hjnormey/DOLPHIN
1
0
Fork 0
Code Issues Pull Requests 1 Actions Packages Projects Releases Wiki Activity

GREEN sprint: S6 asset picker + EsoF regime gate + AEM shadow-complete [S6 sizing suspended — overfitting] #1

Open
hjnormey wants to merge 6 commits from exp/green-s6-esof-aem-shadow-2026-04-21 into master
pull from: exp/green-s6-esof-aem-shadow-2026-04-21
merge into: hjnormey:master
Conversation 1 Commits 6 Files Changed 14 +1309 -30
hjnormey commented 2026-04-25 14:02:13 +02:00
Owner
Copy Link

Summary

GREEN S6/EsoF/AEM sprint — exp/green-s6-esof-aem-shadow-2026-04-21.

V7 stays authoritative (most likely cause of GREEN's ROI-positive delta vs BLUE). AEM stays shadow-only. All three sprint features are toggle-gated; one config flip reverts GREEN to pre-sprint behaviour.

Commits (6)

1 · ce7f3ce — AlphaAssetSelector bucket-ban kwargs (BLUE no-op)

alpha_asset_selector.py — Adds asset_bucket_ban_set / asset_bucket_assignments kwargs. Banned-bucket assets are skipped so the next-ranked asset takes the slot — capital preserved vs a 0× sizer which wastes the slot. Both default None → BLUE unchanged.

2 · 0da46d8 — Orchestrator single-site: S6 + EsoF + int-leverage gate

esf_alpha_orchestrator.py — All sizing mults applied at one location (_try_entry ~line 565). EsoF regime gate at _try_entry top; S6 per-bucket mult; integer-leverage gate (leverage_int=1 FIXED pending winrate analysis, leverage_raw preserved for CH); notional≤0→None guard; NDPosition extended with observability fields.

3 · af5156f — EsoF NEUTRAL→UNKNOWN rename + backward-compat alias

esof_advisor.py, esof_gate.py, prod/docs/ESOF_LABEL_MIGRATION.md — 637-trade retrospective shows mid-band is worst-ROI (signals in conflict, not neutral). NEUTRAL alias kept in all dicts so historical CH rows resolve.

4 · 48dcf3f — AEM per-bucket MAE_MULT + shadow logging completeness

adaptive_exit_engine.py, dolphin_actor.py — Shadow-only. MAE_MULT_BY_BUCKET: B3→None (natural winners, disables stop), B4→2.0 (strict), B6→6.0 (extreme vol). Shadow schema extended with Nullable columns; log_shadow() carries V7 head-to-head + naive counterfactual for future demotion analysis.

5 · 36d263e — GREEN config wiring + S6 recompute tooling

green.yml, green_s6_table.yml, prod/scripts/ — green.yml: ban_set=[4], s6_size_table, esof_sizing_table (UNKNOWN→0.25, UNFAV→0.0), use_int_leverage. blue.yml unchanged. Recompute script + Prefect flow (30d cadence, 20%-variance guard). Leverage winrate analysis script (read-only).

6 · aac4484 — GREEN-only feature tests

test_green_only_features.py — Toggles-OFF identity, selector ban slot-rerouting, AEM MAE table values, EsoF rename+alias, int-leverage still at 1x with option comments. Existing 104-test parity file left intact.

Kill-switch

# prod/configs/green.yml
asset_bucket_ban_set: null
s6_size_table: null
esof_sizing_table: null
use_int_leverage: false

Merge gate

Per prod/docs/BRANCHING_CONVENTION.md: merge to master only after integration replay reproduces expected counterfactual, parity tests pass, and 24h staging monitor is clean.

🤖 Generated with Claude Code

## Summary GREEN S6/EsoF/AEM sprint — `exp/green-s6-esof-aem-shadow-2026-04-21`. V7 stays **authoritative** (most likely cause of GREEN's ROI-positive delta vs BLUE). AEM stays **shadow-only**. All three sprint features are toggle-gated; one config flip reverts GREEN to pre-sprint behaviour. --- ## Commits (6) ### 1 · `ce7f3ce` — AlphaAssetSelector bucket-ban kwargs (BLUE no-op) `alpha_asset_selector.py` — Adds `asset_bucket_ban_set` / `asset_bucket_assignments` kwargs. Banned-bucket assets are **skipped** so the next-ranked asset takes the slot — capital preserved vs a 0× sizer which wastes the slot. Both default `None` → BLUE unchanged. ### 2 · `0da46d8` — Orchestrator single-site: S6 + EsoF + int-leverage gate `esf_alpha_orchestrator.py` — All sizing mults applied at one location (`_try_entry` ~line 565). EsoF regime gate at `_try_entry` top; S6 per-bucket mult; integer-leverage gate (`leverage_int=1` FIXED pending winrate analysis, `leverage_raw` preserved for CH); `notional≤0→None` guard; NDPosition extended with observability fields. ### 3 · `af5156f` — EsoF NEUTRAL→UNKNOWN rename + backward-compat alias `esof_advisor.py`, `esof_gate.py`, `prod/docs/ESOF_LABEL_MIGRATION.md` — 637-trade retrospective shows mid-band is worst-ROI (signals in conflict, not neutral). `NEUTRAL` alias kept in all dicts so historical CH rows resolve. ### 4 · `48dcf3f` — AEM per-bucket MAE_MULT + shadow logging completeness `adaptive_exit_engine.py`, `dolphin_actor.py` — Shadow-only. `MAE_MULT_BY_BUCKET`: B3→None (natural winners, disables stop), B4→2.0 (strict), B6→6.0 (extreme vol). Shadow schema extended with Nullable columns; `log_shadow()` carries V7 head-to-head + naive counterfactual for future demotion analysis. ### 5 · `36d263e` — GREEN config wiring + S6 recompute tooling `green.yml`, `green_s6_table.yml`, `prod/scripts/` — `green.yml`: ban_set=[4], s6_size_table, esof_sizing_table (UNKNOWN→0.25, UNFAV→0.0), use_int_leverage. `blue.yml` unchanged. Recompute script + Prefect flow (30d cadence, 20%-variance guard). Leverage winrate analysis script (read-only). ### 6 · `aac4484` — GREEN-only feature tests `test_green_only_features.py` — Toggles-OFF identity, selector ban slot-rerouting, AEM MAE table values, EsoF rename+alias, int-leverage still at 1x with option comments. Existing 104-test parity file left intact. --- ## Kill-switch ```yaml # prod/configs/green.yml asset_bucket_ban_set: null s6_size_table: null esof_sizing_table: null use_int_leverage: false ``` ## Merge gate Per `prod/docs/BRANCHING_CONVENTION.md`: merge to `master` only after integration replay reproduces expected counterfactual, parity tests pass, and 24h staging monitor is clean. 🤖 Generated with [Claude Code](https://claude.com/claude-code)
hjnormey added 6 commits 2026-04-25 14:02:14 +02:00
feat(s6): AlphaAssetSelector bucket-ban kwargs (BLUE no-op) ce7f3ce8ff 
Add `asset_bucket_ban_set` and `asset_bucket_assignments` kwargs to
AlphaAssetSelector.__init__ (both default None → BLUE unchanged).

When active, assets whose KMeans bucket is in the ban set are skipped
inside rank_assets() so the next-ranked asset takes the slot — capital
is preserved rather than wasted by a 0× sizer multiplier.

Plan ref: Task 9 — ban ≠ 0× is the critical caveat from the sprint plan.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
feat(s6/esof): orchestrator single-site S6+EsoF+int-leverage gate 0da46d8635 
All sizing multipliers now applied at one location (esf_alpha_orchestrator.py
line 565 region) so there are no hidden side-effects and BLUE parity is
trivially preserved by leaving the new kwargs at None/False.

- S6 per-bucket notional multiplier via s6_size_table kwarg
- EsoF regime gate at _try_entry top: reads _current_esof_label, skips
  entry if esof_sizing_table maps to 0.0 (UNFAV/MILD_NEG regime)
- Integer-leverage gate: use_int_leverage=True → leverage_int=1 (FIXED,
  pending prod/scripts/analyze_leverage_winrate.py analysis); float
  leverage_raw preserved in NDPosition + return dict for CH logging
- notional <= 0 → return None guard (prevents 0-notional positions)
- NDPosition extended: asset_bucket_id, s6_mult, esof_mult, esof_label,
  leverage_raw fields (BLUE leaves these at defaults)

Plan ref: Task 2 — sizer stays pure (its returned notional is discarded
by line 565 anyway; adding mults inside would be dead code with hidden
side-effects).

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
feat(esof): rename NEUTRAL→UNKNOWN + backward-compat alias af5156f52d 
The mid-band advisory label (constituent signals in conflict) was called
NEUTRAL, implying "benign middle" — but retrospective data (637 trades)
shows it is empirically the worst-ROI regime. Renaming to UNKNOWN makes
the semantics explicit for regime-gate consumers.

- esof_advisor.py: emits UNKNOWN; LABEL_COLOR keeps NEUTRAL alias for
  historical CH rows / stale HZ snapshots
- esof_gate.py: S6_MULT, IRP_PARAMS, Strategy A mult_map all keyed on
  UNKNOWN with NEUTRAL alias (values identical → replays unaffected)
- prod/docs/ESOF_LABEL_MIGRATION.md: migration note, CH/HZ impact,
  rollback procedure

Plan ref: Task 4 — NEUTRAL→UNKNOWN is load-bearing for the EsoF gate
in the orchestrator (0.25× sizing vs 1.0× under old label semantics).

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
feat(aem): per-bucket MAE_MULT table + shadow logging completeness 48dcf3fe13 
AEM stays shadow-only this sprint (V7 drives live exits); changes affect
what AEM *would have done*, logged to CH for future V7→AEM demotion analysis.

adaptive_exit_engine.py:
- Replace single MAE_MULT_TIER1=3.5 with MAE_MULT_BY_BUCKET dict
  (B3→None disables MAE stop, B4→2.0 strict, B6→6.0 wide band)
- evaluate() return dict extended: mae_mult_applied, mae_threshold,
  atr, p_threshold, giveback_k (all params that drove the decision)
- adaptive_exit_shadow schema extended (new Nullable columns added via
  ALTER TABLE IF NOT EXISTS — backward-compat with pre-sprint rows)
- log_shadow() signature extended: v7_action, v7_exit_reason,
  naive_would_have (TP/STOP/MAX_HOLD counterfactual at same instant)

dolphin_actor.py:
- AEM shadow call now passes V7 head-to-head decision and naive
  counterfactual so future retrospective requires no offline replay
- EsoF listener registered on DOLPHIN_FEATURES map (esof_advisor_latest
  key); label fed into engine._current_esof_label before each step_bar
- S6/bucket loaders (_load_s6_size_table, _load_asset_bucket_data) and
  constructor wiring for the new GREEN engine kwargs

Plan refs: Tasks 5, 7, 10 — V7 path untouched, AEM return value is
never gated, CH shadow is best-effort (daemon thread).

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
feat(config/scripts): GREEN config wiring + S6 recompute tooling 36d263eb91 
prod/configs/green.yml:
- asset_bucket_ban_set: [4] (B4 banned at selector level)
- s6_size_table: inline bootstrap multipliers (B0→0.4, B1→0.3, B3→2.0,
  B5→0.5, B6→1.5) matching CRITICAL_ASSET_PICKING S6 scenario
- esof_sizing_table: FAV→1.2, MILD_POS→0.6, UNKNOWN→0.25,
  MILD_NEG→0.0, UNFAV→0.0
- use_int_leverage: true (1x fixed pending winrate analysis)
- s6_table_path: pointer to generated YAML (recompute updates this)
BLUE (blue.yml) carries none of these keys → BLUE math unchanged.

prod/configs/green_s6_table.yml: bootstrap stub with frontmatter
(generated_at, source_branch, n_trades). Regenerated by recompute script.

prod/scripts/recompute_s6_coefficients.py:
  Queries trade_events, maps assets to KMeans buckets, derives per-bucket
  sizing mults. Variance guard: >20% net-PnL move flags bucket in
  dolphin.s6_recompute_log for manual review before promote.

prod/s6_recompute_flow.py:
  Prefect flow wrapping the recompute script. Cadence via
  S6_RECOMPUTE_INTERVAL_DAYS env (default 30). Kill-switch:
  S6_RECOMPUTE_DISABLED=1.

prod/scripts/analyze_leverage_winrate.py:
  Read-only walk of CH trade_events; bins trades by leverage_raw,
  emits per-bin WR/net-PnL/avg-MAE. Output informs the int-leverage
  rounding rule choice (Option 1 round-half-up vs Option 2 banker's round
  vs stay-at-1x). Does not auto-apply a rule change.

Plan refs: Tasks 3, 8, 10.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
test(green): add green-only feature tests aac4484c0f 
New test_green_only_features.py covers the sprint additions; the existing
test_green_blue_parity.py (104 tests) is left untouched as the authoritative
doctrinal baseline.

Tests:
- Toggles-OFF identity: with all GREEN kwargs at BLUE defaults (None/False)
  NDAlphaEngine + NDPosition produce the pre-sprint forms
- S6 selector ban: banned bucket assets skipped; slot rerouted (not wasted)
- AEM MAE table: B3→None disables stop, B4→2.0 strict, B6→6.0 wide,
  unknown bucket falls back to MAE_MULT_TIER1
- EsoF rename: advisor emits UNKNOWN not NEUTRAL; both label keys present
  in gate tables with identical values (historical replay safe)
- Int-leverage gate: source still has leverage_int=1 with both option
  comments present (guards against premature flip before winrate analysis)

Plan ref: Task 11.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
hjnormey commented 2026-04-25 14:16:59 +02:00
Author
Owner
Copy Link

⚠️ Fix committed: S6 coefficients suspended (commit ea65e2d)

Root cause: S6 sizing multipliers (B3→2.0×, B6→1.5×, etc.) were overfitting to the ~600-trade research window ending 2026-04-19. ~100+ trades since that window showed regime reversal — bucket PnL rankings did not generalise out-of-sample.

Applied fix:

  • green.yml: s6_size_table: null — all non-B4 buckets now size uniformly at 1.0× (no-op)
  • s6_table_path commented out
  • B4 ban retained — structural exclusion (only gross-negative bucket, -$100 gross before fees). Not a time-window call.
  • AEM MAE_MULT_BY_BUCKET retained — grounded in asset vol characteristics, not point-in-time PnL.

Infrastructure intact. Routing layer, recompute script (recompute_s6_coefficients.py), and Prefect flow are all present and functional. Re-enable path: run recompute across multiple non-overlapping windows, confirm variance guard (< 20% bucket PnL move between runs) passes consistently, then set s6_table_path in green.yml.

System naming clarification (2026-04-22):

  • BLUE + old-GREEN (Python-based semi-shadow) = in profit, pre-S6, unaffected
  • new-GREEN (Nautilus Trader sprint target) = would have been the first system to run S6 coefficients live — avoided

Post-mortem note added to prod/docs/CRITICAL_ASSET_PICKING_BRACKETS_VS._ROI_WR_AT_TRADES.md.

## ⚠️ Fix committed: S6 coefficients suspended (commit `ea65e2d`) **Root cause:** S6 sizing multipliers (B3→2.0×, B6→1.5×, etc.) were overfitting to the ~600-trade research window ending 2026-04-19. ~100+ trades since that window showed regime reversal — bucket PnL rankings did not generalise out-of-sample. **Applied fix:** - `green.yml`: `s6_size_table: null` — all non-B4 buckets now size uniformly at 1.0× (no-op) - `s6_table_path` commented out - **B4 ban retained** — structural exclusion (only gross-negative bucket, -$100 gross before fees). Not a time-window call. - **AEM `MAE_MULT_BY_BUCKET` retained** — grounded in asset vol characteristics, not point-in-time PnL. **Infrastructure intact.** Routing layer, recompute script (`recompute_s6_coefficients.py`), and Prefect flow are all present and functional. Re-enable path: run recompute across multiple non-overlapping windows, confirm variance guard (< 20% bucket PnL move between runs) passes consistently, then set `s6_table_path` in `green.yml`. **System naming clarification (2026-04-22):** - BLUE + old-GREEN (Python-based semi-shadow) = in profit, pre-S6, unaffected - new-GREEN (Nautilus Trader sprint target) = would have been the first system to run S6 coefficients live — avoided Post-mortem note added to `prod/docs/CRITICAL_ASSET_PICKING_BRACKETS_VS._ROI_WR_AT_TRADES.md`.
hjnormey changed title from GREEN sprint: S6 asset picker + EsoF regime gate + AEM shadow-complete to GREEN sprint: S6 asset picker + EsoF regime gate + AEM shadow-complete [S6 sizing suspended — overfitting] 2026-04-25 14:17:00 +02:00
This pull request is broken due to missing fork information.
View command line instructions

Checkout

From your project repository, check out a new branch and test the changes.
git fetch -u origin exp/green-s6-esof-aem-shadow-2026-04-21:exp/green-s6-esof-aem-shadow-2026-04-21
git checkout exp/green-s6-esof-aem-shadow-2026-04-21
Sign in to join this conversation.
Reviewers
No Reviewers
Labels
No items
No Label
Milestone
No items
No Milestone
Projects
Clear projects
No project
Assignees
Clear assignees
hjnormey
No Assignees
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: hjnormey/DOLPHIN#1
No description provided.
Delete Branch "exp/green-s6-esof-aem-shadow-2026-04-21"

Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?