Function starcBands

STARC Bands (Manning Stoller — Stoller Average Range Channel) — a moving average of the close with bands a multiple of the average true range away:

middle = MA(close, period)                       ${prefix}Middle
upper  = middle + multiplier × ATR(atrPeriod)    ${prefix}Upper
lower  = middle − multiplier × ATR(atrPeriod)    ${prefix}Lower
Copy

Appends three columns. Reads high, low and close (the ATR needs all three; the centre needs the close), each named by an option defaulting to its DEFAULT_OHLCV name — the atr shape.

Three ATR channels ship, and they are genuinely different studies

The package now has keltner, atrBands and this one, and the difference between them is worth stating once because "MA ± k·ATR" describes all three:

So starcBands is not a rename of atrBands: atrBands draws its bands around a column the caller already has and appends no centre, while this study computes a moving average that is not otherwise on the series and therefore appends it (the keltner / bollinger rule — a computed centre is worth a column, a copy of the caller's own field is not).

It is, deliberately, the same arithmetic: composed by hand, starcBands is movingAverage(…) followed by atrBands({ column: <that average> }), and a test pins the two equal bar-for-bar rather than leaving the claim in prose. The study ships because that composition needs a scratch column and the corpus names this one (assessment §6.2) — vocabulary over a kernel, which is what this package is.

Which parameters — the formula is uncontested, the defaults are not

Stoller's construction (a simple average of the close, ±2 ATRs) is not in dispute; the parameter set published with it is. Short centres (5–6 bars) appear beside 20-bar ones, and atrPeriod 15 is the value most commonly quoted with the study. What ships is period 20, atrPeriod 15, multiplier 2, maType 'sma' — the 20 chosen to sit beside keltner's and bollinger's rather than because a vendor defaults to it, and every one of the four is an option. Said here rather than left to be discovered from a chart that does not line up.

maType is the shared MaType menu, spelled maType because the average is an ingredient rather than the output (the keltner side of that split). TA-Lib has no STARC to arbitrate; the oracle is a pandas replication over the same _atr_series reference TA-Lib's ATR is asserted against, so the numbers are TA-Lib's ATR with arithmetic on top.

Warm-up: per column, not per study

The centre starts at the MA's own first bar (period − 1 for the window types on gap-free input, later for the composed ones — see movingAverageColumn); the ATR at bar atrPeriod (true range needs a previous close, so a period-bar average of it lands one bar later than a period-bar window would). The bands therefore start at the later of the two, and the centre is emitted where it is genuinely defined rather than masked back to them — the macd per-column rule. At the defaults that is bar 19 for the centre and bar 19 for the bands (the MA's 19 beats the ATR's 15); at { period: 5, atrPeriod: 20 } the centre lands at bar 4 and the bands at bar 20.

Edges

• In the units of the price, like atr and keltner: scaling every bar scales the whole channel, and adding a constant translates it without changing its width. Both pinned by property tests.
• A leading gap shifts the start rather than emptying the study.
• An interior gap in the close is the Wilder asymmetry, and the two halves genuinely differ. The centre takes the K2 engine's column door, where 'sma' keeps core's count-window contract — minSamples counts rows, and avg skips a missing cell — so the centre is drawn straight through the gap (over period − 1 contributors on the windows containing it). The ATR is Wilder over true range, which reads the previous close, so the gap costs the next bar's true range and the recursion never gives it back: the bands are defined on the gap bar itself and blank from the one after it to the end. Measured and pinned in study-missing-cells.test.ts. The other maTypes follow their own rules (movingAverageColumn): the ema family skips the bar, smma and kama propagate. A caller who needs continuity fills before smoothing. The same asymmetry keltner documents, though keltner's centre smooths a derived typical price and so loses the bar where this one does not.
• No division anywhere, so there is no zero-denominator case: a flat stretch gives ATR = 0 and a zero-width channel, which is the honest reading (contrast bollinger, whose zero-width statistical band is emitted as undefined).