The toolkit is real: 408 Python files, 156,000 lines. But it is not "buy this stock now" advice. It is a calculator for pricing derivatives and measuring risk.
Reading alone would have misled me, so I called all 367 functions. 83 work without an account; 62 need one. Sharpe ratio looks like a pure calculation but quietly calls Goldman's server for the risk-free rate.
The real findings were ours
- Live P&L recorded zero trading costs — and that number was compared against a cost-adjusted expectation to trigger a "strategy is underperforming" alert. The alert could never fire.
- The benchmark used to decide whether to keep the bot running did not match the code — the script that produced it had been crashing for a month and nobody noticed.
- The capital split was justified by a look-ahead test — stocks picked with 2025 hindsight, then backtested from 2023.
One of my own proposals was wrong
I suggested replacing a fixed threshold with a per-stock one. It sounded right. Measured over 12 years, it made results worse — tightening the threshold cut off the best performers. So the rule stayed; only the information was added to the screen.
The most valuable thing I took from Goldman's code was not a formula. It was a clock that makes look-ahead bias impossible rather than merely discouraged.
