path ID. All observations inside / final inside after an exit / final outside, classified at maturity.
Missing-data path / unfinished progress k of L, not a successful result.
Endpoint/all-observation containment; back-inside/any-exit; valid completions/open age.
Loaded-history sensitivity of the scheduled origin phase.
Research interval exceeds its budget; remaining partial records are not full results.
Reversed interval / no eligible close / insufficient warm-up or data.
The labels below match the published script’s English Inputs panel. Defaults reflect the public source at the time of review and may change in a later TradingView update.
On small screens, each input is reformatted as a readable stacked card.
| Input label | Default / options | What it controls and when to adjust it |
|---|---|---|
View | Latest / options: Latest, As-of research | Latest follows the current snapshot. As-of research reconstructs the last admitted confirmed close and disables alerts; use standard time-based candles. |
Research start | 2026-01-01 00:00 UTC | Research starts at 2026-01-01 00:00 UTC by default. Training may precede the interval, but test origins must be admitted; one-step and H-step outcomes need their earlier in-interval models. |
As-of cutoff | 2026-09-01 00:00 UTC | Default cutoff is 2026-09-01 00:00 UTC. Only closes at or before it qualify; check the selected close in exchange time. It does not crop or scroll the chart. |
Research statistics | Rolling audit / options: Rolling audit, Selected interval | Rolling audit uses the recent K bars; Selected interval aggregates eligible tests throughout the interval. Actual valid counts matter, and frozen paths have separate slot retention. |
Second frozen horizon, bars | 5 / min 2, max 50 | Second endpoint horizon H, default five. Extends the model from H bars earlier with its original width; overlapping tests are dependent and do not test whole-path containment. |
| Input label | Default / options | What it controls and when to adjust it |
|---|---|---|
Source | close / options: open, high, low, close, hl2, hlc3, hlcc4, ohlc4 | Price-valued input: close by default; open/high/low, midpoint and OHLC averages are alternatives. External sources can repaint; missing observations invalidate a fit rather than being skipped. |
Regression length | 120 / min 12, max 500 | Number of consecutive observations per regression, default 120. This is separate from the historical workload budget. |
Snapshot updates | Confirmed close / options: Confirmed close, Live preview | Confirmed close holds the closed-bar snapshot; Live preview allows intrabar refitting of that snapshot. Research totals, path observations, events and alerts still use confirmed closes. |
Width method | Residual standard error / options: Residual standard error, Mean absolute residual, Residual percentile | Choose residual standard error, mean absolute residual, or a nearest-rank absolute-residual percentile. None constructs a confidence or prediction interval. |
Residual multiplier | 2 / min 0.1, max 8 | Multiplier for standard error or mean absolute residual only, default two. Percentile width ignores it; it does not imply a normal-distribution coverage guarantee. |
Absolute-residual percentile | 90 / min 50, max 99.5 | Nearest-rank percentile of absolute fitted residuals, default 90. It concerns the fitting window, not a future 90% coverage target; the multiplier is unused. |
Minimum half-width, symbol ticks | 2 / min 0.1, max 100 | Minimum half-width in symbol ticks, default two. Together with a numerical floor it prevents zero division and may dominate flat or nearly exact-line samples. |
| Input label | Default / options | What it controls and when to adjust it |
|---|---|---|
Rolling audit window, bars | 100 / min 20, max 500 | Recent audit window K, default 100 bars. Counts only valid confirmed tests within those bars; overlapping fits are not independent trials. |
Slope measurement span, bars | 20 / min 1, max 250 | Span for expressing fitted price change, default 20 bars, normalized by current half-width. It is neither an annualized return nor a screen angle. |
Flat slope threshold, half-widths | 0.05 / min 0, max 3 | Flatness threshold in half-width units, default 0.05. Width changes can alter normalized slope and its classification. |
Coverage-gap highlight, percentage points | 10 / min 1, max 80 | Highlight threshold for in-fit coverage exceeding next-bar coverage, default 10 percentage points, after a full valid sample. These are different populations, not a significance test. |
Residual persistence highlight | 0.6 / min 0.1, max 0.99 | Threshold for the absolute lag-one residual-product ratio, default 0.6. Describes persistence without a significance test. |
| Input label | Default / options | What it controls and when to adjust it |
|---|---|---|
Current fitted channel | On | Shows the current fitted center and boundaries; the historical fitted segment redraws as the window moves. |
Inner guides | On | Shows dotted inner guides. |
Width fraction | 0.5 / min 0.1, max 0.9 | Inner-guide distance as a fraction of the selected half-width, default 0.5. |
Channel shading | On | Enables channel shading. |
Transparency | 91 / min 75, max 99 | Shading transparency, default 91; higher values are fainter. |
Dashed geometric extension | On | Shows the dashed geometric continuation, not a forecast range. |
Bars | 12 / min 1, max 100 | Geometric extension length, default 12 bars. The EXT segment does not enter research statistics or alerts. |
Previous-model reference | On | Shows the amber model reference computed before the displayed observation was included. |
Visible bars | 18 / min 2, max 120 | Visible span of the preceding-model reference, default 18 bars. Its endpoint at the displayed observation is the one-step reference. |
Older / recent half-window fits | Off | Shows separate regressions on the older and recent halves, with one extra older observation for odd lengths. These diagnose shape, not separate trading systems. |
Latest one-step residual connector | On | Connects the latest observation to its one-step reference center. |
Compact endpoint tags | On | Shows compact endpoint labels such as OLS and ±1w. |
Tag offset, bars | 2 / min 1, max 20 | Horizontal label offset in bars, default two. |
Historical frozen one-step center | Off | Connects successive one-step reference centers. Each point has its own preceding model; this is not one permanently fixed channel. |
Historical frozen one-step boundaries | Off | Shows successive one-step boundary traces, distinct from a scheduled fixed-origin path. |
Latest reference display, bars | 240 / min 50, max 4000 | Reference-trace display span in Latest, default 240 bars. Research uses the admitted interval instead. |
Center line thickness | 3 / min 1, max 5 | Center-line thickness, default three. |
Boundary thickness | 2 / min 1, max 5 | Boundary thickness, default two. |
| Input label | Default / options | What it controls and when to adjust it |
|---|---|---|
Show information panel | On | Shows the information panel. |
Panel page | Model / options: Model, Refit research, Horizon research, Frozen paths | Select Model, Refit research, Horizon research or Frozen paths. All pages share the same eight-row panel geometry and font size. |
Detail | Compact / options: Compact, Detailed | Compact keeps the concise readout; Detailed adds diagnostic values. |
Text size | 12 / min 9, max 18 | Panel text size, default 12, range 9–18. |
Bold values | On | Makes values bold. |
Bold row headings | On | Makes row headings bold. |
Panel symbols and meters | On | Shows supplementary icons and fit meter. Disabling them does not change position, row count or font size. |
Right clearance, chart percent | 3 / min 0, max 12 | Right clearance as chart percentage, default 3%. |
Top clearance, chart percent | 2 / min 0, max 12 | Top clearance as chart percentage, default 2%. |
| Input label | Default / options | What it controls and when to adjust it |
|---|---|---|
Upper / rising | #24E4F2 | Upper boundary/rising color, cyan by default; not a buy instruction. |
Lower / falling | #FF55AB | Lower boundary/falling color, pink by default; not a sell instruction. |
Prior reference | #FFD166 | Previous-model reference color, amber by default. |
Split fits / flat | #AF96FF | Split-fit/flat color, lavender by default. |
Panel background | rgba(16,26,44,0.95) | Information-panel background color. |
Panel text | #F0F6FF | Primary panel text color. |
Secondary text | #C5D4E8 | Secondary text color. |
| Input label | Default / options | What it controls and when to adjust it |
|---|---|---|
Calculation budget, bars | 600 / min 100, max 4000 | Historical budget, default 600 bars. Latest adds warm-up; research treats it as a hard admitted-range limit. Exceeding it blocks the main research display rather than shortening the regression. |
Confirmed descriptive event markers | On | Shows confirmed descriptive event glyphs, not trade signals. |
Retained event markers | 60 / min 0, max 240 | Event-glyph retention cap, default 60; zero hides drawings. |
Event marker size | Tiny / options: Tiny, Small, Normal | Text-only glyph size: Tiny 10px, Small 12px, Normal 14px. No filled symbol background. |
Event display window, bars | 240 / min 20, max 4000 | Recent event display window, default 240, bounded by the calculation budget. It does not change fit length, statistics or alerts. |
Research: no marker spacing filter | Off | Disables research glyph spacing only; count and window caps remain. Numeric events are retained independently of drawing filters. |
Minimum same-side marker spacing, bars | 4 / min 0, max 100 | Display-only spacing applied separately above and below candles, default four bars. Hidden events still affect statistics and configured alerts. |
Marker clearance, ATR units | 0.45 / min 0.1, max 3 | Clearance beyond current and previous two candle wicks in ATR(14), default 0.45, with a four-tick minimum. ATR is placement-only, not a fixed pixel guarantee. |
Event symbols instead of letters | On | Uses !, ↺ and =; off selects O, R and I. Glyph placement is not an execution price. |
| Input label | Default / options | What it controls and when to adjust it |
|---|---|---|
Show historical frozen paths | On | Shows chronological fixed-origin paths through reached confirmed observations. Disabling drawings leaves the path study active. |
Frozen path length, bars | 24 / min 4, max 120 | Path outcome length L, default 24, excluding the origin. Completion requires all L observations; this is separate from endpoint research. |
Gap between paths, bars | 24 / min 0, max 240 | Gap between outcomes, default 24, gives 48-bar origin spacing. With zero gap completion precedes the next start on one close. Training windows can still overlap. |
Path schedule anchor | 2020-01-01 00:00 UTC | Default schedule anchor is 2020-01-01 00:00 UTC. The first loaded qualifying close seeds bar-count scheduling; PHASE* warns that earlier history may change the phase. |
Matured path sample, slots | 32 / min 2, max 200 | Completed-slot retention, default 32 for Latest/Rolling. Missing-data slots consume retention but not valid denominators; Selected interval uses all completed slots. Open paths are not successes or failures. |
Retained drawings, paths | 4 / min 0, max 12 | Visual path cap, default four, maximum twelve, including an open path. Zero hides paths without removing observations. |
Frozen path drawing | Forward only / options: Forward only, Fit + forward | Forward only shows reached outcomes. Fit + forward restores the original training fit as faint dotted lines, known only at its origin, not at every earlier candle. |
Historical line transparency | 48 / min 10, max 85 | Historical-path transparency, default 48. |
Historical line thickness | 1 / min 1, max 3 | Historical-path thickness, default one. |
Text-only path IDs and results | On | Shows F IDs and text results: ALL, BACK, OUT, DATA, or unfinished k/L. Final labels appear only at maturity. |
First-exit residual connector | On | Connects the first sampled outside Source value to the fixed center. It is not a stop order, intrabar crossing or fill. |
Observed-end residual connector | On | Connects the latest evaluated Source value to the fixed center. |
| Input label | Default / options | What it controls and when to adjust it |
|---|---|---|
Numeric history page | Match panel / options: Match panel, Model, Refit research, Horizon research, Frozen paths | Selects numeric history independently, or Match panel follows the visible page. M/R/H/F and the page ID identify field meanings. The 53 fields are available across pages, not simultaneously; switch and export again using the tooltip mapping. |
Create alerts manually in TradingView. They operate in Latest, not As-of research. Path start/completion/exit bits are research records rather than extra standalone alert conditions.
This tool is for analysis and visualization. It is not investment advice, a trade recommendation, a forecast, or a promise of profit or accuracy. Results depend on inputs, symbol, timeframe, and data feed. You remain responsible for every trading decision.
Current fitted history redraws as the window changes. Restored training fits were known only at their origin. Fixed-path coefficients remain fixed within a path, but changed settings, history, schedule phase or feed revisions can change the reconstructed study.
Wide envelopes can have high coverage with little directional information. High R² and a 90th percentile do not establish future predictive skill, calibrated probability or significance. The components do not vote on an optimal model or trade score.
DATA consumes a retention slot but is excluded from valid denominators. k/L is unfinished, not success or failure. Mean first-exit age conditions on exited valid completed paths; it does not assign zero to unexited paths or estimate survival.
PHASE* warns that the anchor is at/before loaded history and more history can move the schedule. Anchor within loaded history with preceding bars available and hold symbol, timeframe, settings and interval fixed. Invalid origins are skipped, not relocated. F IDs are local, not permanent identifiers across recalculations.
RANGE > CAP requires narrower dates or a larger budget. Remaining traces or records may be partial, not a complete interval result. Drawing retention differs from numerical slots, and rolling aggregation differs from interval-wide aggregation.
Bars are not calendar duration. Non-standard candles, external sources, missing observations and feed revisions can change interpretation. No volume, external requests, higher-timeframe data or imported libraries are needed; ATR is only for glyph clearance. Non-overlapping outcomes do not make market data or training windows independent.
This is descriptive model research, not order placement or a strategy simulation. Containment and error skill are not profitability. Compare like-for-like symbol, timeframe, Source, width and valid sample counts.
Clarify fitted history, endpoint coverage and fixed-path interpretation.
No. It is a fitted-window residual rank, not a calibrated future probability or confidence interval.
Not necessarily. Center movement, width expansion or both can accommodate the same observation; inspect matched before/after geometry.
BACK finishes inside after an earlier sampled exit; ALL keeps all sampled outcomes inside. Neither is a profit result or an intrabar-wick guarantee.
No. It reconstructs a selected model but does not crop the chart or provide blind replay. Alerts are disabled.
It is a free open-source TradingView indicator. The publication specifies Mozilla Public License 2.0; check license conditions before reuse.
This guide is grounded in SG Group’s live catalogue, the script’s published TradingView description, and the user-facing inputs in its open source. Labels and defaults can change after a script update.
