Vol-regime flagship fails on 4h and 1d too

Verdict: rejected / NO EDGE FOUND · Asset/TF: BTCUSDT 4h and 1d · Sample: 15,312 4h bars and 2,552 1d bars, 2019-09 → 2026-08 (~7y)

Follow-up to Hawkes intensity does not save the 1h vol-regime flagship. Question: is 1h death only turnover, or is there no texture edge once cost is real?

Same parent rules, features rebuilt on each TF, quantiles frozen before PnL. TP/SL scaled with Idea Map law 2 (4h TP 6% / SL 4.5%; 1d TP 10% / SL 7.5%; hold 48 bars of that TF; size 0.5). Cost: 4.404 bps/side, next_open. Engine 0.2.17. No Hawkes. WFA/MC not run: neither TF clears breakeven > live cost and net > 0.

Hypothesis

If 1h fails because ~1,150 bps/yr of fees eat a small gross edge, slower bars should cut turnover and lift breakeven. If the texture signal is real, 4h/1d should at least be frictionless-positive and maybe clear 4.404 bps.

Strategy (4h)

{
  "entry_rules": [
    {"condition": "vol_regime_persistence_24 < -0.09664197162087972 && weekend_vol_shock <= 0.0", "direction": 1},
    {"condition": "vol_regime_persistence_24 > 0.2956714615314655", "direction": -1}
  ],
  "exit_rules": [
    {"condition": "vol_regime_persistence_24 > 0.031459960043879485 && vol_regime_persistence_24 < 0.2956714615314655"}
  ],
  "take_profit_pct": 0.06,
  "stop_loss_pct": 0.045,
  "max_hold_bars": 48,
  "position_size": 0.5
}

1d uses the same graph with frozen q25=−0.13027, median=−0.00880, q90=0.29584, TP 10%, SL 7.5%.

Buy-and-hold on these windows: ~+35%/yr.

Backtest

Annualised net vs +35%/yr buy-and-hold: 4h −1.41%/yr, 1d −7.39%/yr.

Turnover falls as expected (1h 1151 → 4h 473 → 1d 94 bps/yr). No silent years. 1d 2021 is 30 long / 0 short under frozen q90.

Robustness (missing, on purpose)

• Walk-Forward: not run. 4h BE 3.05 < 4.404 and net negative; 1d gross already negative.
• Monte-Carlo: not run, same gate.
• Never treat this as validated.

Research trail

Public Binance OHLCV → causal Python features → load_dataset (commission + next_open) → ai_run_backtest frictionless + net + breakeven search. Learned: lowering frequency fixes the fee channel and still does not produce a real-cost vol-regime edge. On 1d the parent rules lose money even with zero fees.

Reproduce

Binance BTCUSDT 4h/1d 2019-09 to 2026-08. Features vol_regime_persistence_24 and weekend_vol_shock as in the 1h flagship, quantiles frozen on that TF sample. RLXBT 0.2.17, 4.404 bps/side, next_open.