# Replay recording campaign — 2026-09-09 (2-year Oanda window, M5 + M15, $0.65)

## Purpose and owner decisions

The owner asked for a 2-year replay of the M5 and M15 timeframes that writes down, for
every trade setup the system generates, the indicator weights and values the decision was
made with and the economic-calendar events around it, in a form that a later study can use
to decide whether indicator weights should change, or indicators be dropped or replaced.
Owner decisions (2026-09-09): also record one compact row per ANALYZED bar (not only the
bars that became trades), and restore the Forex Factory calendar scraper to `scripts/`.

Nothing in the signal core changed. The recorder is opt-in (`--record DIR`, `--record-bars`)
and read-only with respect to the signal path; the trade book with recording on is
byte-identical to the book with it off (MD5 below).

## Harness

`backtest/gpu_backtest.py --pair XAUUSD --start 2024-09-05 --end 2026-09-05 --spread 0.65
--workers 12 --primary-tf {5min|15min}`, pair-factory config (no --set/--quality/--strategies/
--confirm-tfs; confirmation TFs from the factory map: M5→15min,1h and M15→5min,1h),
`Backend: NumPy CPU fallback` (CuPy cannot find CUDA on this box). Code at git `28b591d`
plus the recorder (this campaign). Bars fetched once from Oanda into
`backtest/cache/oanda_XAU_USD_{5min,15min,1h}_20240705_20260905.parquet` (62-day warm-up pad)
and read from the cache by every later run. Each timeframe was run twice: a CONTROL run
(`--export-csv book_control.csv`) and a RECORDED run (`--export-csv book.csv --record DIR
--record-bars`). Outputs in `backtest/research/replay_2026-09/{m5,m15}/` (gitignored).

## Books (identical for control and recorded — that is the point)

| book | trades | total R | WR | PF | max DD |
|---|---|---|---|---|---|
| M5 whole window | 781 | +146.07 | 52.9% | 1.52 | 10.1 |
|   first half | 189 | +28.01 | 49.7% | 1.41 | 8.4 |
|   second half | 592 | +118.06 | 53.9% | 1.56 | 10.1 |
| M15 whole window | 793 | +117.62 | 60.4% | 1.41 | 14.8 |
|   first half | 383 | +38.22 | 58.5% | 1.26 | 14.8 |
|   second half | 410 | +79.40 | 62.2% | 1.55 | 9.9 |

Halves split the entry-time range in two; M5 trades cluster in the second year.

### Behaviour-neutrality check

| TF | control MD5 | recorded MD5 | manifest `book_csv_md5` |
|---|---|---|---|
| M5 | 29120a1ee32979431d062f210d41a9d1 | 29120a1ee32979431d062f210d41a9d1 | same |
| M15 | e62d3ded56f35572f52a96f604a99cf4 | e62d3ded56f35572f52a96f604a99cf4 | same |

Replay time M5 114 s / M15 35 s on 12 workers (control runs 110 s / 35 s — recording costs
seconds). `backtest/replay_scorecard.py DIR --check` passes all eight self-consistency
invariants on both runs: trade counts match the manifest; every trade has its primary block
and both confirmation blocks; base weight × regime multiplier = effective weight and
contribution = vote × effective weight × confidence/100 on every one of the 17,869 (M5) /
18,070 (M15) indicator rows; the composite score is reproduced from the stored rows on every
trade and timeframe; every regime multiplier equals the table in the manifest; the bar table
has exactly one row per analyzed bar and contains every trade's bar.

### Cross-check against the owner's 3-year baseline books (same window, pre-2026-09-07 code)

| TF | 2-yr run | baseline `sizing_2026-09/base_*_3yr_065.csv` cut to the window | shared entries |
|---|---|---|---|
| M5 | 781 / +146.07R | 769 / +153.58R | 757 (24 only new, 12 only baseline) |
| M15 | 793 / +117.62R | 788 / +115.78R | 774 (19 only new, 14 only baseline) |

The baseline files are dated 2026-09-05 and predate the 2026-09-07 outlier-filter change
(which moved the 3-yr M5 book 857→870 trades, +161.3→+153.6R), so the differences are the
expected ones; byte identity was never expected here. The behaviour-neutrality proof is the
MD5 pair above, on identical code.

## What was recorded (storage decision: Parquet, one folder per run)

Parquet keeps exact typed numbers and UTC timestamps, compresses well (M5: 20 MB for 86k
bars × 172 columns), and pandas reads a file in one line; the repo already uses it for the
bar cache. SQLite would need JSON blobs or many tables with string timestamps; CSV loses
types and is several times larger. DuckDB can query the parquet files directly if SQL is
ever wanted.

| file | grain | M5 rows | M15 rows |
|---|---|---|---|
| `manifest.json` | CLI, window, spread, TFs, engine + gate fingerprints, git SHA, backend, regime weight table in force, observed base weights, composite thresholds, bar counts + skip reasons, cache file names, calendar fingerprint, table row counts, book MD5 | 1 | 1 |
| `trades.parquet` | one row per ACCEPTED trade: the 63 book columns + per-timeframe regime / composite / confidence / ATR / vote sums, penalty tags, setup fields, all warnings and reasons, fingerprints, calendar scalars | 781 | 793 |
| `indicators.parquet` | one row per trade × timeframe × indicator that voted: vote, value, confidence, base weight, regime multiplier, effective weight, contribution, agree/oppose/neutral, regime | 17,869 | 18,070 |
| `events.parquet` | one row per trade × calendar event from 4 h before entry to exit (or 4 h after entry), plus all-day rows on the entry's New-York date, with phase before / during_hold / after_exit / all_day | 7,803 | 12,932 |
| `bars.parquet` | one row per ANALYZED primary bar (after weekend/session filters, before the setup gates): final score/signal/quality, has_setup, per-timeframe regime/composite and per-indicator vote/value/confidence/effective weight/contribution as float32, plus forward moves 1 h / 4 h / 24 h and MFE/MAE in ATRs on the raw bars | 86,192 | 29,917 |

Candidates before the 5-bar cooldown: M5 1,279 → 781 accepted; M15 1,535 → 793. An
indicator that did not fire (divergence, candlestick) has NO row / NaN — never a zero vote.
Reader: `backtest.replay_record.load_run(dir)`. First analysis pass:
`backtest/replay_scorecard.py DIR --what-if --bars` (full output saved as `scorecard.txt`
in each run folder).

## First look (descriptive statistics only — NOT evidence for a change)

Everything below is in-book: it cannot see trades a change would have created, and it
ignores the MTF/penalty/gate chain that sits between the composite and the entry. The 2-year
window is a strong gold up-trend, so the "no vote" baseline itself drifts up about +0.37 ATR
per 4 h and +1.9 ATR per 24 h on M5 (+0.15 / +0.9 on M15); any BUY-vote number must be read
against that baseline, not against zero.

Trade-level (primary timeframe):
- Bollinger opposes the entry on 731 of 781 M5 trades (94%): a mean-reversion reading
  inside a trend-following entry — the "constant bias" seen in the 2026-08-04 audit — and
  those trades carry +152R. RSI is neutral on 73% of M5 entries.
- M5 trades where divergence OPPOSED the entry did best (158 trades, +0.37R mean, 63% wins)
  and trades where it was absent did worst (546, +0.13R); the same "inverted" pattern as the
  2026-08-04 in-book audit, which replayed WORSE when "fixed" — the reminder that this table
  is a hypothesis generator.
- What-if (M5): removing ema_trend would drop 206 accepted trades carrying +36.8R below the
  entry band; removing MACD 211 / +21.2R; ADX 127 / +22.6R. On M15 hardly any trade loses
  direction from a single-indicator removal (the entries sit further above the band); the
  changes show up as strong/weak band moves instead.
- Calendar: M5 trades held through at least one high-impact USD release did better
  (69 trades, +0.42R mean vs +0.19 overall; M15 224 trades, +0.23 vs +0.15) — the same
  direction as the 2026-09-06 addendum. Entries within 30 min after a release are slightly
  negative on both TFs (40 and 35 trades — too few to act on).

Bar-level (all analyzed bars, drift NOT yet removed):
- Every indicator's SELL-side votes were followed, on average, by a move AGAINST the vote
  on both timeframes (e.g. M5 ema_trend −0.27 ATR / 4 h, divergence −0.63); this is the
  up-drift plus the known "SELL negative everywhere" finding. The next analysis step is to
  subtract the per-bar baseline (or split by regime) before scoring votes.
- Above-baseline BUY callers on M5 at 4 h: ADX +0.50, keltner +0.50, RSI +0.45,
  candlestick +0.46, ROC +0.46, ema_trend (+0.67 vs baseline +0.32 among its own neutral
  bars); below baseline: divergence +0.30, Bollinger +0.36.
- Bars what-if (M5): dropping divergence would newly qualify 2,268 bars whose later move
  averaged +0.26 ATR / 4 h and drop 3,611 bars averaging −0.26 — the one single-indicator
  removal where both sides point the same way. Dropping ema_trend or ADX points the other
  way (the bars they add did worse, the bars they remove did better).

## Pre-registered rule for anything that comes out of these files

A candidate change (weight, regime multiplier, dropped or replaced indicator) is a
hypothesis until it passes CLAUDE.md rule 1 on a full rerun through
`analyze_from_dataframes`: positive total R in BOTH halves of a ≥12-month window, PF ≥ 1.3
at $0.65, trailing exits, next-open fills, plus the 2024-06→2025-03 out-of-sample window for
marginal edges, compared with a CONTROL produced on the same harness. Nothing in this
campaign changes any number.

## Suggested next analysis (not done here)

1. Drift-adjust the bar table (subtract the neutral/absent baseline per regime and hour) and
   re-score each indicator's standalone hit rate by timeframe role.
2. For the 2–3 most suspect indicators (divergence, Bollinger, RSI on M5), run the real
   rule-1 rerun with `--set technical.regime_weights_override=...` or `disabled_indicators`
   and compare against the control books in this folder.
3. Re-run this recording after any config change so the folders stay comparable (the
   manifest carries the engine fingerprint and the regime table for exactly that purpose).

Tests: Module 33 "backtest replay recording" (12 tests); suite green.