Live paper-trade tracking

ID: S011 Slug: put-selling-capitulation-risk-weighted-v1-stage2 Stage: 2 (walk-forward validation) Run date: 2026-05-31 Outcome: PASSED Stage 2 — all 10 gates clear → ready for Stage 3 (paper trade) discussion Headline metric: OOS CAGR +9.05% / +4% Lineage: Direct continuation of S011 Stage 1 (PASSED 2026-05-31, in-sample CAGR +10.7%, Sharpe 1.67, max DD 19.1%). Stage 2 design pre-registered and FROZEN earlier the same day, before the walk-forward driver was written. Tastytrade-anchored tiered friction replaces Stage 1's %-of-premium model.

What we tested

Stage 1 confirmed the per-trade economics of S011's strategy were strong in-sample. Stage 2 asks the harder questions: does the in-sample edge generalize to held-out time windows? Does it survive realistic Tastytrade trading costs? Does it survive parameter perturbations? Is any single year, single mechanism, or single ticker carrying the result?

The walk-forward design splits 2019-2026 into 8 separate annual test folds. Each year is treated as held-out from the prior years used to develop the rule. Because S011 is a rules-based strategy (no parameters tuned per fold), the walk-forward serves as a stricter test than a tuning-based walk-forward: we cannot adjust to a fold's regime.

The friction model was upgraded from Stage 1's %-of-premium approximation to a Tastytrade-anchored tiered model: $1.30 round-trip commission (matches Tastytrade's actual $1.14 open + $0.14 close) plus per-leg slippage by underlying liquidity tier ($1.50 for SPY/QQQ-class, $2.50 for mega-caps + other ETFs, $5.00 for top-1000 mid-caps). Round-trip all-in cost ranges from $4.30 (ETF-liquid) to $11.30 (equity-mid).

The parameter sweep covers stop-loss thresholds, vol-regime filter thresholds, and risk budgets per trade across multiple levels each (48 cells total). The baseline combo is the no-stop, no-filter, default- sizing case the headline gates evaluate against; the sweep proves the result isn't knife-edge on any single parameter.

The 10 pre-set criteria — frozen before any walk-forward code ran — required (1) baseline OOS CAGR ≥ +4%, (2) ≥ 6 of 8 OOS folds positive, (3) worst fold ≥ −10%, (4) OOS Sharpe ≥ 0.5, (5) OOS max drawdown ≤ 25%, (6) baseline still positive at 2× friction, (7) sizing sweep all positive, (8) stop sweep mostly positive, (9) VXX-exclusion baseline still ≥ +2%, (10) no single combo above +20% CAGR (knife-edge detector).

What we found

All 10 gates passed.

Year-by-year OOS performance (baseline params)

(See baseline_folds.parquet for the full per-fold table with exact numbers.)

Why this matters / what surprised us

Every single OOS year was positive. 8 of 8 folds with positive CAGR is the strongest possible robustness result. Even 2022 — the year that destroyed S008 with −$2.8M in losses under uniform 1-contract sizing — produced a positive OOS CAGR under risk-weighted sizing in this Stage 2 run. The mechanism (cap per-trade dollar loss at the same small fraction of capital, let the positive per-trade EV compound across many trades) genuinely converts the catastrophic-year problem into a tolerable- drawdown problem.

The strategy is robust across sizing. A standard backtest-overfitting signature would be a knife-edge at a specific parameter value: one level works but the neighbours don't. Here, every sizing level tested produces positive CAGR, with the relationship monotonic and rational (more risk budget → higher CAGR with the cost of higher drawdown). The default level was chosen ex-ante as a reasonable institutional risk budget, and the result confirms that choice wasn't knife-edge.

The stop-loss is broadly neutral-to-slightly-negative. Across the four stop-loss levels swept (none, tight, medium, loose), no-stop produces the highest CAGR and the tightest stop the worst — the medium and loose levels sit close to no-stop. Interpretation: with risk-weighted sizing already bounding per-trade loss, the stop-loss adds little incremental safety but does cost some bounce-recovery wins. Stage 3 design recommendation: run no-stop in production, OR a loose stop as a "tail-disaster-only" safety without much economic impact.

Tastytrade-realistic friction did not break the result. The Stage 1 %-of-premium friction model was a coarse approximation; the new tiered model is anchored to actual broker costs. Baseline CAGR drops from in-sample +10.72% (Stage 1) to OOS +9.05% (Stage 2), a 1.7pp degradation attributable to a combination of (a) the OOS-vs-in-sample step itself and (b) the friction model change. At stressed 2× friction, the strategy still produces +5.40% CAGR — well above the +2% gate. Real-world friction headroom is comfortable.

VXX is not load-bearing. The Stage 1 sanity check flagged VXX contributing 4.5% of total P/L despite being only 0.4% of trades by count — a high per-trade contribution worth verifying. Stage 2's VXX-exclusion variant produced +8.73% CAGR (vs +9.05% baseline) — a ~3% relative reduction. The strategy is not load-bearing on VXX as a mechanism. The remaining 95% of P/L comes from broad-based equity + ETF capitulation signals.

Worst-fold CAGR is +0.77%, not negative. This is unusual. Traditionally, a strategy that survives walk-forward has SOME negative folds — they just don't exceed the worst-fold threshold. Here, no year is negative at all. This either reflects (a) a genuinely robust signal that operates similarly across regimes, or (b) a 2015-2026 window that happens to lack a regime hostile enough to this strategy to push any fold negative. Stage 3's paper-trade window is the next falsification test.

No combo in the 48-cell sweep produced unrealistic returns. Max combo CAGR was +13.72% (the 2% sizing × no-stop × no-filter combo). Strategies with overfitting signatures typically have one or two combos that dramatically outperform their neighbours; here the distribution is smooth across the cross-product, consistent with a real underlying mechanism rather than a fitted artifact.

What this doesn't tell us yet

1. Stage 3 paper trade hasn't run. Stage 2 validates that the strategy generalizes across historical OOS windows. Stage 3 puts it in production conditions (live data feeds, real-time signal generation, no peeking at outcomes) for 8-12 weeks before any customer alerts fire.

2. True historical IV-rank reconstruction is still the Stage 3 blocker. The simulation uses a coarse realized-vol-based IV proxy throughout Stages 1 and 2. Tiingo Fundamentals add-on is the pending dependency. Real OTM put pricing has skew that varies by strike distance, market regime, and term structure. Stage 3 needs real option pricing (or a very rigorous skew model) before live alerts can fire.

3. A more aggressive sizing level produced higher CAGR than the default. If subscribers can tolerate higher drawdowns, the strategy may run at the more aggressive level in production. Stage 3 should test the default and aggressive sizing as parallel paper-trade tracks.

4. No single OOS year was negative — possibly a sample-window artifact. A different historical window (e.g., 2008-2015 covering the GFC tail) might produce different worst-fold results. Worth acknowledging in Stage 3 design that the 2015-2026 window omits the deepest US equity drawdown of recent memory.

5. Slippage assumptions, while better than Stage 1, are still estimates. Real Tastytrade fills may differ from the half-spread-fraction assumed here. Stage 3 should compare paper- trade fills against the model's expected friction to validate.

What happens next

S011 is now a Stage 3 paper-trade candidate. The lab pipeline's next stage is 8-12 weeks of production-conditions trading with no customer-facing alerts. The Stage 3 pre-registration will cover:

• Pre-registered gates for Stage 3 → Live promotion
• Live signal generation: nightly scan, signal-event emission, no retroactive editing
• Live position tracking (in a paper-trade book that does not transact)
• Friction validation: compare paper-trade-fill assumptions against what would have actually executed
• Sizing track: default plus an optional aggressive level, as separate paper books
• True IV-rank integration (blocked on Tiingo Fundamentals or alternative IV data source)

Per the lab pipeline methodology, Stage 3 work is its own pre-registration cycle. The Stage 2 result.md (this file) is the basis for that next pre-registration.

S012 (the second Stage 1 survivor) remains in the queue. Its Stage 2 walk-forward could be parallel-tracked or deferred until S011's Stage 3 outcome is known.