ENSO cycles are inherently very disturbed, since they range from 2-7 years; periods no one can reliably predict yet.
Not bulk ENSO noise warned, but pervasive and itself chaotic sampling and sensor noise from ocean waves, clouds, temperature fluctuations, sea-bottom features, coastal irregularities, sensing array gaps, you name it. Lots of noise. Of course, if you are right, you'll predict lunisolar driven ENSO periods regardless of all concerns. But if your predictions don't match actual outcomes, at least a long list of suspected noise sources is already going.
Machine learning? Great. Lets see how statistical mechanics and knowledge based heuristics compare with a Black Box. Don't expect an exact match to your model. All models overlaid could beat any one model. Weather is best predicted by multi-models.
Many unidentified squiggles in the data will turn out to be stochastic superpositions of factors, not discrete "straightforward" events. "Noise" is indeed punting, but "deterministic chaos" is running the ball.