"Shouldn’t lead the term: tcos(6.192 + 153t) to an increasing signal?"
That is true, but the value of t ranges from 1950 to 2000, so that this amplification is very slight, like 2 to 3% of the signal amplitude. I am not sure if this can be perceived visually amongst the fluctuations, but that is what the tool is finding.
So I have no control over the machine learning fit that Eureqa executes, which is good and bad I suppose. Good because it doesn't add any human bias, but bad in that there is no physics involved at this level. For example, Eureqa is not going to say that those numbers are related to lunar monthly cycles, but it is up to the human to figure out the physical mechanisms and decide whether something is just a coincidence.
I am sure that this has some relevance to the recent Azimuth blog post on models and machine learning.