Binary search limits on multi-peak average grids

Two decimals on one database

A worked system-optimization example tuned only two moving-average inputs: one exponential-average decimal for buying and one for selling.

The buy decimal was stepped from 0.1 to 0.3 by 0.05 and the sell decimal from 0.15 to 0.35 by 0.05, with the same historical database applied to every pair. That listing is a parameter-step-grid.

Several peaks and a valley

Results on that two-parameter moving-average grid formed several peaks, a valley along one sell-decimal line, and local dips rather than a single hill that rose and fell in one direction.

The map is a multi-peak-surface: more than one local high, plus valleys or dips, so a climb from one start need not reach another high.

Varying more than two parameters together was expected to make the optimization surface still more complex.

The binary path depends on the start

Binary optimization is a successive-step search that moves from a chosen starting pair and increment rather than scoring every cell on a parameter grid.

A binary search with an initial increment of 0.05 can miss other peaks. The path depends on that increment and on the starting buy and sell values.

Four illustrated starting pairs were (0.10, 0.15), (0.15, 0.35), (0.30, 0.30), and (0.30, 0.20). Each was described as unable to reach at least one of the other observed peaks.

Binary optimization was said to reduce computing time but to be appropriate only when the objective surface has a single peak.

The sample used to set the pair

Even if a search method is sound, choosing rules on historical prices remains a robustness problem because later data seldom match the sample used to set the parameters.

A contemporaneous test that found no benefit from optimization was faulted for not stating the search procedure or whether the testers faced a single-peak surface.