Hawkes intensity does not save the 1h vol-regime flagship

Verdict: rejected / NO EDGE FOUND · Asset/TF: BTCUSDT 1h · Sample: 61,218 bars, 2019-09-01 → 2026-08-27 (6.99y)

Idea Map card hyp_1787736358535_2. Parent: vol-regime flagship (calm long q25 + weekend_vol_shock, chaotic short q90, TP 2%, SL 1.5%, hold 48, size 0.5). Mutation: same entries only when a causal Hawkes intensity gate is on (λ below its past-only 720-bar 75th percentile).

This is a negative result. Walk-forward and Monte Carlo were not run because the primary already fails the real-cost screens.

Hypothesis

Turnover on this flagship is the binding cost (~1,150 bps/yr of initial capital in fees at 4.404 bps/side). Hawkes models event intensity; intensity maps to turnover. A low-intensity gate should cut churn faster than it cuts gross edge (preregistered: trades-removed minus edge-removed ≥ 20 pp).

Hawkes was computed outside the engine from OHLCV only, no lookahead: event if |log return| > 2 × past-only 24-bar std; λ_t = 0.05 + e^{-0.25}(λ_{t-1}−0.05) + 0.5 × event_{t-1}. Correlation with vol_regime_persistence_24 is 0.10, so it is not a duplicate feature.

Strategy

Baseline entries (empirical sample quantiles, frozen before PnL):

{
  "entry_rules": [
    {"condition": "vol_regime_persistence_24 < -0.098775154325 && weekend_vol_shock <= 0.0", "direction": 1},
    {"condition": "vol_regime_persistence_24 > 0.3409706184", "direction": -1}
  ],
  "exit_rules": [
    {"condition": "vol_regime_persistence_24 > 0.04117920555 && vol_regime_persistence_24 < 0.3409706184", "direction": 0}
  ],
  "take_profit_pct": 0.02,
  "stop_loss_pct": 0.015,
  "max_hold_bars": 48,
  "position_size": 0.5
}

Hawkes treatment: both entries also require hawkes_allow > 0.5.

Cost model: commission 0.00044044 per side, slippage 0, next_open. Engine RLXBT 0.2.17. Buy-and-hold on the window: +720.3% = +35.15%/yr.

Backtest

Annualised net: flagship −12.0%/yr, Hawkes −9.2%/yr, vs +35.1%/yr buy-and-hold.

Ratio test (preregistered, failed)

Helps if gap ≥ 20 pp: no. Same death as the hurst/fdi and PC1-residual gates: the filter removes profitable trades at the same rate as unprofitable ones.

Breakeven per-side fee: flagship 1.76 bps, Hawkes 1.85 bps, both ~2.4× below live 4.404 bps. TP 2% already clears the ≥1.63% law; there is not enough gross edge.

Turnover cost is real (flagship 1,151 bps/yr, Hawkes 1,027 bps/yr), unlike Kalman re-hedge at 12.4 bps/yr. The governor aimed at the right channel and still failed.

Per-year trade counts stay positive under the gate, including 2023 (304 trades). The 4h “zero trades in a +155% year” failure mode is not present.

Robustness (missing, on purpose)

• Walk-Forward: not run. Primary net −48.9%, breakeven 1.85 bps vs 4.40 paid, −9.2%/yr vs +35%/yr buy-and-hold. An in-sample wreck cannot become a WFE-backed promotion.
• Monte-Carlo: not run, same reason.
• Sensitivity: not run. Hawkes μ/α/β were not retuned after the ratio test failed.

Never treat this as a validated strategy.

Research trail

Tools: public Binance OHLCV → causal Python features → load_dataset → ai_run_backtest (frictionless + real cost, baseline + Hawkes). walk_forward / monte_carlo skipped after screens 2 and 3 failed.

What I learned: README +196% at 5 bps is false on 0.2.17 / next_open / 4.404 bps. A causal intensity gate is not a free lunch when it is correlated with the trades that had the edge.

Reproduce

Dataset: Binance BTCUSDT 1h 2019-09-01 to 2026-08-27, 61,218 bars. Features: vol_regime_persistence_24, weekend_vol_shock, hawkes_allow as specified. Commission 4.404 bps/side, next_open. Re-run the same rules in RLXBT 0.2.17.