# ETH RSI2 Price TWAP deep evaluation diff

Parameter under test: `trend_sma=50, rsi_threshold=3.0, exit_rsi=55.0, stop_loss_pct=0.008, max_hold_bars=48, entry_offsets=(0.002, 0.005, 0.008), entry_valid_bars=3, fill_buffer=0.0`.

| evaluation | window | first_candle | last_candle | bars | trades | gross_return | maker_taker_net_return |
| --- | --- | --- | --- | --- | --- | --- | --- |
| main_report_10y_eval | total | 2019-12-25 00:00 | 2026-04-29 16:45 | 222500 | 321 | 582.29% | 255.57% |
| main_report_10y_eval | 3y | 2023-04-29 16:45 | 2026-04-29 16:45 |  | 321 |  | 90.59% |
| main_report_10y_eval | 1y | 2025-04-29 16:45 | 2026-04-29 16:45 |  | 321 |  | 45.84% |
| main_report_10y_eval | 6m | 2025-10-29 16:45 | 2026-04-29 16:45 |  | 321 |  | 24.50% |
| main_report_10y_eval | 3m | 2026-01-29 16:45 | 2026-04-29 16:45 |  | 321 |  | 24.72% |
| search_script_3y_eval | total | 2023-04-30 17:00 | 2026-04-29 16:45 | 105120 | 170 | 28.36% | 5.65% |
| search_script_3y_eval | 3y | 2023-05-01 05:30 | 2026-04-29 16:45 |  | 170 |  | 5.65% |
| search_script_3y_eval | 1y | 2025-04-29 16:45 | 2026-04-29 16:45 |  | 170 |  | 10.70% |
| search_script_3y_eval | 6m | 2025-10-29 16:45 | 2026-04-29 16:45 |  | 170 |  | 6.40% |
| search_script_3y_eval | 3m | 2026-01-29 16:45 | 2026-04-29 16:45 |  | 170 |  | 12.47% |

## Readout

- `main_report_10y_eval` matches `scripts/generate_ultrashort_report.py`: load up to 10 years, run the shared `explore_ultrashort` candidate gross, then apply maker_taker cost with `cost_weight` per closed trade.
- `search_script_3y_eval` matches `scripts/search_eth_price_twap_variants.py`: load 3 years, run its local price-TWAP segment, and deduct maker_taker cost inside each close.
- The prior search grid did not include the main-report ETH deep parameter. It searched `trend_sma in {80,160}`, `rsi_threshold in {5,8,10}`, `entry_valid_bars in {2,4}`, and offset sets `(0.001,0.003,0.005)` / `(0.003,0.006,0.009)`, not `trend_sma=50`, `rsi_threshold=3`, deep offsets `(0.002,0.005,0.008)`, `entry_valid_bars=3`.
- Therefore the sign mismatch is primarily parameter universe plus total-sample scope: the main report's cost table is 10-year total, while the search table was a different 3-year grid sorted by maker_taker `net_calmar` then `net_annualized_return`.

Adopt the main-report evaluation口径 for reporting selected strategies: shared `explore_ultrashort` runner, 10-year cached history for total cost sensitivity, and explicit recent horizons for 3y/1y/6m/3m.