hidalgo

Fast Rust economic-complexity kernels (ECI, PCI, product proximity, density) with Python bindings. Reproduces the Hidalgo-Hausmann eigenvalue method used by TradeWeave, verified against its published rankings.

Use

import numpy as np, hidalgo
out = hidalgo.bundle_from_rca(rca_matrix, threshold=1.0)
out["eci"], out["pci"], out["proximity"], out["density"]
out["kept_countries"], out["kept_products"]  # indices mapping back to your rows/cols

bundle_from_exports(exports, ...) computes RCA for you first; bundle_from_m(m, ...) takes a pre-binarized matrix.

Build

uv venv && source .venv/bin/activate
uv pip install numpy
maturin develop --release

Performance

Benchmarked on the live TradeWeave RCA data (year 2024, 226 countries x 6429 products, Apple Silicon M5). The NumPy path (two np.linalg.eig calls on the reflections matrices plus proximity) takes 30732 ms; hidalgo completes ECI, PCI, proximity, and density in 242 ms, a 126.9x speedup. The gain comes from skipping np.linalg.eig, which computes the full spectrum of the roughly 6000x6000 product reflections matrix just to extract one eigenvector. hidalgo uses matrix-free deflated power iteration instead, which needs only matrix-vector products and converges in tens of iterations. Reproduce with: python tests/bench_compare.py (set HIDALGO_TRADEWEAVE_DATA if your parquet tree is not at ~/tradeweave/data/parquet).

Correctness

Verified by pytest tests/:

• Algorithm proof: hidalgo agrees with NumPy np.linalg.eig on the same matrix (Spearman = 1.0 for ECI, 0.99999999 for PCI).
• Deployed parity: reproduces TradeWeave's published eci_rankings / pci_rankings order on the HS92 core catalog (Spearman > 0.99999), and proximity matches product_proximity to ~5e-5.

Provenance note: pci_rankings.parquet is a hybrid. The HS92 core (the HS6 codes in products.parquet) is scored by the eigenvector method (what hidalgo implements). The extended HS96-HS22 products are scored by a different formula in TradeWeave's data/compute_ext_rankings.py ("PCI = mean ECI of exporters with RCA >= 1"), so parity is checked on the core universe only.

Layout

• crates/hidalgo-core pure-Rust math (only rayon), matrix-free second eigenvector via shifted deflated power iteration.
• crates/hidalgo-py PyO3 bindings, built with maturin.