Four-span MACD lookbacks as perishable parameters

The indicator uses four lookbacks

The indicator is a four-span oscillator: an exponentially smoothed difference of two simple moving averages of the close, then compared with a further exponential signal average of that difference.

An n-period simple average is mapped to an equivalent exponential smoothing factor of 2 divided by n plus 1.

The signal-difference is a zero-line event

Subtracting the signal average from the oscillator turns the usual line-cross into a zero-line event. The result is the signal-difference. Under the zero-line-rule, a rise through zero is a buy and a fall through zero is a sell.

One fully specified example uses simple-average lengths of 17 and 56, a 12-period exponential smoother, and a 32-period exponential signal average.

Lookbacks are local and they expire

A single quartet of lookbacks is not expected to be jointly useful across a large stock universe, and some names stay unresponsive after a search over timespans. That is the instrument-local-parameters assumption.

A quartet that appears best for one stock can change as new observations arrive or as older bars fall out of the history used to select it. That tendency is rolling-window-decay.

A local-bracket-search can move each of the four current lengths one unit up and one unit down, which produces 81 combinations.

Fills and the long-only test design

The procedure assumes a fill at the signal-day-close, while also noting that a substantial part of the later price path can arrive well after that close.

The historical experiment used a long-only-test-design on about 50 stocks from 24 March 1986 to 31 March 1988, with warmup leaving roughly 21 months of evaluated signals and with no stop, interest, or commission terms.

The published procedure omitted short sales and treated that omission as unfinished work that would likely change which names appear responsive. Under a long-only-test-design, names whose dominant sample path was down can stay out of view.