Does combining two indicators actually help?

Running 2,878 pairs guarantees winners. If every filter were pure noise, about 1439 of them would still beat their trigger, some by a wide margin, and the top of any ranking would look compelling. That is not a flaw in the test, it is arithmetic, and it is the reason most published indicator combinations are worthless: somebody searched until something looked good and then stopped.

So the column that decides anything is the last one. Every window has its final 30% scored separately, and a pair is only interesting if it still beats its trigger there. 582 of 802 did, against roughly 401 expected from chance. That gap is real and it is the one result on this page worth acting on.

One caveat we cannot test away. The out-of-sample third is the most recent third, so it sits next to the data it is being compared against. A filter that suits one market regime can survive the split simply because the regime did. Treat the survivors as worth investigating, not as settled.

Not which indicators they are. How much they refuse. Splitting every pair by the share of the trigger's own entries the filter let through:

A filter that agrees with almost everything is worse than no filter. The 2,276 pairs that let through 70% or more of the trigger's entries helped only 27% of the time — 22 standard deviations below a coin toss, so this is not noise. An indicator that rarely disagrees adds its own lag and its own cost and contributes no information, and that describes most of the pairs in this study.

The selective end looks better, and the sample is too small to bank it. The 112 pairs rejecting more than 70% of entries helped 49% of the time, which sounds decisive and is only 0.2 standard deviations from chance on 112 pairs. Selective filters also produce fewer trades, and fewer trades means noisier estimates in both directions. It is the most promising thing on this page and it is not yet a result.

What the two rows together do say, with more confidence than any individual pair in the table: if a second indicator is going to earn its place, it earns it by declining setups, not by nodding along.

56,239 backtests of 2,878 indicator pairs over 14,326,491 trades, each pair scored against the indicator that fires it rather than against zero. The full dataset is downloadable and free to reuse under CC BY 4.0 — which asks only that you say where it came from.

The filter is not required to fire on the same candle as the trigger. Crossovers are rare, and asking two of them to land on the same bar would test coincidence rather than agreement. What a trader means by “the filter agrees” is the indicator's standing opinion — its most recent signal — and that is what is used.

The pairing is asymmetric on purpose. “Enter on RSI, but only while the trend indicator agrees” is a different rule from the reverse, and both appear in the table separately.

Everything else matches the single-indicator laboratory: published settings, entries at the next bar's open, the stop four median bars wide with the target at twice it, the real per-pair spread charged on every trade, and the last 30% of each window scored separately. Pairs taking fewer than 60 trades on a chart are not recorded at all.

It runs on 4 instruments rather than the laboratory's fifteen, because 2,878 pairs on six timeframes is already 56,239 backtests. The instruments are a major, a second major, a yen major and the most volatile cross we carry, so a result that holds on one and fails on the rest is visible in the table.