Crisis and Stress Scenarios

These reweight the resampler toward a labelled historical window rather than injecting a synthetic shock. The mechanism is a fivefold weighting applied at each eligible block restart, so a path is more likely to begin drawing from the stressed window but the days it draws are still real observed days with their real cross-asset structure intact.

Both replay arms need the block_bootstrap process. That process is retired and no plan grants it, so you cannot request it. The engine skips a replay arm on any other process and reports the reason with the result. One route still reaches these arms: a panel too short for the default generator falls back to block_bootstrap, with the fallback disclosed on the run.

A crisis window must contribute at least ten trading days of the frozen history before the scenario is considered runnable. A window your data cannot support, such as COVID when the calibration history starts after 2020, is reported in skipped_scenarios with the reason rather than silently dropped.

The weighting is a versioned non-probabilistic intervention rule. It is fixed by product decision and is deliberately not derived from the observed crisis arrival rate described below.

Four opt-in arms cover risks that ordinary return resampling cannot express cleanly, because they are mechanical events rather than draws from a return distribution.

The slow variance layer of the regime processes has a median fitted half-life of about 414 sessions, so a simulated path reaches a stressed level over months. Across 300,000 measured paths it produced no entry to three times volatility by session 5. The 2008 and COVID episodes crossed three times within 4 and 9 sessions of their last calm window. Since mc-v1.24.0, this opt-in scenario applies a deterministic fast-entry state, fhs-stress-fast-entry-v1, to the conditional variance of every holding:

Here counts simulated sessions from zero. Every path opens at three times its conditional volatility, and the excess log variance halves every 21 sessions: the volatility multiplier is about 2.54 at session 5 and 1.73 at session 21. The state adds to the slow layer in logs where a process has one, inside the same bound, and the conditional volatility clamp still applies. It stays out of the GJR recursion, so the declared half-life is its whole persistence.

• The state uses no random numbers, so the scenario arm keeps the base arm's common random numbers and the paired delta measures the variance path alone.
• The engine accepts only points from a grid declared before any result was read: an initial multiplier of 2, 3 or 4 and a half-life of 10, 21 or 63 sessions. The preset uses 3 and 21.
• It applies to fhs_gjr_garch_vt, fhs_gjr_garch, fhs_ewma, fhs_gjr_regime and fhs_gjr_regime_uv_fallback. On gbm_lognormal and the simple-return processes it is reported in skipped_scenarios with the reason, because those processes have no variance state to raise.
• The artifact records the mechanism, its version and its parameters in scenario_assumptions.fast_entry_stress, with event_probability: null.

It is a stress severity, not an event probability and not a calibrated forecast. The rolling-origin harness has not validated it. A run that selects no stress scenario does not change. The API and the MCP submit_monte_carlo tool accept it.

Every run carries the digest-verified BSE SENSEX episode registry that sits behind its named stress scenarios. The frozen window runs from 1997-07-02 through 2026-07-28, contains 7,160 returns, and spans 29.0711 calendar years.

Each episode publishes its start, trough and recovery dates, depth and duration. The Poisson interval is included specifically to show how little eight events identify, even under the generous assumption that crises arrive at a constant rate:

The upper bound is more than four and a half times the lower one. That width is the reason the rate is not converted into an annual crisis probability, does not enter the generator, and does not alter scenario sampling. The observed rate and empirical severity describe a frozen broad-market sample. They are not a statement about the probability, timing or severity of a future portfolio crisis.

A missing or damaged snapshot fails closed to the same non-probabilistic scenario contract rather than to a softer one.