Master-LGBM without ETF — Robustness Check

2026-05-19 · status: done · 34.0s

Hypothesis: If ETF flow carries no causal forward information (B4 finding: residual IC +0.075, CI includes 0), Master-LGBM without the etf_flow feature delivers a nearly identical R² lift over HAR-RV. A drop > 1 pp would be an alarm signal that the feature does carry lookahead after all.

Verdict: MASTER-LGBM ROBUST — Removing etf_flow changes the R² lift by only -0.14 pp (10.60 → 10.74). Both versions are well above the 3 pp promote gate. The Master-LGBM result is NOT driven by an ETF lookahead. Top features remain: log_rv_7d_ann, iv_ann, stablecoin_d7.

Key metrics

Approach

Replicate of the Master-LGBM setup (Exp #22) with a single change: feature etf_flow removed from the all-features set. Compares the OOS R² against the HAR-RV baseline and against the original LGBM-with-ETF. If the robustness check passes, the R² lift without ETF should be similarly high.

Pooled OOS comparison

Marginal contribution of the ETF feature

• LGBM-with-ETF lift: +10.60 pp over HAR-RV

• LGBM-without-ETF lift: +10.74 pp over HAR-RV

• Loss from removing ETF: -0.14 pp

Per-window R² (13 windows)

Top-10 features in the ETF-free LGBM (mean gain)