ml_robuststats

A small robust-statistics toolkit for building stable, outlier-resistant indicators.

Provides MAD-based robust z-scores (median/MAD, breakdown point 0.5), robust variance,
winsorization, percentile rank, multi-window boolean consensus, and inverse-variance
"precision" weighting (ridge-regularized, capped, normalized) for combining several
factors into one composite so quiet/reliable inputs get more weight and noisy ones fade.

Intended as a shared building block imported by other scripts, not as a standalone chart
indicator. MPL-2.0.

Library "ml_robuststats"

robMedian(src, len)
Parameters:
src (float)
len (simple int)

robMAD(src, len)
Parameters:
src (float)
len (simple int)

robZ(src, len)
Parameters:
src (float)
len (simple int)

robVar(src, len)
Parameters:
src (float)
len (simple int)

winsor(x, cap)
Parameters:
x (float)
cap (simple float)

pctRank(src, len)
Parameters:
src (float)
len (simple int)

consensus(sShort, sMid, sLong, mode)
Parameters:
sShort (bool)
sMid (bool)
sLong (bool)
mode (simple string)

precisionWeights(variances, wCap, ridgeFrac)
Parameters:
variances (array<float>)
wCap (simple float)
ridgeFrac (simple float)

dot(values, weights)
Parameters:
values (array<float>)
weights (array<float>)

版本注释

ml_robuststats is a dependency-free library of robust-statistics primitives for Pine v6. It exists to fix a common weakness in indicator maths: the ordinary mean/standard-deviation z-score is dominated by a single fat-tailed bar — one gap, news spike, or expiry print stretches the σ scale and distorts every band and threshold built on it. The estimators here use the median and the median-absolute-deviation instead, which tolerate up to half the sample being outliers before they break down, so bands and scores keep a consistent meaning across regimes.

Everything is exported for reuse via import. Each function takes a series float source (and a simple int window where relevant), so they slot in wherever you currently compute a z-score, a percentile, or a weighted composite.

What's inside
robMedian(src, len) — rolling median. The robust centre of a window.
robMAD(src, len) — rolling median absolute deviation. The robust spread of a window.
robZ(src, len) — robust z-score: (x − median) / (1.4826 × MAD). The 1.4826 factor makes MAD ≈ σ under Gaussian data, so a robust z reads on the same scale as a classic z but shrugs off outliers. Returns 0 when the window has essentially no spread. This is the drop-in replacement for a mean/stdev z-score.
robVar(src, len) — robust variance, (1.4826 × MAD)². Useful as the per-factor variance input to precision weighting.
winsor(x, cap) — clamp a value to ±cap. A second layer of outlier control to apply before compositing several factors.
pctRank(src, len) — percentile rank 0–100. Self-calibrating strength that needs no fixed threshold; naturally bounded and outlier-tolerant.
consensus(sShort, sMid, sLong, mode) — multi-window boolean agreement across a short/medium/long signal, with mode "Any" | "Majority" | "All". Turns three single-window reads into one agreement gate.
precisionWeights(variances, wCap, ridgeFrac) — inverse-variance (precision) weights from an array of per-factor variances, ridge-regularized, capped per factor, and normalized to sum to 1. Quiet, reliable factors earn more weight; noisy ones fade — a principled replacement for hand-set weights. wCap prevents any single factor from dominating; ridgeFrac stabilizes the weights when a variance is near zero.
dot(values, weights) — safe dot product of a values array with a weights array (handles size mismatch and na), for turning factors + precision weights into one composite score.
How to use

Import the library, then call what you need. A robust z on any series:

//version=6
indicator("Example — robust z", overlay = false)
import Market_Logic_India/ml_robuststats/1 as rs

len = input.int(100, "Window")
z = rs.robZ(close - ta.vwap(hlc3), len) // robust vs (x - mean)/stdev
plot(z, "robZ", color = color.new(#5b9cf6, 0))
hline(0)

A precision-weighted composite of several factors:

import Market_Logic_India/ml_robuststats/1 as rs

f1 = rs.robZ(factorA, 100)
f2 = rs.robZ(factorB, 100)
f3 = rs.robZ(factorC, 100)

var array<float> vals = array.new<float>(3)
array.set(vals, 0, f1), array.set(vals, 1, f2), array.set(vals, 2, f3)

var array<float> vars = array.new<float>(3)
array.set(vars, 0, rs.robVar(factorA, 100))
array.set(vars, 1, rs.robVar(factorB, 100))
array.set(vars, 2, rs.robVar(factorC, 100))

w = rs.precisionWeights(vars, 0.6, 0.10) // cap 0.6, ridge 10%
composite = rs.dot(vals, w)
plot(composite, "composite")
Notes
Non-repainting: all estimators use only closed historical bars — no request.*, no future references. The last bar updates intrabar and settles on close.
Warm-up: the rolling estimators need len bars of history before they are meaningful; treat the start of history as warm-up.
Types: pass simple int windows and simple float caps (plain inputs or constants), not series values. consensus takes a simple string mode.
precisionWeights and dot operate on array<float> — build the arrays each bar (or as var arrays you overwrite).
Concept credits

The median / MAD robust z-score (with the 1.4826 consistency constant), winsorization, inverse-variance precision weighting, and percentile-rank strength are standard robust-statistics and meta-analysis techniques. This library is an original Pine v6 packaging of those public methods for reuse; it is not affiliated with, nor endorsed by, any originator.

License

Mozilla Public License 2.0 — as required for TradingView libraries (open source). Free to import and build on.

版本注释

v3

Added:
pctCap(src, pct, len)
Parameters:
src (float)
pct (simple float)
len (simple int)

Pine脚本库

秉承TradingView的精神，作者已将此Pine代码作为开源库发布，以便我们社区的其他Pine程序员可以重用它。向作者致敬！您可以私下或在其他开源出版物中使用此库，但在出版物中重用此代码须遵守网站规则。

Market_Logic_India

免责声明

这些信息和出版物并非旨在提供，也不构成TradingView提供或认可的任何形式的财务、投资、交易或其他类型的建议或推荐。请阅读使用条款了解更多信息。

Pine脚本库

Market_Logic_India