blackswan 0.4.0

Stress-test mathematical and financial logic in Python code — find the exact line where models break.

blackswan

A stress-testing engine for financial and mathematical Python code.

BlackSwan finds the exact source line where your model breaks under extreme conditions — before production does. It is a debugger for numerical fragility, not a linter, not a simulator.

---

Installation

pip install blackswan

Requires Python 3.11+ and NumPy 1.26+.

---

Quickstart

# Stress-test a function with the liquidity crash scenario
python -m blackswan test models/risk.py --scenario liquidity_crash

{
  "status": "failures_detected",
  "runtime_ms": 2840,
  "iterations_completed": 5000,
  "summary": {
    "total_failures": 847,
    "failure_rate": 0.1694,
    "unique_failure_types": 1
  },
  "shatter_points": [
    {
      "line": 36,
      "severity": "critical",
      "failure_type": "non_psd_matrix",
      "message": "Covariance matrix loses positive semi-definiteness when pairwise correlation exceeds 0.91. Smallest eigenvalue: -0.0034.",
      "frequency": "847 / 5000 iterations (16.9%)",
      "causal_chain": [
        { "line": 8,  "variable": "correlation", "role": "root_input" },
        { "line": 31, "variable": "corr_matrix",  "role": "intermediate" },
        { "line": 36, "variable": "cov_matrix",   "role": "failure_site" }
      ],
      "fix_hint": "Apply nearest-PSD correction (Higham 2002) after correlation perturbation, or clamp eigenvalues to epsilon."
    }
  ]
}

---

CLI Reference

test — stress-test a model

python -m blackswan test <file> [options]

Options:
  --scenario     Preset scenario name (required)
  --function     Target function name (auto-detected if omitted)
  --iterations   Number of Monte Carlo iterations (default: 5000)
  --seed         Random seed for reproducibility (default: 42)
  --adversarial  Use genetic algorithm to search for worst-case inputs
  --population   GA population size (default: 100, requires --adversarial)

Exit codes: 0 = no failures, 1 = failures detected, 2 = engine error.

fix — generate a deterministic guard for a failure line

python -m blackswan fix <file> --line N --type TYPE

Requires pip install blackswan[fixer] (adds libcst).

--type	Guard applied
division_instability	max(denominator, 1e-10) epsilon clamp
non_psd_matrix	Nearest-PSD correction via np.linalg.eigh + np.maximum
ill_conditioned_matrix	Conditional np.linalg.pinv fallback when cond > 1e12
nan_inf	np.nan_to_num(result, posinf=…, neginf=…) guard

Output JSON:

{
  "status": "ok",
  "line": 36,
  "original": "    cov_matrix = np.cov(returns.T)",
  "replacement": "    cov_matrix = np.cov(returns.T)",
  "extra_lines": [
    "    vals, vecs = np.linalg.eigh(cov_matrix)",
    "    cov_matrix = vecs @ np.diag(np.maximum(vals, 1e-10)) @ vecs.T"
  ],
  "explanation": "Clamps negative eigenvalues to epsilon, restoring PSD property."
}

status is "ok", "error", or "unsupported". Exit code is always 0 (the caller decides what to do with the result).

Available Scenarios

Name	Description
liquidity_crash	Spread widening, vol expansion, correlation stress, turnover collapse
vol_spike	2–4× volatility multiplier, mild correlation increase
correlation_breakdown	Pairwise correlation shift +0.20–+0.50
rate_shock	+100–+300 bps interest rate shock with spread widening
missing_data	Random NaN injection and partial time series truncation

---

Python API

from blackswan.engine.runner import StressRunner
from blackswan.scenarios.registry import load_scenario

scenario = load_scenario("liquidity_crash")
runner = StressRunner(scenario)

result = runner.run(
    calculate_portfolio_var,
    base_inputs={"weights": weights, "vol": vols, "correlation": 0.0},
)

print(f"Failures: {len(result.findings)}")
for finding in result.findings:
    print(f"  Line {finding.line} [{finding.severity}]: {finding.message}")

Adversarial API

from blackswan.engine.adversarial import EvolutionaryStressRunner

runner = EvolutionaryStressRunner(
    scenario,
    n_generations=20,
    population_size=100,
    elite_fraction=0.2,
)
result = runner.run(fn, base_inputs)

Custom Detectors

from blackswan.detectors.numerical import LogicalInvariantDetector

# Weights must always sum to 1 (+/-1e-6)
invariant = LogicalInvariantDetector(
    assertion=lambda inputs, output: abs(output.sum() - 1.0) < 1e-6,
    name="weights_sum_to_one",
)
runner = StressRunner(scenario, extra_detectors=[invariant])

---

Failure Detectors

Detector	Catches	Always active?
NaNInfDetector	NaN or Inf in any output	Yes
DivisionStabilityDetector	Denominator approaching zero	Yes
MatrixPSDDetector	Covariance/correlation matrix non-PSD	Auto (on matrix code)
ConditionNumberDetector	Ill-conditioned matrices (cond > 1e12)	Auto (on linalg.inv)
BoundsDetector	Outputs outside configurable plausible ranges	Auto
ExplodingGradientDetector	Output growth > 100x input perturbation	Auto
RegimeShiftDetector	Structural breaks in output distribution	Auto
LogicalInvariantDetector	User-defined assertions	On demand

Detectors are auto-tagged to relevant source lines via AST analysis.

---

Custom Scenarios

Create a YAML file and pass its path as --scenario:

name: my_scenario
description: "Custom stress test for credit portfolio"
perturbations:
  - target: spread
    type: multiplicative
    distribution: lognormal
    mu: 2.0
    sigma: 0.4
    constraints:
      min: 1.0
      max: 5.0
  - target: correlation
    type: additive
    distribution: uniform
    low: 0.15
    high: 0.40
global_constraints:
  - target: correlation
    min_value: -1.0
    max_value: 0.95
defaults:
  iterations: 5000
  seed: 42

python -m blackswan test models/credit.py --scenario path/to/my_scenario.yaml

---

What BlackSwan Supports

Works well on:

• Pure functions with NumPy / Pandas inputs and outputs
• Explicit variable assignments
• NumPy array operations, linalg, random
• Pandas DataFrame column operations
• Single-file scripts and focused modules

Out of scope for V1:

• Deeply nested class hierarchies with side effects
• Config-driven logic loading parameters from databases
• Multi-threaded or async computation pipelines
• C extensions or Cython modules

---

Links

• GitHub
• Architecture
• Scenarios
• Changelog
• Bug Tracker

---

License

MIT