ai / ai engineering / rust / backtesting / algorithmic trading / performance optimization / developer tools / public portfolio challenge
Why Is My Backtest So Slow? GPT-6 Astra Cut Mine to 51 Seconds.
I run a public portfolio challenge, and every iteration on it started the same way. Kick off a minute-by-minute backtest across 30 strategies and 24 underlyings, wait about ten minutes, change one number, wait another ten. So I tried to make the engine itself faster. Over three months I put four frontier models on the Rust code, Claude Opus 5, Claude Fable 5, GPT-5.6 Sol and Grok 4.6, and every one of them missed the same thing. Opus told me the physics would not allow it to go faster. I believed it, and I started pricing out hardware.
Then I gave GPT-6 Astra the same problem. Its first suggestion was naive enough that I cursed at it and told it to try harder, and what it did next was profile the engine instead of guessing at it. It found a preload spending 125 seconds building 7 GB of data that nothing ever read, which took preparation from 130 seconds to 6. Then it went after my allocation solver and told me the code I had written by hand was an instance of a named problem, the continuous quadratic knapsack problem, citing Cominetti, Mascarenhas and Silva from 2014. A hundred-iteration binary search became one division and seven lines. On the workload that solver touches, eight-ETF risk parity over a full year of minute bars, warm execution went from 470 seconds to 51 on the same five machines I already owned. The ten minutes in this video is my own wait on the public portfolio challenge, which is the reason I started; the 51 seconds is that measured eight-ETF run. I never bought the hardware I was pricing.
Transcript
0:00GPT-6 Astra just made my trading platform the fastest retail trading platform out there.
0:06It started with my public portfolio challenge. I would sit down, kick off a backtest, and wait ten minutes for it to finish. Then I would change one number, run it again, and wait another ten.
0:17So I tried to make the engine itself faster. I threw four different AI models at this code for three months, and every one of them missed the same thing. Opus told me the physics would not allow it to be faster. I believed it, and I started pricing out hardware.
0:31Then OpenAI released GPT-6 Astra, and I gave it the same problem. Its first suggestion was so naive that I cursed it out and told it to try harder. I was sure I was watching another hype cycle.
0:42Then it stopped guessing and started profiling. It found a preload spending 125 seconds building 7 GB of data that nothing ever read. That was the boring part.
0:53Then it went after my allocator. It told me the code I wrote by hand was a known problem with a name, the continuous quadratic knapsack problem, and it cited a paper from 2014. A hundred guesses collapsed into one division. The fix was seven lines.
1:08That backtest took ten minutes. It now takes 51 seconds on the same five machines I already owned. I can sit down and make every change I want in one session.
1:18One of those four models told me this could not be done. That is the impact of GPT-6 Astra. Follow me for more.
Join the conversation
Loading conversation…