Hassan El Mghari told AI Engineer his open-model system forecast eight quarters of Costco profit within 4.4% of actuals.
I call that accuracy an accounting accident because his father and auditor Omar El Mghari showed two bad assumptions canceled out.
Hassan said he turned 214 Costco filings into 3,000 timestamped facts and forced the model to emit formulas as structured JSON.
Hassan said 16 runs of GPT-OSS-120B cost 19 cents and only five passed his self-correction checks before predicting a $560 million two-year profit lift.
Omar argued the model taxed every member for a United States and Canada hike after the July 2024 increase from $60 to $65 and $120 to $130.
Omar said a 15% executive-mix guess against a disclosed 45% turned a lucky 1.7% miss into a 28% miss when fixed, yet the cutoff method stays valid because pre-hike GPT-OSS-120B and Llama-3.3 70B did not know the hike unlike GLM-5.3 and Chimera-3, and GPT-OSS-120B is a 117B mixture-of-experts model with 5.1B active parameters per token built for reasoning and agents, per Modelfit specs as of November 2026.
If you want to sell forecasting to a CFO, don't quote the error rate until an auditor has broken your assumptions.
