# Overnight event nights — RESULTS

**Date:** 2026-09-16  
**Pre-reg:** `studies/2026-09-16-overnight-event-nights-prereg.md` (frozen before any event-conditioned overnight number was computed).  
**Script:** `studies/run_overnight_event_nights.py`  
**Price:** `<data-root>/fp-data/glbx/{NQ,ES,GC}/1m/2020..2026.csv`  

Top-10 NDX by current market cap (approximate weights): NVDA, AAPL, GOOGL, GOOG, MSFT, AMZN, SPCX, META, AVGO, TSLA  
Heavy names (weight ≥ 5%): NVDA, AAPL, GOOGL, GOOG, MSFT, AMZN  
Event rows loaded: 94406

Window and friction match the 2026-09-15 tape study. Quad-witching open dates excluded.
Bonferroni: NQ/ES 7 arms α=0.00714 (|t|≳2.69); GC 6 arms α=0.00833 (|t|≳2.64).

## NQ
n nights 1642 · unconditional net $134.39 · t 1.82 · win 56.8% · Σ gross $243660
**Positive control (pre-FOMC):** n=53 net $972.70 t=3.12 → PASS
| arm | n | mean gross $ | mean net $ | win % | t | share of gross | placebo p(≥obs) | train net | train t | confirm net | confirm t | verdict |
|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---|
| A_prefomc | 53 | 986.70 | 972.70 | 69.8 | 3.12 | 21.5% | 0.014 | 927.41 | 2.79 | 1041.71 | 1.69 | supported |
| B_amc_heavy | 100 | 625.65 | 611.65 | 58.0 | 1.65 | 25.7% | 0.061 | 115.04 | 0.33 | 1269.95 | 1.76 | inconclusive |
| C_amc_light | 78 | 685.26 | 671.26 | 59.0 | 1.67 | 21.9% | 0.047 | 735.44 | 1.85 | 583.73 | 0.74 | inconclusive |
| D_fed_speech_after16 | 69 | -368.55 | -382.55 | 50.7 | -0.74 | -10.4% | 0.931 | -107.04 | -0.24 | -520.30 | -0.70 | inconclusive |
| E_cpi_pce_nfp | 261 | 305.77 | 291.77 | 60.2 | 1.40 | 32.8% | 0.189 | 40.61 | 0.18 | 659.02 | 1.65 | inconclusive |
| G_policy_geo | 413 | 180.84 | 166.84 | 57.6 | 1.07 | 30.7% | 0.357 | -32.31 | -0.20 | 339.85 | 1.33 | inconclusive |
| H_quiet | 583 | 106.86 | 92.86 | 56.8 | 0.79 | 25.6% | 0.661 | 56.11 | 0.54 | 159.12 | 0.59 | inconclusive |

Exclusive hierarchy (first match wins):
| arm | n | mean net $ | t | share of gross |
|---|---:|---:|---:|---:|
| A_prefomc | 53 | 972.70 | 3.12 | 21.5% |
| B_amc_heavy | 89 | 490.55 | 1.26 | 18.4% |
| C_amc_light | 58 | 692.81 | 1.66 | 16.8% |
| D_fed_speech_after16 | 61 | -521.38 | -0.95 | -12.7% |
| E_cpi_pce_nfp | 187 | 125.97 | 0.50 | 10.7% |
| G_policy_geo | 177 | 173.34 | 0.77 | 13.6% |
| H_quiet | 583 | 92.86 | 0.79 | 25.6% |

## ES
n nights 1642 · unconditional net $51.18 · t 1.25 · win 55.1% · Σ gross $131650
**Positive control (pre-FOMC):** n=53 net $424.54 t=2.80 → PASS
| arm | n | mean gross $ | mean net $ | win % | t | share of gross | placebo p(≥obs) | train net | train t | confirm net | confirm t | verdict |
|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---|
| A_prefomc | 53 | 453.54 | 424.54 | 66.0 | 2.80 | 18.3% | 0.047 | 497.17 | 2.64 | 313.86 | 1.22 | supported |
| B_amc_heavy | 100 | 210.88 | 181.88 | 59.0 | 1.04 | 16.0% | 0.200 | -79.88 | -0.44 | 528.85 | 1.63 | inconclusive |
| C_amc_light | 78 | 454.65 | 425.65 | 57.7 | 2.04 | 26.9% | 0.019 | 517.67 | 2.08 | 300.17 | 0.83 | suggestive |
| D_fed_speech_after16 | 69 | -155.98 | -184.98 | 44.9 | -0.68 | -8.2% | 0.897 | -205.63 | -0.69 | -174.65 | -0.45 | inconclusive |
| E_cpi_pce_nfp | 261 | 116.43 | 87.43 | 59.4 | 0.78 | 23.1% | 0.385 | 8.18 | 0.06 | 203.31 | 1.07 | inconclusive |
| G_policy_geo | 413 | 78.84 | 49.84 | 56.4 | 0.56 | 24.7% | 0.471 | -58.30 | -0.53 | 143.79 | 1.07 | inconclusive |
| H_quiet | 583 | 94.02 | 65.02 | 54.2 | 1.01 | 41.6% | 0.406 | 75.67 | 1.01 | 45.82 | 0.38 | inconclusive |

Exclusive hierarchy (first match wins):
| arm | n | mean net $ | t | share of gross |
|---|---:|---:|---:|---:|
| A_prefomc | 53 | 424.54 | 2.80 | 18.3% |
| B_amc_heavy | 89 | 109.90 | 0.60 | 9.4% |
| C_amc_light | 58 | 568.41 | 2.42 | 26.3% |
| D_fed_speech_after16 | 61 | -273.67 | -0.96 | -11.3% |
| E_cpi_pce_nfp | 187 | -15.56 | -0.11 | 1.9% |
| G_policy_geo | 177 | 59.91 | 0.44 | 12.0% |
| H_quiet | 583 | 65.02 | 1.01 | 41.6% |

## GC
n nights 1631 · unconditional net $158.21 · t 2.11 · win 55.2% · Σ gross $297180
**Positive control (pre-FOMC):** n=52 net $99.27 t=0.24 → FAIL
| arm | n | mean gross $ | mean net $ | win % | t | share of gross | placebo p(≥obs) | train net | train t | confirm net | confirm t | verdict |
|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---|
| A_prefomc | 52 | 123.27 | 99.27 | 55.8 | 0.24 | 2.2% | 0.595 | 198.26 | 1.15 | -46.86 | -0.05 | inconclusive |
| D_fed_speech_after16 | 69 | 212.17 | 188.17 | 52.2 | 0.38 | 4.9% | 0.467 | -136.61 | -0.63 | 350.57 | 0.48 | inconclusive |
| E_cpi_pce_nfp | 252 | 493.17 | 469.17 | 61.5 | 2.47 | 41.8% | 0.033 | 211.30 | 1.70 | 842.21 | 1.97 | suggestive |
| F_eia_wasde | 682 | 112.62 | 88.62 | 53.1 | 0.76 | 25.8% | 0.799 | 40.71 | 0.61 | 159.98 | 0.59 | inconclusive |
| G_policy_geo | 410 | 266.95 | 242.95 | 54.9 | 1.68 | 36.8% | 0.279 | -50.93 | -0.51 | 494.28 | 1.95 | inconclusive |
| H_quiet | 584 | 254.11 | 230.11 | 56.7 | 1.97 | 49.9% | 0.264 | 27.84 | 0.36 | 590.33 | 2.01 | inconclusive |

Exclusive hierarchy (first match wins):
| arm | n | mean net $ | t | share of gross |
|---|---:|---:|---:|---:|
| A_prefomc | 52 | 99.27 | 0.24 | 2.2% |
| D_fed_speech_after16 | 61 | 405.84 | 0.81 | 8.8% |
| E_cpi_pce_nfp | 183 | 512.17 | 2.41 | 33.0% |
| F_eia_wasde | 433 | -92.20 | -0.66 | -9.9% |
| G_policy_geo | 177 | 270.07 | 1.26 | 17.5% |
| H_quiet | 584 | 230.11 | 1.97 | 49.9% |

## Pipeline control

- NQ pre-FOMC: PASS n=53 t=3.12 net $972.70
- ES pre-FOMC: PASS n=53 t=2.80 net $424.54
- GC pre-FOMC: FAIL n=52 t=0.24 net $99.27

## Honest reading

Same frozen arms as the pre-reg. This table is the rerun after Alpaca wire was filled 2024-01 → 2026-09 (80 pages/month cap). Arm G n on NQ moved 360 → 413; verdicts did not change. GDELT remains a thin 2020 slice (DOC API still 429 at 20s).

**The overnight drift does not live only in identifiable event nights.** On NQ, quiet nights are 583 of 1,642 (36%) and still carry 26% of total gross dollars, mean net +$93, t = 0.79. The pre-FOMC night is real, large, and the only Bonferroni-supported cell: 53 nights, +$973 net, t = 3.12, 21.5% of all gross from 3.2% of nights. That is concentration, not an explanation of the rest.

What does **not** explain the drift (same prominence as the one that does):

- **AMC top-10 earnings (B, C):** positive point estimates on NQ (+$612 / +$671) but t ≈ 1.65; heavy names do not replicate in 2024–2026. **Inconclusive.** Today's top-10 includes SPCX — survivorship.
- **Fed speeches after 16:00 ET (D):** NQ −$383, t = −0.74. If anything the other way. **Inconclusive.**
- **CPI / PCE / NFP mornings (E):** NQ +$292, t = 1.40. **Inconclusive.** On GC this is suggestive (t = 2.47) but below Bonferroni.
- **EIA / WASDE (F, GC):** t = 0.76. Exclusive hierarchy **negative** (−$92). Does not explain gold overnight.
- **Policy / geopolitics (G):** t = 1.07 NQ after the fuller wire. **Inconclusive.**
- **Quiet (H):** still positive, not significant. The leftover *is* most of the sample.

ES tracks NQ: pre-FOMC supported (t = 2.80); AMC-light is *suggestive* (t = 2.04) and dies in confirm. GC pre-FOMC **fails** the pipeline control (t = 0.24) — the 2026-08-25 robust cell was NQ, not gold.

Pipeline: NQ pre-FOMC passed (required). ES passed. GC did not; GC arms are tabulated, not promoted.

No re-binning. Unconditional NQ net $134 / t 1.82 is the same order as the tape study ($168 / t 2.27); this sample excludes quad-witching opens.

## Robustness — PIT NDX top-10 (not a new arm family)
Membership reconstructed from FMP historical-nasdaq-constituent; weights = market-cap share of members that day (`features/ndx_top10_daily.jsonl`). Same window/friction. Frozen primary B/C still use today's list.

### NQ
| arm | n | mean net $ | win % | t | train net | confirm net |
|---|---:|---:|---:|---:|---:|---:|
| B_pit_heavy | 93 | 576.65 | 60.2 | 1.42 | 294.80 | 852.49 | inconclusive |
| C_pit_light | 78 | 461.58 | 56.4 | 1.29 | 65.72 | 1352.25 | inconclusive |

### ES
| arm | n | mean net $ | win % | t | train net | confirm net |
|---|---:|---:|---:|---:|---:|---:|
| B_pit_heavy | 93 | 238.61 | 60.2 | 1.26 | 58.77 | 414.62 | inconclusive |
| C_pit_light | 78 | 191.35 | 59.0 | 1.08 | 21.00 | 574.65 | inconclusive |


## Robustness — SEC 8-K Item 2.02 timestamps (not a new family)
1,751 of 2,044 earnings rows stamped from EDGAR 8-K 2.02. 265 are **BMO** (06:00–11:00 ET) and were mis-labelled AMC in the frozen primary. This table uses only exact AMC 8-K times on today's top-10.

### NQ
| arm | n | mean net $ | win % | t |
|---|---:|---:|---:|---:|
| B_exact_amc_heavy | 100 | 611.65 | 58.0 | 1.65 | inconclusive |
| C_exact_amc_light | 57 | 658.46 | 57.9 | 1.55 | inconclusive |

### ES
| arm | n | mean net $ | win % | t |
|---|---:|---:|---:|---:|
| B_exact_amc_heavy | 100 | 181.88 | 59.0 | 1.04 | inconclusive |
| C_exact_amc_light | 57 | 554.99 | 57.9 | 2.33 | suggestive |