Function donchian

Donchian channel (Richard Donchian) — the highest high and lowest low of the trailing period bars, and their midpoint:

upper  = max(high, period)      ${prefix}Upper
lower  = min(low, period)       ${prefix}Lower
middle = (upper + lower) / 2    ${prefix}Middle
Copy

Appends three columns; each is undefined for the first period − 1 rows. Reads high and low (no close), each named by an option defaulting to its DEFAULT_OHLCV column.

Definition

There is only one: upper is exactly what rollingMax gives over high and lower what rollingMin gives over low — this study is those two edges in one rolling scan (via highestLowestValues, the kernel it shares with stochastic and williamsR), plus the midpoint. Verified against pandas rolling(n).max() / .min() in the oracle fixture; TA-Lib has no Donchian to compare with.

Edges

• In the units of the price, like atr: scaling every price scales the channel with it. It does not normalise.
• A missing high or low is skipped, not propagated: the edge is the extreme of the cells the window does hold, and only a window with none reads undefined. This is core's reducer policy and the same contract rollingMax / rollingMin already ship — including over another study's output, where the channel starts on that study's first bar with whatever the window holds by then, rather than period − 1 bars later.
• upper and lower bound close by construction only when low ≤ close ≤ high holds on the input bars; the study does not check that, and a close outside its own bar's range will sit outside the channel too.