I didn't alter the sentiment of the trades though, the sentiment of every single one of those 58M trades is exactly as it was.
but by changing the trades, (making the $1000 never happen) you are changing the sentiment, which then effects future sentiment. and that is the exact problem, by not considering that sentiment will change because of your actions. you have failed to take that into account..
as ciyam said.. it wont work with live data as that involves live human emotion and change of sentiment in reaction to your model
What you're talking about is the market dynamics during a pump (FOMO) and dump (fear), which is correct.However, you're wrong. The test is done so that the sentiments are tested, which gives a worst case scenario for the price stability mechanism. In a live testing, these extremes will be much less pronounced and hence the stability will be easier to keep. It does not invalidate the algorithm, it backs it up since it will likely not see such pressures in live environment.