Drafts & Failed Studies — Milkman Trades

drafts & failed.

studies that didn’t make the core playbook

These pages live outside the core playbook for a reason. Every study here was reviewed against the same two filters used for the keepers: quality (sample size, edge magnitude, caveats acknowledged) and actionability (entry/exit clarity, real-time identifiability, fit for 0DTE/short-dated SPX/SPY).

Listing them transparently — rather than burying them — keeps the catalog honest. The point of a research site is to publish what worked and what didn’t.

Draft Studies — In Progress 11 studies

Real signal but a weak frame. These need tweaks — sample-size deepening, reframed metrics, rebuilt code, or merging into adjacent studies — before an edge can be confirmed. Treat as research drafts, not playbook material.

SPX vs SPY Morning Divergence New →

Exploratory first cut. Median |PO_SPY − PO_SPX| at 09:30 is ~40 oscillator units (about two full Saty zones); fast cloud disagrees on 38% of days. Mechanism: SPY’s pre-market bars cushion its EMAs; SPX’s RTH-only EMAs absorb the gap all at once at 09:30. Across 4,527 RTH days (2008–2026), SPY’s reading is the leading signal — |gap| ≥ 1% cohorts produce a 0.54% full-day spread between symmetric sign(d_open) buckets, and the fast-cloud-mismatch cuts deliver a 0.42% spread. Needs VIX/regime stratification, FOMC/OpEx exclusion, and a 2010+ re-cut before it becomes tradable.

ATR Level Cascade New →

First-pass map. 33,153 RTH first-touch events across 25y SPY tagged with continuation / retrace / last and time-of-day bucket. Solid construction (3-min bars, no look-ahead) and clear time-of-day shape — 76–78% continuation pre-10:00 collapsing to last-dominated post-15:00. Needs regime-stratification (high/low ATR days, gap days, FOMC) and a paired entry/exit framework before it earns a tradable slot. Use the explorer as a path-handicapper for now, or the static Saty-style chart for the same-session probability map.

Bilbo Golden Gate Restored 2026-07-11 →

Look-ahead bias found, fixed, and rebuilt. The 1h PO was sampled from the not-yet-closed hour bar (up to 60 min of forward leakage). The join was fixed 2026-04-26 and the page now carries the corrected rerun: bear Low+Falling 90.3% → 85.1%, bull High+Rising 77.1% → 71.0%, mid-zone slope splits inverted (no signal). Back in the day-trading list, and the satellites are rebuilt too: 10m-vs-60m (10m same-direction extreme runs BELOW baseline) and the continuation ladder (bear marquee full-ATR 66% → 63.8%, story holds).

Trigger Box — Credit Spread Win Rates →

Mislabel. The 93.6% figure is short-strike no-touch probability, not a realized P&L on credit spreads. Real spreads carry credit, gamma, and settlement risk that the page doesn’t model. Needs reframing to “no-touch probability” with explicit caveats.

Bilbo Box Breakout (raw) →

Near-zero edge. 51% / +0.03R across 50,889 events. Raw signal isn’t tradable on its own. Either retire or repackage as the negative finding that motivates Bilbo Box × HTF PO Bracket Exits.

Bilbo Box × HTF PO Bracket Exits →

Sound frame, shaky entry. The bracket-exit framework is reasonable, but the underlying signal it brackets (raw Bilbo Box Breakout) sits in drafts for near-zero edge. Needs a stronger entry trigger before it’s tradable — or a rebuild that doesn’t depend on the raw box.

Bilbo Continuation Rebuilt 2026-07-11 →

Rebuilt from code. The full 61.8 → 200% extension ladder is now regenerated by scripts/rerun_bilbo_satellites.py with the point-in-time-safe 1h PO join. Bear marquee full-ATR reach 66% → 63.8% (vs 32.4% baseline) — the “let bearish Bilbo winners run” story holds with slightly smaller numbers.

4h PO Reversal Anatomy →

Stale code. Still under re-verification. Event log and velocity buckets came from a script no longer in the repo. Rebuild before promoting to a tradable study.

Daily 21 EMA Reversion →

Underpowered. 50 episodes is too thin to stand alone. Useful as context for swing setups, not a standalone tradable. Needs explicit “close-confirmed / next-session executable” framing.

Premarket ATH Morning Fade →

Too rare. Real edge per event but frequency is only 3.7% (post-rerun n=243). Standalone page doesn’t earn its slot — consolidate with Call-to-Put Reversal under a unified “morning fade” cluster.

Multiday Put Trigger Reversion →

Needs baseline. Reads as a thinner cousin of Call-to-Put Reversal. Needs an explicit weekly-trigger definition and a baseline comparison before it earns a slot beside the keepers.

Failed Studies — No Edge Found 3 studies

These were tested and didn’t hold up. They should be ignored as playbook material. If you want to revisit any of them, expect to redo from scratch with a different framing — the current versions don’t carry tradable edge.

TenAM Traffic PO Divergence →

Honest negative finding. n=36,829 PO divergences with regular signals under-performing baseline. Hidden divergences carry only a marginal, time-window-conditional edge. Don’t trade off this page.

10m vs 60m PO Rebuilt 2026-07-11 →

Rebuilt from code, conclusion sharpened. Both columns now regenerate from scripts/rerun_bilbo_satellites.py with point-in-time-safe joins (10m = previous completed bar). A same-direction 10m PO extreme runs below baseline (−9.9pp bull, −3.3pp bear) — only the 60m PO carries the Bilbo signal.

4h PO × OpEx Pin Release →

Over-claiming. Headline cohort n=13. Either deepen the sample dramatically or retire the page. Current version sells more confidence than the data supports.

← back to core playbook