AI Created a Profitable SPY Strategy. Then We Tried to Break It transcript

AI found a profitable SPY strategy in 30 seconds, then we tried to break it. The rules were intentionally simple. SPY above its 200-day moving average, RSI 3 below 20 by next open, and exit after a rebound. I wanted to be fair to the idea, so we froze those rules before touching the later data. No tuning after the fact. At first, it looked excellent. 253 trades, 75.9% winners, a 2.60 profit factor, and an average trade of 0.60% $10,000 compounded across the closed trades grew to roughly 44,000.

But, the pretty equity curve was not the point. The real test was what happened next. First, we split the history. From 1993 through 2014, profit factor was 2.93. From 2015 onward, it dropped to 2.18. That is weaker, but still solid across 104 trades. I actually like seeing some deterioration here. Perfect validation would make me more suspicious, not less. This is a retrospective holdout, not a live test. Next came the overfitting test I care about most.

We changed the RSI length and the entry threshold. All 20 nearby combinations remained profitable. Some were better, some worse, and the strongest cells had fewer trades. But, there was no single magic setting holding the whole strategy together. The profitable area was broad. We also moved the trend filter from 150 to 250 days. The full sample profit factor stayed between 2.58 and 2.73, and the later period stayed above two.

Again, no cliff around the original 200-day setting. That is what you want to see from a simple rule. Then, we changed the exit completely. One day, three days, five days, 10 days, 21 days, and the original rebound exit. Every version stayed profitable. The original exit was best, but the other exits still worked. That suggests the entry condition itself may contain a real historical tendency rather than the result depending on one clever exit.

Then we deliberately made execution worse. At 10 basis points of round trip friction, profit factor was still 2.25. Even at 20 basis points, it was 1.94. For SPY, that is a deliberately harsh stress test. Costs hurt the result as they should, but they did not erase it. Finally, we split the test into four market eras. The weakest was 2010 to 2019 with a profit factor of 1.85. That is the number I care about here. It tells us the full result was not created by one lucky decade.

The strategy was not equally strong everywhere, but every major period remained profitable. So, did the AI strategy survive? Under these tests, yes, but not perfectly. Later data was weaker. Costs reduced the edge. Some periods were mediocre. That is exactly why I find the result interesting. It looks less like a perfect back test and more like something that deserves further research. AI can generate ideas incredibly fast.

The difficult part is still deciding which ideas deserve any trust. Historical tendency, not a forecast. If you like this kind of rule-based research, you are very welcome to join our school community, rule-based trading. We publish one new strategy every day. The link is school.com/rulebasedtrading.