TA Library

Introduction

kScript v3 ships a complete, reference-validated technical-analysis library. Every indicator below is tested bar-for-bar against an independent textbook implementation (relative error at most 1e-6 across the validation corpus, with exact na warmup positions), in both of the engine's execution paths. When kScript says RSI, it computes Wilder's RSI to the decimal.

The catalog

Moving averages & smoothing sma, ema, alma, swma, wma, vwma, hma, rma (legacy recurrence; see notes)

Oscillators & momentum rsi, wpr, cmo, tsi, macd (line/signal/histogram), stoch (%K/%D), cci, mfi, change, roc, mom

Trend & volatility adx, psar, supertrend (line/direction), tr, atr, hl2, hlc3, ohlc4, hlcc4, bb (basis/upper/lower), keltner (basis/upper/lower), stdev, variance

Volume obv, vwap (cumulative, or calendar-anchored via vwap(anchor="day"|"week"|"month")), cum

Statistics correlation (Pearson), median, percentile (nearest-rank, 0..100), linreg (least-squares endpoint)

Bars & events highest, lowest, highestbars, lowestbars, pivothigh, pivotlow, valuewhen, barssince, rising, falling, crossover, crossunder, fixnan, isna, nz

timeseries basis = sma(source=trade_data, period=20)
var bands = bb(source=trade_data, period=20, mult=2)     // multi-output
timeseries st = supertrend(factor=3, atrPeriod=10)

Conventions (read this once)

These rules are uniform across the library; knowing them explains nearly every edge case:

Warmup is na. An indicator with a p-bar window emits na until it has a full window of finite values. No fabricated early values. Compose with isna/nz as needed.

Indicators compose. Any indicator accepts any series as source, including other indicators (rsi(source=sma(...))) and microstructure-derived series. Warmup propagates correctly through the composition.

Values forward-fill; events do not. Reading a value series past a gap or a shorter timeline forward-fills the last value (so math keeps working). Event conditions (valuewhen, barssince, the crosses) are stricter: a missing condition is false, never a carried-over true, so stale data cannot fabricate signals.

stdev vs stddev. stdev is the textbook one: strict full finite window, population divisor; any missing sample in the window yields na. stddev is the legacy lenient variant (skips non-finite samples, divides by the finite count), preserved unchanged for old scripts. New code should use stdev. The same strict-window rule covers bb, variance, correlation, median, percentile, linreg.

Gaps in stateful indicators are honest. supertrend emits na on a bar whose inputs are not finite (and re-converges after), rather than carrying a stale flip.

rma note. The legacy rma builtin uses a recurrence based on the previous source value rather than the previous RMA, so it is not Wilder smoothing. rsi, atr, and adx implement true Wilder smoothing internally; for textbook smoothing of an arbitrary series, build on ema/sma.

Multi-output indicators

macd, bb, keltner, supertrend, and stoch return multiple streams; plot family functions accept them directly and each stream is addressable by name:

Named streams support history indexing (macd.signal[1]) and feed fillBetween directly:

var bands = bb(source=trade_data, period=20, mult=2)
fillBetween(bands.upper, bands.lower, "#0ea5e9", 0.12)
plotLine(bands.basis, color="#64748b")

Over microstructure

Every windowed function in the library also runs over order-flow-derived series, which is where the library stops being a Pine clone:

timeseries delta = fp.cells.map((c) => c[2] - c[3]).reduce((s, x) => s + x, 0)

timeseries deltaRsi = rsi(source=delta, period=14)      // delta-RSI
timeseries cvd = cum(delta)                              // cumulative volume delta
timeseries imbalanceMa = ema(source=bookImbalance, period=21)

The validation suite covers these compositions explicitly (27 windowed functions over a footprint-derived series, plus end-to-end microstructure-only indicators: CVD, delta-RSI, largest-bucket drift, book-imbalance EMA, POC migration speed).

Accuracy as a contract

The reference suite (independent plain implementations, never the engine itself) is part of the build: a change that drifts any indicator from its textbook formula fails CI. v3 in fact corrected eight long-standing indicators to their textbook forms; if you maintain very old scripts, see the migration notes for the early-bar value changes that correction implies.