Ml4t Stress Test

Stress Testing

A strategy backtested on 2015-2023 has never seen a regime where equities and bonds fall simultaneously. Without stress testing against 2008, 2020, and 2022, you are implicitly betting that those regimes will not recur.

The Problem

Backtests cover only the historical sample, which may exclude the scenarios most relevant to survival. A momentum strategy backtested from 2010 onward has never experienced the 2009 momentum crash (-46% in one month). Stress testing applies known crisis scenarios and hypothetical shocks to the current portfolio, revealing exposures that summary statistics hide. This is not optional - it is how you discover that your "diversified" portfolio has a hidden correlation spike that produces a -30% month.

The Pattern

WRONG

import numpy as np

# Only looks at backtest period (2016-2023) - misses major crises
returns = backtest_returns  # 2016-2023 daily
max_loss = returns.min()
print(f"Worst day: {max_loss:.1%}")  # -3.2%, looks safe
# But GFC 2008 would have been -15% in a single week for this portfolio

CORRECT

import numpy as np

# Define crisis scenarios as asset-class shocks
SCENARIOS = {
    "GFC 2008":       {"equity": -0.50, "bond": +0.15, "credit": -0.30, "vol": +3.0},
    "COVID Mar 2020":  {"equity": -0.34, "bond": +0.08, "credit": -0.15, "vol": +4.0},
    "Rate Shock 2022": {"equity": -0.20, "bond": -0.15, "credit": -0.10, "vol": +1.5},
    "Correlation Spike":{"equity": -0.25, "bond": -0.10, "credit": -0.20, "vol": +2.0},
}

# Apply each scenario to current portfolio weights
weights = np.array([0.40, 0.30, 0.20, 0.10])  # equity, bond, credit, vol
asset_classes = ["equity", "bond", "credit", "vol"]

for name, shocks in SCENARIOS.items():
    pnl = sum(weights[i] * shocks.get(ac, 0) for i, ac in enumerate(asset_classes))
    survives = pnl > -0.20  # survival threshold
    print(f"{name:25s} PnL: {pnl:+.1%}  {'OK' if survives else 'BREACH'}")

Factor Stress Testing

def factor_stress(weights, factor_betas, factor_shocks):
    """Stress via factor exposures rather than asset classes.

    factor_betas: (n_assets, n_factors) from regression
    factor_shocks: dict of factor_name -> shock magnitude
    """
    shocks = np.array([factor_shocks[f] for f in factor_names])
    asset_impacts = factor_betas @ shocks
    return weights @ asset_impacts

# Example: what if momentum factor drops 3 sigma?
loss = factor_stress(weights, betas, {"momentum": -0.15, "value": 0.05})

Hypothetical Scenarios to Always Include

Guardrails

• Historical scenarios are a floor, not a ceiling - always include a "2x worst" hypothetical
• Correlations increase under stress - use stressed correlations, not normal-regime estimates
• Test at current positions, not average or target weights
• Update scenario library when new crises occur (each one reveals a new failure mode)

Checklist

• At least 3 historical crisis scenarios applied (2008, 2020, 2022)
• At least 1 hypothetical scenario (correlation spike or liquidity freeze)
• Survival threshold defined (e.g., max -20% in any scenario)
• Factor exposures stress-tested (not just asset-class proxies)
• Scenario library reviewed and updated within last 12 months