forex-scalping.com

Stacking indicators also makes the settings harder to optimise without overfitting. Each layer adds parameters, like the RSI period, the smoothing period and the smoothing type. With three or four parameters, it's easy to find a combination that fits the past nicely but means little going forward. If you use indicator on indicator setups, I'd limit the number of parameters and test them on data you didn't use for tuning. If the results fall apart on new data, the combination was probably just fitted to the old. A simple walk forward test is enough to show it.

Is there a case where indicator on indicator clearly helps? I think volume is one. A moving average of tick volume, compared with the current tick volume, gives a sense of whether activity is above or below normal, which is useful context. That's applying a simple average to a raw series, not stacking oscillators, and it's easy to interpret. The trouble mostly starts with smoothing things that are already smoothed. A simple layer on raw data is often fine. Several layers on processed data is where the lag and complexity pile up.

The normalising point is a good one, and it separates the useful versions from the pointless ones. I tried putting a 10 period moving average on a 14 period RSI because the raw line was too jumpy. It did calm things down, but the crosses came about two or three bars after the turns I cared about, which on M5 is 10 to 15 minutes. By then half the move was gone.

Here's the test I use now. Take the plain indicator with a slightly longer setting, like RSI 18 instead of 14, and compare the signals with the stacked version over the same 200 trades. In my case the longer RSI gave nearly the same exits, with about a third fewer false flips, and no extra lag.

For a worked example, if each signal trade averages 6 pips of target and the lag costs 1.5 pips of entry, that's a quarter of the profit gone to a smoothing choice. On 0.5 lots that's about 7.50 dollars a trade, and over 200 trades it's 1,500 dollars.

Is there a case where you'd still stack? Maybe volatility normalised by ATR? And do you check for lag by comparing the signal bar to the turn bar?