Walk-Forward OOS Certification

The deterministic out-of-sample engine. Walk-forward, validation mode, anchored, 5 folds, 2022→today — the OOS folds are the verdict, not a single backtest.

The study call

mcp__nexustrade__run_walk_forward_study (tools are deferred — load via ToolSearch). Always run preview_only: true FIRST to see the fold calendar + cost, confirm it spans 2022→today (must include both the 2022 bear and the April-2025 selloff — this book's deepest drawdown is an April-2025 event, so the folds have to span it), then flip to false and record the study_id.

{
  "portfolio_id": "<SUBJECT or a clean clone>",
  "global_start_date": "2022-01-01",
  "global_end_date":   "<TODAY>",
  "fold_count": 5,
  "walk_forward_mode": "anchored",
  "mode": "validation",              // per-fold independent OOS
  "oos_width_days": 252,
  "embargo_days": 14,
  "interval": "Day",
  "inner_mode": "backtest_only",     // certify the FIXED book — no per-fold optimizer
  "preview_only": true               // flip to false to actually run
}

• Certify a fixed book: inner_mode: "backtest_only" (engine_kind ga/default — a sweep is rejected for a fixed-config cert).
• Re-optimize (only on FAIL or a structural change): engine_kind: "sweep", inner_mode: "optimize", certification: true, plus gene_intents. GA overfits and is rejected for deploy certification — sweep is the certified path. See sweep-reoptimization.
• certification: true applies the activity-floor + percentChange ≥ 0 + sortino ≥ 0.5 policy and defaults validation_percent to 50.

Pre-flight before certifying a live book

1. get_portfolio — capture the strategy set, conditionFieldAudit, and the exit ladder.
2. Strip the LaunchAgent. A LaunchAgent strategy can't run historically — it fires once and logs thousands of cooldownSkip events, contaminating folds. Certify a LaunchAgent-free clone (rebalance + close rules only) and say you did so.
3. Spread-shape check the live structures (see options-structure-rules). A long-dated vertical spread is a hard violation → the book cannot certify as-is; go straight to re-opt.
4. Convexity-cap footgun: if multiple "always" take-profit closes exist (e.g. P/L ≥ 20% / 50% / 200%), the lowest binds first and caps every winner at ~+20%, defeating the outright-call rationale. Flag it — it's the most likely reason returns trail a let-winners-run book.

Monitor the study

Poll list_walk_forward_studies / get_walk_forward_study_results <study_id> until terminal (COMPLETE / ERROR / CANCELLED). Parse with jq — payloads are large, don't dump them. Watch for known failure modes and surface them (don't paper over — that's a bug-protocol trigger):

• Silent-fail / ERROR — no fold stats / empty validationAggregate. Report it; don't invent a verdict.
• Wrong window — fold calendar not starting 2022-01-01 / skipping the bear.
• Backtest contamination — LaunchAgent firing in folds (should have been stripped in pre-flight).
• Zero-trade / NaN folds — upstream position-marking corruption; report which folds, never average over NaNs.
• A study genuinely hung past its estimated cost/time → stop, summarize, wait for the human (the engine may be under repair in another tab).

The mandatory per-fold checks (these ARE the verdict)

1. Per-fold OOS table — for each fold: train window, OOS window, OOS return, OOS maxDD, OOS Sortino, median deployment, distinct underlyings, participation. Plus the validationAggregate.
2. Independent re-backtest spot-check — backtest_portfolio ($25k) on the full cycle (2022-01-01→today), the 2022 bear (2022-01-01→2023-01-01), and the last 12 months; audit_backtest_posture the full-cycle run. Study folds must agree with a fresh standalone backtest. Baseline ≠ SPY for an options book — use a material underlying or equal-weight universe B&H, and note it.
3. Degradation — train vs OOS per fold. Large train→OOS collapse = overfit, even if the full-cycle backtest looks great.
4. Drawdown honesty — OOS maxDD next to OOS return per fold; lead with risk. At this leverage a fold can post a huge return and a 50–77% drawdown.
5. Breadth gate at FIXED $25k (breadth-audit).
6. Reproducibility / field audit — conditionFieldAudit matches intended knobs; compare_backtests {tolerance_bps: 0} on a re-run. Verify by field, never display name.
7. Spread-shape compliance (options-structure-rules) — hard reject on violation.

Calendar alignment (variant vs base)

The walk-forward engine clips the fold calendar to the indicator's / data's coverage, so a variant's study can silently get different folds than the base's. Always run a base CONTROL with the same global_end_date as the variant's clipped calendar and compare fold-for-fold. Report per-fold signal coverage next to per-fold OOS. A variant that only wins where the signal exists is a finding, not a cheat — but say it.