NUMERICAL AUDIT · 22 OCTOBER 2026
Empyrical: constant returns and floating-point Sharpe ratios
Exactly constant returns can produce huge finite ratios instead of the documented signed infinity. A focused fix passes all 860 package tests.
Xamit Kadirbekov · AI-assisted investigation and writing
Reproduced; PR open; not accepted
Upstream report #124 · Reproduction archive · Github Catalog · Huggingface · Linkedin · Youtube
Empyrical: constant returns and floating-point Sharpe ratios
22 October 2026 · GERO Numerical Assurance · Xamit Kadirbekov
REPRODUCED IN THE STATED SCOPE; FIX PROPOSED. The upstream PR is open. Acceptance and a corrected release are not claimed.
Exactly constant returns can produce huge finite ratios instead of the documented signed infinity. A focused fix passes all 860 package tests.
Upstream submission: https://github.com/stefan-jansen/empyrical-reloaded/pull/124
Closes #123. Related contract: #80/#81.
For identical nonzero returns, floating-point centering can produce a tiny nonzero sample standard deviation. For example, 252 copies of 0.001 currently produce an excess Sharpe around 1.53e15 and Sharpe around 2.44e16, rather than the intended positive infinity.
Normalize the computed deviation to zero only when the observed values are exactly equal and the deviation is finite. This preserves undefined NaN deviations, does not threshold near-constant data, and keeps the existing annualization and signed-infinity convention. Apply the helper to both Sharpe functions.
The 13 public-API regression cases cover several lengths and positive/negative constants, plus constant and varying DataFrame columns together. The existing insufficient-observation tests are unchanged.
Validation (Python 3.14, NumPy 2.5.4, pandas 3.0.6):
- Base
a54620df2077f56e8ef126d9b230b2ec2e1fadf6with new tests:PYTHONPATH=src python -m pytest tests/test_zero_dispersion_ratios.py -q→ 8 failed, 5 passed. - Fix
6d255c8:PYTHONPATH=src python -m pytest tests/ -q→ 860 passed. - Black 24.4.2 and flake8 on the changed source and new test file: pass.
All runs used one local CPU worker. AI assistance: Codex reviewed the candidate fix, repeated the actual public-API regressions and complete suite, and prepared this patch description. Independent human review is not claimed; maintainer review is requested.
Scope and attribution
This is an AI-assisted numerical correctness investigation using synthetic inputs and, where stated, the upstream test fixtures. There is no evidence here of executed trades, client losses, a security exploit, or effects in a commercial deployment. Independent human review is pending. This report is CC BY 4.0; upstream code retains its original license.