Task
Standardize pinned deep-window scenarios on 1998-2026 instead of 2000-2026, so the deep baseline consistently captures the full dotcom round-trip (1998-99 run-up + 2000-02 bust) rather than starting near the peak.
Scope (already scoped 2026-06-21 — smaller than it sounds)
A grep for start_date 2000-01-01 across trading/test_data/ finds exactly one pinned spec:
trading/test_data/walk_forward/exit-timing-surface-deep-2000-2026.sexp
The canonical deep baseline is already 1998-2026 (goldens-sp500-historical/sp500-1998-2026.sexp); other pinned goldens use 1998-12-22 / 1998-01-01 / 2010. So this is a one-file migration, not a sweep.
Do NOT touch the ~23 dev/backtest/scenarios-* hits — those are historical research run outputs (records of past runs), not pinned specs.
Steps
- Change the window in
exit-timing-surface-deep-2000-2026.sexp to (start_date 1998-01-01), and rename the file (window is in the name) to ...-1998-2026.sexp. Update any references (dune / scripts / docs) to the old name.
- Universe consistency: if the spec pins a PIT-2000 universe (SP500 as-of-2000 or top-N-2000), switch to the as-of-1998 snapshot (
top-N-1998.sexp / sp500 1998 PIT) so membership matches the new start. Verify the data floor covers 1998 for those names (fetch-historical-data skill).
- Re-pin expected bands: changing the window changes the data → the
expected metric bands must be re-derived from a fresh run (this is a re-pin, NOT a pure date swap). Run via snapshot mode if N>=1000.
dune runtest for the walk_forward goldens; confirm green.
Why queued, not done
Surfaced during the 2026-06-21 barbell breadth-confirm work (the breadth cell intentionally stays 2000-26 for blend-comparability with grid cell A — see dev/backtest/barbell-breadth-2026-06-21/). This migration is independent and re-pins goldens, so it gets its own PR.
Origin: user request 2026-06-21.
Task
Standardize pinned deep-window scenarios on 1998-2026 instead of 2000-2026, so the deep baseline consistently captures the full dotcom round-trip (1998-99 run-up + 2000-02 bust) rather than starting near the peak.
Scope (already scoped 2026-06-21 — smaller than it sounds)
A grep for
start_date 2000-01-01acrosstrading/test_data/finds exactly one pinned spec:trading/test_data/walk_forward/exit-timing-surface-deep-2000-2026.sexpThe canonical deep baseline is already 1998-2026 (
goldens-sp500-historical/sp500-1998-2026.sexp); other pinned goldens use 1998-12-22 / 1998-01-01 / 2010. So this is a one-file migration, not a sweep.Do NOT touch the ~23
dev/backtest/scenarios-*hits — those are historical research run outputs (records of past runs), not pinned specs.Steps
exit-timing-surface-deep-2000-2026.sexpto(start_date 1998-01-01), and rename the file (window is in the name) to...-1998-2026.sexp. Update any references (dune / scripts / docs) to the old name.top-N-1998.sexp/ sp500 1998 PIT) so membership matches the new start. Verify the data floor covers 1998 for those names (fetch-historical-dataskill).expectedmetric bands must be re-derived from a fresh run (this is a re-pin, NOT a pure date swap). Run via snapshot mode if N>=1000.dune runtestfor the walk_forward goldens; confirm green.Why queued, not done
Surfaced during the 2026-06-21 barbell breadth-confirm work (the breadth cell intentionally stays 2000-26 for blend-comparability with grid cell A — see
dev/backtest/barbell-breadth-2026-06-21/). This migration is independent and re-pins goldens, so it gets its own PR.Origin: user request 2026-06-21.