Research note
Saved optimizer runs: proposals, not proof of edge
Global vectors and per-stage schedules retain native results and application stamps. No actual saved row or overall verdict was inspected.
Backtest verdict: Unknown — no frozen result recorded in the inventory. Saved optimizer proposals and application metadata are shared research, not evidence of a universal passing strategy.
Open this research tool in Backtesting
The rule
The global optimizer searches a weight vector using IC, regime-alignment and anchor-penalty terms. The per-stage optimizer proposes schedules with stage selection, minimum-uplift controls and explicit operator application. These are configuration searches, not a fixed long/short entry/exit strategy. Their underlying replay window, forward horizon, folds and constraints belong to the saved proposal.
Opening defaults such as lookback, candidate count or fold count do not describe an older run whose request fields were never stored. There is no common execution cost, slippage, sizing, stop or target implied by the optimizer UI. Operator Run and Apply are explicit controls, separate from reading a result.
Methodology
Evidence as of: the INVENTORY review dated November 5, 2026; this is the consolidation planning baseline, not a new observation window. Publication does not extend any study's sample.
Use stored native JSON, creation stamps, application stamps and fold/stability arrays as the historical record. Global and schedule IDs occupy separate stores. Twenty global rows and ten schedules are page sizes, not retention or sample totals; cursor archives reach older records without execution. Six-regime historical label strings and skipped/null proposals remain unchanged.
Current-versus-suggested compares today's live configuration to a saved proposal; it is not a stored historical baseline. The inventory inspected no actual saved result. Exact dates, benchmark/basis, costs, candidate selection and validation outcome remain not recorded for this article. Fold uplift alone does not eliminate selection bias.
Results
| Study / variant | Inventory evidence | Window / basis | Strategy return / CAGR | Benchmark | Max drawdown | Closed trades / observation N | Significance / IC | Native verdict |
|---|---|---|---|---|---|---|---|---|
| Global vector proposals / folds / stability | D09 / F17 / H20 | not recorded | not recorded | not recorded | not recorded | not recorded | not recorded | Unknown — no frozen result |
| Per-stage schedules / skipped proposals | D09 / F18 / H21 | not recorded | not recorded | not recorded | not recorded | not recorded | not recorded | Unknown — no frozen result |
| Application / active overrides | D09 / H22 | not recorded | not recorded | not recorded | not recorded | not recorded | not recorded | Unknown — application is not validation |
The linked workspace retains the complete native output and notes when available. Missing evidence is not a zero return, a successful empty result or a failed strategy.
Verdict and limitations
Unknown — no result/verdict recorded in this inventory. Applying weights does not prove edge, and positive uplift does not establish a trading PASS. Missing original settings remain “not recorded” rather than being replaced by current defaults. Archive reads neither apply proposals nor change retention.
Sources and live research
INVENTORY D09, F17–F18/H20–H22. Sources to check: frontend/components/optimize/OptimizerRunsBoards.tsx, frontend/components/RegimeSchedulePanel.tsx, frontend/lib/optimizer-page.ts, backend/app/scoring/weight_optimizer.py, regime_weight_optimizer.py and the existing serializers in backend/app/api/routes.py.
Read global saved records and read stage schedules. A specific claim needs its already-saved result/provenance; no production audit or optimization was performed.