signal-platform/docs/research/execution-recovery.md

# Execution recovery (2026-07-18) — when you fill is the edge you leave on the table

Report: `reports/execution_recovery_matrix.json` / `.md`  
Follows Phase A A4 ([phase-a-matrix.md](phase-a-matrix.md)).  
Arms: `close_control` · `next_open` · `stale_close` · `next_open_gap2`.

Mechanics verified before sign-off: `stale_close` re-anchors stop to
fill − 1.5×ATR(signal day); no same-day stop after a close fill; each t−1
candidate keeps its own gate/rank cross-section (no lookahead). Numbers trusted.

---

## Results

| Arm | Full Sharpe | Full CAGR | Full DD | Train Sharpe | Val Sharpe |
|---|---:|---:|---:|---:|---:|
| **close_control** | **1.77** | **48.3%** | **21.6%** | 1.75 | 1.68 |
| **stale_close** | **1.57** | **40.6%** | **21.8%** | 1.36 | **1.74** |
| next_open | 1.20 | 30.0% | 28.2% | 0.94 | 1.44 |
| next_open_gap2 | 1.06 | 25.2% | 27.8% | 0.72 | 1.41 |

### Monotonicity (the strongest evidence)

In every window, Sharpe recovers as the fill moves **toward** the signal:

| Window | next_open → stale_close → close |
|---|---|
| Train | 0.94 → 1.36 → 1.75 |
| Validation | 1.44 → 1.74 → 1.68* |
| Full | 1.20 → 1.57 → 1.77 |

\*Val close 1.68 is within SE of stale 1.74 — not a break of the story.

A dead edge does **not** produce a monotone gradient in fill timing. A live edge
that is progressively surrendered to execution delay does. Combined with full-period
**DD recovery** (28.2% → 21.8% ≈ close’s 21.6%), this confirms the diagnosis:
**when you fill**, not decaying alpha.

### Auto-flag `recover: False` is not a null

Pre-registered recovery applied the “≥ close − 0.5 SE” bar to **`stale_close`**,
the **lower-bound** arm (one full session of lag). Flags:

| Flag | Result |
|---|---|
| near_close_control (val) | **True** (Δ +0.06) |
| beats_next_open | **True** (val Δ +0.30) |
| train_ok | **False** (1.36 vs need ≥ ~1.39 — miss by ~0.03) |

The floor missed “≥ close − 0.5 SE” by 0.03 while the live design is expected to
sit **above** the floor. The flag worked correctly on the wrong object.

**Log sentence:** *Partial recovery proven; full recovery needs same-day fill.*

### Live outcome bracket

| Bound | Arm | Meaning |
|---|---|---|
| Floor | stale_close ~1.57 full | One-session-stale close fill (conservative) |
| Ceiling | close_control ~1.77 full | Same-day close; optimistic only by final ~15 min of signal info |

Real near-close execution (scan ~15:30–15:40 ET on a ~99% complete bar, MOC by
15:50/15:55) is signal-at-partial-bar filled at the same close the control uses.
**Live truth is bracketed [1.57, 1.77]** with residual uncertainty of ~15 minutes
of staleness, not 24 hours — expect near the **top** of the bracket.

`stale_close` alone already justifies the schedule change. **No further fill-timing
simulation on this snapshot** — the bracket is the result. You cannot simulate
15:45 partial bars from daily data, and you do not need to.

### Gap-cap — dead (third tail-trim instance)

`next_open_gap2` worse than plain next_open on every window (full Sharpe 1.06 vs
1.20). **262** full-period gap-ups skipped — they were continuations.

This is the **third independent instance** of the same lesson:

1. Take-profit exits (gate target as TP)  
2. FIP as an in-book filter  
3. **Gap-up entry caps**

Any rule that trims the right tail improves cosmetic quality metrics and destroys
P&L. **Standing method rule:** momentum filters must be presumed guilty of
tail-trimming until shown otherwise.

---

## Decisions locked

1. **Biggest lever is execution scheduling**, not a strategy rewrite.  
2. **Until near-close execution is live:** grade strategy promotion under
   `fill_mode=next_open`; keep close-fill as historical control.  
3. **After near-close ships:** grade under a close-like / near-close fill mode
   (actual live path).  
4. **No more sim arms** on fill timing for this snapshot.  
5. **Gap-cap:** do not ship.  
6. **Strategy work** (nasdaq_all, fip_id, sector) waits until the execution path
   is decided — those experiments must be graded under the fill mode you will trade.

---

## Ops design — settle before touching the scheduler

### 1. One canonical qualifying scan per day (non-negotiable)

Post-stop gate-reset was validated on **one full-universe observation per day**.
Adding a near-close scan *alongside* the 07:00 scan would let fail→qualify
transitions complete twice as fast and **silently change** the validated re-entry
policy.

**Move** the R:R scan to near-close. **Leave** sentiment / fundamentals /
OHLCV-backfill at 07:00 (or existing early slots). Preserve **scan before
Outcome Eval** (scan late session, eval after close) so the documented
strict-gate-reset live analogue is unchanged.

### 2. Cron in `America/New_York`, not `Europe/Berlin`

DST offsets shift on different dates. A Berlin-fixed wall time drifts ~1 hour
off the US close for a week or two twice a year.

### 3. Fill mechanics and feed honesty

- NYSE MOC cutoff **15:50 ET**; Nasdaq **15:55 ET**.  
- Scan ~**15:30–15:40 ET** on a latest-price snapshot; place entries by 15:50.  
- For now: paper-trade entries marked at the **actual close**.  
- Document Alpaca entitlement: if only 15-minute-delayed SIP, a 15:35 scan sees
  ~15:20 prices — immaterial for a 12-month signal, but write it down so nobody
  treats it as a bug.

### 4. Partial-bar plumbing

- Scan synthesizes day-t’s in-progress bar from the snapshot.  
- Nightly OHLCV job overwrites with the final bar via existing idempotent upsert.  
- Forward paper record marks entries at the actual near-close fill so the live
  track measures the new execution honestly.

---

## Implementation status

| Item | Status |
|---|---|
| Research evidence | **Closed** — this doc + matrix report |
| Scheduler move (R:R scan → NY near-close) | **Not started** — blocked on ops design above |
| Partial-bar scan path | **Not started** |
| Paper fill-at-close marking | **Not started** (may already mark at close; verify when shipping) |