# Claude Validator: Wider-Universe Rotation Walk-Forward (Falsified)

Reviewer: `l_ClaudeBTCPhaseLiquidMoneyOps1`
Date: 2026-07-16
Hypothesis (Taras): the regime, not the tune, limits returns; a wider symbol
universe with pre-fold selection (rotation) should beat the active6 baseline
(+7.44% chained May->mid-Jul, carry model).

## Setup

- Universe: active6 (PYTH HYPE SOL SEI ROSE RENDER) + 10 fresh-backfilled
  symbols (GRT LDO BLUR XRP IMX UNI AVAX TIA CRV ENA), BingX USDT-M 1m,
  2026-01-15..2026-07-16 (`_reports\liquid_money_rotation_wf_20260716\
  fast_cache_rotation_extra10_20260115_now.npz`).
- Params per fold: causal tunes (same tuner, 1x/1x, fee 0.0005,
  slip 0.000938, min step 2.5). extra10 fully tuned on Feb->Apr; Feb->May
  completed only GRT/LDO before the job was stopped twice, so Jun/Jul folds
  use the freshest available causal (Apr-vintage) params for most extra10 --
  staler but leak-free. active6 uses proper per-fold tunes.
- Selection at fold start (causal): `sleeve30` = last-30d return of the
  carry-replayed sleeve before the fold; `mom30` = trailing 30d price return;
  K in {4,6,8}; plus `all16` and `baseline_a6`.
- Execution: carry model (replay from 2026-02-14, score fold segment),
  EW across selected sleeves, folds May / Jun / Jul(->16th), chained.

## Results (fold total%/MDD%, chained)

| rule          | May          | Jun          | Jul         | chained |
|---------------|--------------|--------------|-------------|---------|
| sleeve30_top4 | -10.13/-13.3 |  1.83/-6.7   | 2.11/-0.9   | -6.54%  |
| mom30_top4    |   1.53/-5.3  | -2.61/-6.7   | 2.06/-1.5   |  0.92%  |
| sleeve30_top6 |  -5.23/-6.3  |  0.12/-5.0   | 2.31/-0.8   | -2.93%  |
| mom30_top6    |  -1.45/-5.6  | -3.00/-5.4   | 3.16/-0.8   | -1.39%  |
| sleeve30_top8 |  -2.73/-4.3  | -1.00/-5.8   | 0.57/-1.4   | -3.15%  |
| mom30_top8    |   0.64/-2.4  | -5.78/-7.3   | 1.27/-1.0   | -3.97%  |
| all16         |  -1.72/-3.3  | -4.77/-5.8   | 1.25/-1.2   | -5.23%  |
| **baseline_a6** | **3.26/-5.2** | **1.70/-5.4** | **2.31/-0.8** | **+7.44%** |

## Conclusions

1. **Hypothesis falsified.** Every rotation rule and the full 16-symbol EW
   underperform the plain active6 baseline. The May fold is clean evidence
   (all params equally fresh there): all16 -1.72% vs active6 +3.26%.
2. `sleeve30` (chasing recent sleeve performance) is actively harmful
   (worst at K=4). Recent sleeve returns anti-predict next-month sleeve
   returns in this family -- consistent with the earlier finding that the
   per-wheel allocator adds little over identity weights.
3. The Jun-Jul regime hit the wider universe too; rotation among more
   symbols of the same family does not escape it. In-sample the extra10
   sleeves looked fine (BLUR +126%, LDO +91% on Feb->Apr), OOS they drag.
4. Caveat: Jun/Jul folds used Apr-vintage params for 8 of 10 extra symbols
   (tune jobs were stopped twice). This biases against rotation modestly,
   but the fully-fresh May fold already shows the effect.
5. Stopping here deliberately: with only 3 folds, testing more selection
   rules would be overfitting the validation itself.

## Standing Recommendation (unchanged, now better supported)

Active6 + monthly retune + continuous capital (carry) + EW, ~+2.9%/mo OOS,
MDD ~-5%, portfolio kill-switch at -10%. The in-sample +177.8% chart remains
an upper bound, not a forecast.

## Artifacts

- `_reports\liquid_money_rotation_wf_20260716\rotation_walkforward_results.csv`
- `_reports\liquid_money_rotation_wf_20260716\fast_cache_rotation_extra10_20260115_now.npz`
- `_reports\liquid_money_rotation_wf_20260716\tune_extra10_train_feb_apr|_feb_may` (feb_may partial: GRT, LDO)
- Prior context: REVIEW_liquid_money_feb_jul_retune_walkforward_20260716.md