dennisthiessen/signal-platform
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Retire completed GTL tuning harnesses

Browse Source
This commit is contained in:
dennisthiessen
2026-07-13 16:40:29 +02:00
parent 1d84a40c04
commit 8e09f239c8
16 changed files with 59 additions and 2057 deletions
Download Patch File Download Diff File Expand all files Collapse all files
README.md
+7 -59
View File
@@ -459,65 +459,14 @@ metrics. Keep the SSH tunnel open only while creating the snapshot; the backtest
run itself is local/offline. `backtest_snapshots/` and generated backtest reports
are git-ignored.
### One-command GTL tuning run
### Archived GTL tuning decision
The Gate Target Ladder has a research-only, single-variable matrix for testing
whether its useful screening behavior can be made more explicit. It runs 20
complete backtests **sequentially** so each arm gets the full worker pool:
```bash
# macOS/Linux
.venv/bin/python scripts/run_gtl_tuning_matrix.py \
backtest_snapshots/prod.sqlite --workers 14
```
The arms cover GTL history length, grid density, pivots, price-traffic scoring,
proposal merging, target-zone width, candidate count, and maximum target
distance. Every arm includes the full-period production book, the fixed
`2024-07-01` train/test split, a candidate-level paired audit, and the same
pre-registered robustness screen. This is one long command, not a Cartesian
parameter search; expect total runtime to be roughly 20 times one full local
backtest.
Progress is checkpointed after every arm to
`reports/backtest-YYYYMMDD-gtl-tuning-matrix.json` and the matching `.md` table.
The large per-arm reports are removed only after successful consolidation. If
the run fails, they remain available for diagnosis; pass `--keep-arm-reports`
to retain them after success too. No arm changes live scanner defaults or
deploys anything.
The replacement matrix found no winning single constant. Its evidence-selected
follow-up keeps control geometry and isolates retained versus added cohorts in
one 13-arm confirmation/union run:
```bash
# macOS/Linux
.venv/bin/python scripts/run_gtl_confirmation_matrix.py \
backtest_snapshots/prod.sqlite --workers 12
```
It first verifies exact control parity with the completed tuning matrix, then
checkpoints consolidated JSON and Markdown reports under
`reports/backtest-YYYYMMDD-gtl-confirmation-matrix.*`.
The confirmation matrix's only near-hit was strength-1000 intersection: it
improved full/train/post-2024 Sharpe but missed the unchanged drawdown guardrail
by 0.3 percentage points. The final narrow sensitivity check is:
```bash
.venv/bin/python scripts/run_gtl_strength_sensitivity.py \
backtest_snapshots/prod.sqlite --workers 12
```
It checks eight coarse scales around 1000, verifies exact control and
strength-1000 replication, and requires two adjacent scales to pass every
original guardrail before calling the result stable.
Final result: control and strength-1000 replication both passed, but there was
no adjacent passing plateau. Scale 1500 passed in isolation while lowering CAGR
and setup expectancy; its neighbors failed. The research decision is therefore
to keep the frozen GTL unchanged and evaluate any future challenger only on new
forward data. See the [full research record](docs/research/sr-levels-and-exits.md#gtl-tuning-matrix).
The completed replacement, cohort-composition, and strength-sensitivity
matrices found no stable improvement over the frozen Gate Target Ladder. The
temporary matrix runners and tuning hooks have been retired; their three compact
consolidated report pairs remain in `reports/` as the decision audit. Keep the
GTL unchanged and evaluate any future challenger only on new forward data. See
the [full research record](docs/research/sr-levels-and-exits.md#gtl-tuning-matrix).
### Reading a local backtest report
@@ -557,7 +506,6 @@ Research-only flags, all off by default (the default report is byte-identical to
| `BACKTEST_HOLDOUT_SPLIT=YYYY-MM-DD` | Adds a `holdout` section: train (entries before) vs test (entries on/after), as disjoint books |
| `BACKTEST_MIN_RR_SWEEP=1` | Sweeps the activation R:R floor against portfolio Sharpe. Combine with `BACKTEST_HOLDOUT_SPLIT` to sweep out-of-sample |
| `BACKTEST_SR_VARIANT=<arm>` | Research-only S/R detector/gate arm; see `docs/research/sr-levels-and-exits.md` for the detector and hidden-feature matrices |
| `BACKTEST_GTL_CONFIG=<json>` | Parameterizes only the `gtl_tuning` research arm; normally set by `run_gtl_tuning_matrix.py` |
| `BACKTEST_ENTRY_START=YYYY-MM-DD` | Restrict candidate entry dates to a validation window |
| `BACKTEST_ENTRY_END=YYYY-MM-DD` | Restrict candidate entry dates to a training window |
| `BACKTEST_SR_AUDIT=1` | Add momentum-slice candidate rows for paired S/R cohort comparison |