Skip to content

Risk · Market Regime & Stress Testing

Two complementary research aids sit on top of the covariance-based risk model (Σ = B F Bᵀ + D). The risk model tells you the distribution of outcomes; regime and stress pin a single interpretation or adverse path.

Market Regime — alphaforge.risk.regime

Each trading day is labelled along two independent axes inferred only from trailing information (no future data leaks):

  • trend — Bull / Bear: is the cumulative index above its trailing MA?
  • volatility — High / Low: is rolling annualised vol above its own median?

The four combined labels are Bull/LowVol, Bull/HighVol, Bear/LowVol, Bear/HighVol.

from alphaforge.risk.regime import classify_regime, regime_statistics

regime = classify_regime(market_returns)          # pd.Series of str labels
stats  = regime_statistics(strategy_returns, regime)   # per-regime ann return/vol/Sharpe
  • classify_regime degrades to empty (no labels, no crash) when there is not enough history to decide.
  • factor_performance_by_regime(ic_series, regime) reports Rank-IC diagnostics per regime so you can see which factors work in which environment.

Stress Testing — alphaforge.risk.stress

Answers "what happens to the portfolio if factor X moves by a given amount?" Every scenario is a shock to a risk-model factor, so the result is internally consistent with the risk decomposition:

from alphaforge.risk.stress import run_scenarios, DEFAULT_SCENARIOS

book = run_scenarios(weights, risk_result)   # dict[str, StressResult]
worst = min(book.values(), key=lambda r: r.pnl_pct)

Default scenarios include market_drawdown_10pct, momentum_crash_2sigma, value_selloff_5pct, volatility_spike_3sigma, quality_rotation_5pct, liquidity_dryup_2sigma. Two shock kinds are supported:

  • factor — a fixed move, e.g. market → -10%.
  • factor_sigma — k · σ of the factor's standalone volatility.

sector_shock(weights, industry, sector, shock_pct) applies a direct, asset-level shock to one GICS sector when no factor model is needed.

Both modules are fully deterministic and reproducible; the copilot surfaces them through the regime and stress tools.