Indicator Combinations Tested: A Coin Flip Confirmed Better Than All 56 Real Pairs

“Never trade a single indicator.” It is the most repeated piece of advice in retail forex, and the one nobody has ever put a number on. Add a second indicator, wait for both to agree, and your accuracy improves — that is the claim, and it is usually presented as too obvious to test.

So I tested it. All of it. Eight indicators, every ordered pair, 15,208 base signals and 62,804 filtered ones on the identical 16,926 EURUSD 5-minute bars I have used in every study on this site, from the moving average crossover test to the support and resistance test.

The result is not that combinations are slightly overrated. The result is that a random number generator confirmed better than MACD did.

How do you test “confirmation” without arguing about it?

Every indicator gets reduced to one bit at every bar: bullish or bearish. No neutral zone, no interpretation, no discretion.

A signal is the bar on which a state flips. Entry at that bar’s close, exit 24 bars — two hours — later, 1 pip of cost per trade, which is what a EURUSD spread actually costs according to my spread and swap measurement.

A confirmed signal is the same flip, kept only if a second indicator was already in the same state on that bar. That is exactly what confluence means in practice — you do not wait for the second indicator to move, you check whether it agrees. Everything else about the trade is unchanged, so the only difference between the two numbers is the filter.

Which indicators actually agree with each other?

Before asking whether confirmation helps, it is worth asking how often it is even available. Here is the share of bars on which each pair held the same state, from 16,676 bars where all eight were defined.

Two things fall out of that table immediately.

The top row is not a combination at all. CCI above zero asks whether the typical price is above its 20-period mean; the Bollinger midline asks whether the close is above its 20-period mean. They agreed on 95.1% of bars because they are the same question. Waiting for CCI to confirm your Bollinger signal is waiting for a mirror to nod.

The bottom row is stranger. The 50/200 cross and MACD are both trend-following, both built from moving averages, and they agreed less often than a coin flip. That is not a contradiction — one compares a 50-period average to a 200-period average, the other a 12 to a 26 — but it does destroy the idea that “trend indicators confirm each other.” On this data they disagreed 56% of the time.

What did confirmation actually do to the win rate?

Here is the whole study in one table. Each indicator traded alone, then the same signals filtered by each of the seven others, aggregated.

That column on the right is the entire value of indicator confluence on this dataset. The best filter for each indicator — cherry-picked after the fact, which no live trader gets to do — bought between 0.1 and 0.9 percentage points.

And that is the flattering view. Across all 56 pairs:

Unfiltered, the eight indicators produced 15,208 trades at 44.2% and -1.13 pips each. Filtered through each other, the same signals produced 43.1% and -1.32 pips. Confirmation cost 1.1 points of accuracy, 0.19 pips per trade and a third of the trading opportunities.

The single worst pair is the one that should worry anyone running a trend template: the 50/200 cross confirmed by Parabolic SAR dropped from 48.5% to 43.9%, and its average trade from +0.71 pips to +0.02. The best standalone indicator in the study was taken apart by its confirmation. I saw the same fragility when I measured SAR on its own in the Parabolic SAR strategy test.

Does a coin flip confirm as well as a real indicator?

This is the control that makes the rest of the study mean something. I replaced the confirming indicator with a seeded random sequence — every bar independently assigned bullish or bearish with no reference to price at all — and filtered all eight indicators through it.

Five random filters, all five between 43.9% and 44.9%, averaging 44.5%. The 56 real confirming indicators averaged 43.1%. Throwing away half your signals at random produced a better result than filtering them through a second indicator.

I want to be precise about what this does and does not prove. It does not prove randomness is useful — every row in that table still loses money after costs. It proves that the confirming indicator contributed no information. Whatever the second indicator was telling you, a coin was telling you the same thing, or slightly more of it. This is the same discipline that reframed my reading of signal services in the rule versus random entry test: a filter has to beat what you would get for nothing.

What happened when all eight agreed?

If two indicators are not enough, the natural escalation is consensus. So I counted, on every bar, how many of the eight were bullish, and traded only when the count reached a threshold.

Read down the first four rows and the trend is the opposite of the folklore: more agreement, worse results, from 44.0% at four of eight down to 41.9% at seven of eight. Every extra indicator you demanded agreement from cost you accuracy.

Then the last row jumps. Unanimity — all eight bullish or all eight bearish — produced 89 trades at 46.1% and the only positive average outside the moving average cross. Before anyone builds a system on it: that is 89 trades in three months, a total profit of 9.5 pips, and 46.1% off 89 samples has a margin of error wide enough to drive a truck through. It is an interesting row, not an edge.

It is also less rare than it sounds. All eight agreed on 2,663 of 16,676 bars — 16.0% of the time. Unanimity across your entire indicator panel happens for roughly one hour in six.

Why does agreement measure redundancy instead of truth?

Put the two halves of this study side by side and the mechanism is obvious.

Indicators that agree a lot — CCI and Bollinger at 95.1%, RSI and Bollinger at 87.7% — agree because they are computed from the same thing. Confirmation between them is free, constant, and carries no information, which is precisely why filtering changed nothing: RSI confirmed by Bollinger kept 73% of signals and moved the win rate by +0.3 points.

Indicators that disagree a lot — the 50/200 cross and MACD at 43.6% — disagree because they measure different horizons. Demanding that they confirm each other does not combine two views, it deletes the trades where the horizons differ, which is most of them: that filter kept only 79% of a base with 101 signals, and cut the win rate by a point.

So confirmation has two modes and neither one helps. Either your second indicator is redundant, in which case it does nothing, or it is genuinely different, in which case it removes your signals for reasons unrelated to whether they would have worked. There is no third mode where it removes the losing trades and keeps the winning ones. The same logic showed up when I checked candlesticks against oscillators in the candlestick patterns versus indicators comparison, where confirmation agreed 46% and contradicted 46%.

Where the real difference was hiding

One column in the first table is worth more than everything above it. The 50/200 moving average cross produced 101 signals in three months and a positive average of +0.71 pips after costs. The Stochastic produced 4,197 and lost 4,530 pips.

The separation in this entire study was never between one indicator and two. It was between slow and fast. The eight indicators sorted by signal count sort almost perfectly by result, and no amount of cross-confirmation moved any of them more than a point. If you want a lever that matters, it is the timeframe of your signal, not the number of things agreeing with it — which is what I found from the other direction in the best hours to trade EURUSD study, where hour 16 UTC alone ran 51.2% against a 44.2% baseline.

What I would actually do with this

• Stop counting indicators. Going from one to two cost 1.1 points of accuracy and 36% of signals on average across 56 pairs. There is no version of this where the count is the variable.
• Check agreement before you add anything. If a candidate indicator agrees with your existing one more than about 80% of bars, it is the same measurement. CCI and the Bollinger midline hit 95.1%.
• Beware the most-praised filter. Parabolic SAR was the best-behaved confirmer in some pairs and the single most destructive in others, including a 4.6-point hit to the only profitable base in the study.
• Test any filter against a coin flip. It takes ten lines of code and it is the only way to know whether your filter carries information. Mine did not.
• Spend the effort on signal frequency instead. 101 slow signals beat 4,197 fast ones by a margin no confluence rule came close to.

Frequently asked questions

What is the best indicator combination for forex?

On 16,926 EURUSD 5-minute bars, none of the 56 pairs I tested produced a profitable combination after a 1-pip cost. The best pairing by accuracy gain was the Stochastic confirmed by the ADX directional lines at +0.9 points, and it still lost 1,772 pips. The only positive base was the 50/200 moving average cross alone, at +0.71 pips per trade across 101 signals.

Do two indicators confirming each other improve accuracy?

Not on this data. The unfiltered win rate was 44.2% and the confirmed win rate across all 56 pairs was 43.1%, so confirmation lowered accuracy by 1.1 points while discarding 35.9% of signals. Forty-two of the 56 pairs were worse than the indicator alone.

How many indicators should I use on one chart?

Fewer than you think, and the number is not the point. Requiring agreement from more indicators made things steadily worse — 44.0% at four of eight, 41.9% at seven of eight. Only full unanimity across all eight was positive, and that was 89 trades worth 9.5 pips in total.

Which indicators are redundant with each other?

CCI above zero and the Bollinger midline agreed on 95.1% of bars, RSI above 50 and the Bollinger midline on 87.7%, and RSI and CCI on 87.6%. All three pairs are variations of “is price above its 20-period mean.” Adding one to the other changes nothing.

Is RSI plus MACD a good combination?

They agreed on 76.6% of bars, and trading only when both flipped into agreement gave 1,114 trades at 43.0% and -1,428 pips. MACD alone was 43.3%. The pair did slightly worse than either half. I go through the MACD half in detail in my MACD crossover backtest.

Why did a random filter beat a real indicator?

Because the real indicators removed signals systematically — they cut the trades where two horizons disagreed, and those trades were not worse than average. A coin flip removes trades without any bias at all, which turns out to be the less damaging way to throw work away. Five random seeds scored 43.9% to 44.9% against 43.1% for real confirmation.

How were these combinations tested?

Eight indicators reduced to a bullish-or-bearish state at every bar, entry at the close of the bar where a state flipped, exit 24 bars later, 1 pip of cost, no stop and no target, across 16,926 EURUSD 5-minute bars from June 3 to August 26, 2026. Every pair uses the identical rule, so the only difference between any two rows is the filter. The settings grid behind it is in my moving average settings guide.