How to choose parameter ranges.

The best starting point for any range is the input's own default and the logic behind it. Most strategy authors ship defaults that are at least reasonable, so treat that value as the center of gravity and open a window around it rather than guessing from scratch.

Let the meaning of the input set the shape of the window. A fast moving average that defaults to 20 belongs in a band of small-to-medium lengths, not stretched out to 500, because a 500-bar average is no longer fast in any useful sense. A risk multiplier that defaults near 2 should explore tighter and looser stops around that, not values so small the strategy is stopped out instantly or so large the stop never triggers. Ask what the parameter does, then pick a range where every value in it is still doing that job.

Here is the trade-off that governs every range you set. The number of values an input contributes is roughly its span divided by its step. Halve the step and you double the values. And because the optimizer has to consider combinations across inputs, the totals multiply: the full grid is the values of the first input times the values of the second times the third, and so on. Add one more parameter or tighten one step, and the search space can balloon.

That total is not abstract. Each combination an exhaustive search examines is one optimization cycle, and cycles are your budget. The table below shows how quickly the count climbs for a moving-average length and a stop multiplier, the two inputs we work through in detail further down.

You rarely need a fine step across the whole range on the first pass. The efficient workflow is two stages. Start coarse: keep the range wide but the step large, so a cheap run sketches the broad shape of the space and reveals which region looks promising. Then go fine: shrink the range to that promising window and tighten the step to resolve the best value inside it.

A coarse sweep of a moving-average length from 10 to 200 in steps of 10 is only 20 cycles, yet it tells you whether the strategy prefers short, medium, or long averages. If everything good clusters between 40 and 60, your second pass searches just 40 to 60 in steps of 1, another small run that pins the winner precisely. Two cheap, focused passes beat one giant blind one almost every time.

Two methods automate exactly this narrowing for you. Bisection repeatedly tests around the middle of a range and discards the half that performs worse, zeroing in without an exhaustive scan. The compound bisection-then-TPE-refine method goes further: it brackets the promising region with bisection, then hands that smaller region to the TPE sampler to polish the final value, doing the coarse-to-fine routine in a single run.

A few range mistakes waste cycles or quietly corrupt a run. Steering clear of them is half the skill of setting good ranges.

Once your ranges are set, the search method decides how efficiently they get explored. Read how to choose an optimization method to pair the right algorithm with your grid, then dig into the TPE Bayesian sampler and bisection to see how they find good settings without testing every combination.

New to any of the terms here? The optimization glossary defines parameter space, step, cycles, and the rest in plain language, and the getting started walkthrough shows where the min, max, and step fields live in the strategy dialog.