# COMPUTE_BUDGET.md

| Resource | Plan | Measured / status |
|---|---|---|
| Compute node | M3U96a (Mac Studio M3 Ultra, 28 cores, 96 GB RAM) — FinData API on localhost | load 1.5–1.9 before start |
| Scratch | `~/wp39-cache` on the internal SSD (319 GB free; FinData hard floor 60 GB) | streaming design keeps < 5 GB |
| API exports | ≈ 2 × 4,300 symbol-chunks, ≈ 1 s each server-side, 4 concurrent | pilot measured 1.0 s / 860 k rows |
| Fundamentals calls | 5,451 × 2 | ms each |
| Feature computation | Polars per symbol (≈ 1.7 M rows → 4 frequency tables), 8 worker processes | to be measured in the pilot |
| Models | logistic ridge (seconds) ; LightGBM 24 trials × 5 information sets × (1 chronological + 11 walk-forward folds) ≈ 1,440 fits, early stopping; ≈ 2–4 h on 28 cores | — |
| Bootstrap | 1,000 stationary-bootstrap draws over 23 test quarters per contrast — seconds | — |
| LaTeX | laptop (TeX Live, `latexmk`) | available |
| AI usage | none in the pipeline (no LLM features, no neural B7/B8); Claude Code used only as the engineer | owner instruction 2026-11-02 |

Pilot (Phase 2 gate): 300 firms stratified by first-fundamentals year and size, full history → validates throughput,
feature sanity and label base rates before the full run.