Optimized vs Default Indicator Settings: The Optimizer Won 3 of 4 and Still Lost to the Spread

There are two ways to choose the numbers inside a strategy. You can use the ones in the textbook, which everyone else is also using, or you can search history for the ones that would have worked best. The first is accused of being arbitrary, the second of being curve fitted. I ran both on the same 16,926 EURUSD 5-minute bars and let the second half of the data decide.

How were defaults and optimised settings compared?

• The defaults: the settings that appear in almost every tutorial and in MetaTrader’s own dialogs — MA 50/200 crossover, RSI 14 with a 30/70 reversal, Donchian 20-bar breakout, Bollinger 20 period at 2.0 standard deviations.
• The optimised settings: the best total-pips setting in each family, chosen from a 129-set grid using only the first half of the data, exactly as described in my in-sample versus out-of-sample test.
• The judge: the second half, 15 July to 26 August 2026. Neither side saw it while being chosen. A 24-bar hold, no stop, no target, same measurement as every other test in this series.

Which won on data neither side had seen?

On the scoreboard the optimiser wins, and I want to be fair to it: three families out of four, and a 140 pip swing on the total. That is a real result, and it is the opposite of the conclusion most people expect from a curve fitting article.

But look at where the margin comes from before you draw anything from it.

Why did the defaults lose, exactly?

Almost all of the deficit is one rule. RSI 14 buying a cross back above 30 and selling a cross back below 70 lost 177.8 pips across 216 signals, at 44.0% winners. Take that single line out and the remaining three defaults finish at +91.7 pips combined — ahead of where the whole optimised set landed.

So the honest reading is not “optimisation beats defaults.” It is: one very popular default is bad, and both methods agreed it was bad. The optimiser searched sixteen RSI variants and its winner still lost 156.6 pips. It improved the loss by 21 pips, which is not a fix, it is a rounding of the damage. When a family has no edge, tuning the numbers inside it changes nothing that matters — the same conclusion I reached comparing Stochastic against RSI for scalping, where neither oscillator produced anything worth trading on 5-minute bars.

The finding I did not expect: the optimiser chose the default

On Bollinger bands the grid contained 16 combinations — four lookbacks crossed with four band widths. The setting that maximised in-sample pips was 20 period at 2.0 standard deviations, which is the setting John Bollinger published and the one MetaTrader fills in for you. The optimiser ran the whole search and came back with the textbook.

That is worth sitting with, because it cuts both ways. It is the strongest evidence in this study that the classic numbers are not arbitrary — twenty and two really is near the top of its family on real EURUSD data. It is also, from the optimiser’s point of view, its most useless day of work: +519.6 pips in-sample and +4.4 out-of-sample, a 99.2% collapse, delivered by a setting you would have had for free.

Was the optimised edge stable or lucky?

The three optimised wins are not equal in quality, and the difference is visible in how each was chosen.

Only the first is a genuine win. The Donchian pick was selected from a family in which nothing made money in the first half; the optimiser picked the smallest loss, the market turned friendlier, and it printed +100.1. Crediting optimisation for that is like crediting a raincoat for the weather clearing. The default 20-bar Donchian also flipped from -497.2 to +65.4 over the same two halves, which tells you the regime did the work, not the parameter.

MA 5/150 is the exception that shows what a real result looks like: positive on the half it was chosen from and positive on the half it was not, from a family whose rank correlation between halves was +0.574 — by far the highest of the four. Structure carries. My EMA against SMA comparison found the same thing from the other direction: the type of average barely mattered, the length mattered a great deal.

What does one pip of spread do to the comparison?

Both columns above are gross. Using the roughly one pip round-turn cost I measured in my spread and swap breakdown, the picture changes shape entirely, because the two approaches trade at very different frequencies.

Of every rule in this comparison, default or optimised, one ends the out-of-sample period in profit after costs. Across the full 128-setting grid, the share of profitable settings drops from 67.2% to 21.1% once a pip is charged. The frequency of a rule turns out to be a bigger lever on the final number than the settings inside it — the default Donchian traded 228 times to make 65 pips and handed back 162 in costs.

So which should you use?

Neither question is the right one. The measurable facts are: defaults are not sacred, since one of the four lost 177.8 pips; optimised settings are not an upgrade, since three of four picks fell apart or were rescued by the regime; and the difference between the two methods was smaller than the difference between the two halves of the summer.

What separated the winners from the losers here was not the choice method. It was whether the rule survived its own trade count. That is the same conclusion I keep landing on, whether I am testing ADX against Supertrend as a filter or trading one session versus all day: taking fewer, larger trades beats tuning the entry.

What I would actually do with this

• Start with the default and try to break it. On Bollinger bands a full search returned the default anyway. That is one afternoon saved.
• Only optimise structure. Lengths and speeds held up between halves. Threshold levels like RSI 30 did not.
• Check whether the family works before tuning it. The optimiser turned a -177.8 pip rule into a -156.6 pip rule. Nothing inside a broken family is worth finding.
• Compare after costs or not at all. Optimised beat defaults by 140 pips gross and by far less once frequency was paid for.
• Distrust a win that came from a family where everything lost. The Donchian pick’s +100.1 was the regime, not the setting.

Frequently asked questions

Are default indicator settings any good?

Three of the four I tested were reasonable and one was actively harmful. MA 50/200 made +21.9 pips out-of-sample, Donchian 20 made +65.4, Bollinger 20/2.0 made +4.4, and RSI 14 at level 30 lost 177.8. Notably, a 16-way search of Bollinger settings returned the default as its winner, so at least one classic number is classic for a reason.

Is it worth optimising indicator settings at all?

Only for structural parameters, and only with held-back data. Optimised settings beat defaults in 3 of 4 families here, but one of those wins came from a family where every setting lost money in-sample, and the overall gap shrank to almost nothing after spread.

Why did both RSI versions lose money?

Because the family had no edge on this instrument and timeframe, and no parameter fixes that. Reversal entries against 5-minute EURUSD moves gave back more on the losers than they collected on the winners regardless of period or level, which matches what I found testing Stochastic against RSI on the same bars.

What settings did you actually test?

129 in total: 88 MA crossover pairs, 16 RSI reversal combinations, 9 Donchian lengths and 16 Bollinger period-and-width pairs. 128 produced at least 30 signals in each half of the data and were kept.

What data was this run on?

16,926 EURUSD 5-minute bars, 3 June to 26 August 2026, the same set behind my moving average crossover test and my Ichimoku versus trend following comparison. The split point is bar 8,463, 15 July.