The Omega Block

Correlation-network features that improve systemic tail-loss attribution beyond downside beta and conditional MES.

Companion repository for the preprint (PDF in paper/; SSRN: https://ssrn.com/abstract=7363442). Five per-asset features computed from daily prices alone; in pre-registered paired ablations with a noise-placebo floor, the block adds 1.3 to 3.4 points of out-of-fold R2 to next-year tail-loss attribution over baselines up to a 31-column replica of a vendor feature set, replicated on two universes and two nearly disjoint periods, while the placebo worsens the same model everywhere.

Try it in two minutes

Runs the full acceptance protocol on a public price panel, in your browser, with no setup. Swap in your own prices and rerun; nothing leaves the notebook.

Run the acceptance test on YOUR data

python validate_omega_block.py --prices your_prices.csv \
       [--events your_fomc_dates.csv] [--receipt result.json]

prices.csv: wide CSV, first column Date, one column per asset. The harness reruns the full protocol (5 asset folds, paired ablation, same-width noise placebo, block bootstrap) and prints an explicit PASS or FAIL. Without an event calendar, only the static features are tested.

Signal-trial mode (no formulas needed)

python validate_omega_block.py --prices your_prices.csv \
       --features omega_features_465.csv

Tests the block as a plain numbers file: nothing but values enters your environment. omega_features_465.csv covers 465 S&P constituents, 2016-2025. Verified blind: +0.036 vs a -0.016 placebo (cert_demo.json).

Evaluating against a baseline you cannot see

PROTOCOL.md states the evaluation method separately from this case: the run happens in your environment, your baseline replaces ours, the acceptance rule is fixed before the run, a same-width noise placebo sets the floor, and the receipt carries only aggregates. It also states what the protocol does not solve.

Replication registry

See REGISTRY.md. PASS and FAIL results are equally welcome; use --receipt to produce an aggregates-only, shareable JSON (nothing is ever transmitted automatically).

Contents

• validate_omega_block.py acceptance harness (features mode, receipt)
• omega_features_465.csv signal-trial feature file (SHA-256 committed)
• REGISTRY.md public replication registry
• paper/ the preprint PDF
• prereg/ written pre-registrations, before each run (working documents, in Spanish)
• results/ result logs for every number in the paper
• figures/ the paper's figures and their generators

License and patent notice

Code is released under the MIT License. The quantities Q1-Q5 and their use as inputs to risk-attribution models are covered by USPTO provisional applications 64/112,912, 64/121,656 and 64/140,524 (Omega-S family). Research and evaluation use is encouraged; commercial use in third-party risk products requires a license (acedo@biomemakers.com).