# Walk-forward exam — commit 1c264d5 (+ uncommitted walk-forward feature)

Rolling-fold out-of-sample evaluation of `backtest/gpu_backtest.py
--walk-forward` for XAU/USD, window 2019-01-01 → 2026-08-07, all 4
primary TFs at $0.65 round-trip spread, train 12mo / test 3mo / step 3mo
(26 folds per TF). One continuous replay per TF with the fixed HEAD
config — folds are slices of the book by entry time, no per-fold
re-optimization. News off (backtester default).

Captured 2026-08-07. The working tree at capture was commit `1c264d5`
plus the then-uncommitted walk-forward implementation (gpu_backtest.py,
test_all.py Module 23, CLAUDE.md); `manifest.json` records `dirty: true`
and the sha256 of `git diff` at capture. `scripts/compare_baseline.py`
passed 32/32 against `baseline/da61cff` with this exact tree immediately
before these runs — the feature leaves default (non-walk-forward) books
byte-identical, and the 2019→2026 data fetch touched no pinned parquet.

Per run dir: `trades.csv` (full book incl. train-period trades),
`summary.txt` (backtester stdout — full report + the walk-forward fold
table and aggregated OOS stats), `equity.csv` + `metrics.json` (derived
from the book via scripts/baseline_common.py; these cover the WHOLE
window, not just OOS). `run.log` is stderr, provenance only.

`manifest.json` records the exact argv/command per run, sha256 of every
input parquet (the five 2019→present files, also in
`backtest/cache/data_manifest.json`) and every output file, backend, and
the walk-forward parameters.

Headline aggregated OOS (union of test folds, each trade once,
2020-01-01 → 2026-07-01):

| TF  | OOS trades | Total R | exp/trade | PF   | Max DD (R) | folds PF<1 |
|-----|-----------:|--------:|----------:|-----:|-----------:|-----------:|
| M5  |        991 | +126.54 |    +0.128 | 1.35 |      19.16 |    11 / 26 |
| M15 |       2001 | +139.15 |    +0.070 | 1.18 |      55.06 |     9 / 26 |
| H1  |        638 |  +33.28 |    +0.052 | 1.15 |      23.84 |    13 / 26 |
| H4  |        188 |  +33.31 |    +0.177 | 1.52 |       5.66 |     9 / 26 |

Flagged folds cluster in 2020H2–2023 on every TF; from fold ~16
(test 2023-10) onward the test folds are almost uniformly positive.
Consistent with the 2026-08-04 Oanda re-pin finding that all TFs earn
mainly 2024+. These runs are an OOS robustness read on the shipped
config over deep history — NOT a new baseline and NOT a rule-1
validation event (window, cost and provider differ from the pinned
books; nothing was tuned here).