FinanceX Portfolio Manager
Market
Ask · No manager
+ Attach news
Add a specific article URL or paste article text. Attached news runs in Deep mode and becomes a cited evidence row.
0/3 attached
StockClick to open · / to focus
Benchmarked against published research
FinQA (Chen et al. 2021) · FinanceBench (Islam et al. 2023) · N=50 questions / 100 answers per row| Benchmark | Accuracy | vs. published |
|---|---|---|
| FinQA (open-book) | 75.0% | Human 91.2% · FinQANet 61.2% · GPT-4 63–78% |
| FinanceBench (open-book) | 70.0% | GPT-4 + oracle document 50–85% |
| FinQA (closed-book) | 0.0% | expected — no filing access, exact line-items aren’t memorizable |
| FinanceBench (closed-book) | 15.0% | GPT-4 closed-book 9–19% (in range) |
“Open-book” means the model is given the source filing directly — the same condition the published GPT-4/FinQANet numbers above use. This measures answer and trust-verification quality against a filing in hand, not FinanceX’s own ability to find the right filing unassisted, which is a separate, harder capability. Full methodology and raw results: benchmark/FINDINGS.md.