Overview & Experiments 55 Synthesis Roadmap Lookahead Audit

HAR-RV-J — Jump-augmented Vol Forecast

Dropped

2026-05-19 volatilityjumpsforecasthar-rv-j

Hypothesis

Decomposing trailing realized vol into a continuous (bipower variation) and a jump component improves out-of-sample log-vol forecast quality vs the production HAR-RV model by ≥3 pp R²(log) at the 4h horizon.

Verdict

**DEAD** — jump decomposition adds +0.22 pp R² (effectively zero). Vol persistence is already captured by trailing RV; jumps add no separate signal at this cascade structure.

best_model

HAR-RV-J

R2_4h_HAR_RV

+0.6571

best_lift_pp

+0.2194

n_windows_4h

R2_4h_HAR_RV_J

+0.6593

IC_4h_HAR_RV_CJ

+0.7936

R2_4h_HAR_RV_CJ

+0.6541

har_rv_j_wins_4h

R2_4h_persistence

+0.5470

har_rv_cj_wins_4h

HAR-RV-J — Jump-augmented Vol Forecast

2026-05-19 · status: dropped · 5.6s

Hypothesis: Decomposing trailing realized vol into a continuous (bipower variation) and a jump component improves out-of-sample log-vol forecast quality vs the production HAR-RV model by ≥3 pp R²(log) at the 4h horizon.

Verdict: DEAD — jump decomposition adds +0.22 pp R² (effectively zero). Vol persistence is already captured by trailing RV; jumps add no separate signal at this cascade structure.

Key metrics

metric	value
R2_4h_persistence	`+0.5470`
R2_4h_HAR_RV	`+0.6571`
R2_4h_HAR_RV_J	`+0.6593`
R2_4h_HAR_RV_CJ	`+0.6541`
best_lift_pp	`+0.2194`
best_model	`HAR-RV-J`
IC_4h_HAR_RV_CJ	`+0.7936`
har_rv_j_wins_4h	`18`
har_rv_cj_wins_4h	`10`
n_windows_4h	`21`

Approach

Decompose realised vol RV at each cascade scale (1h, 4h, 1d) into a continuous component (bipower variation BV) and a jump component J = sqrt(max(RV² − BV², 0)). Compare four models on out-of-sample log(σ_{t+h}) prediction:

Persistence — forecast = log(rv_h) (zero parameters, baseline)
HAR-RV — log RV regressed on log(rv_1h, rv_4h, rv_1d) — current production
HAR-RV-J — HAR-RV plus one jump feature log(1+J_h)
HAR-RV-CJ — full continuous/jump decomposition: log(bv_) + log(1+J_)

Walk-forward: 12mo train / 3mo test, embargo = max(1440, horizon) minutes. Hourly resampled panel, σ-scale features (consistent with ml/forecast/).

Pooled OOS comparison

horizon	model	IC	R2_log	MAE_log	n
1h	persistence	0.8218	0.6433	0.2817	45,984
1h	har_rv	0.8363	0.7041	0.2593	45,984
1h	har_rv_j	0.8369	0.7051	0.2587	45,984
1h	har_rv_cj	0.835	0.695	0.2608	45,984
4h	persistence	0.7457	0.547	0.3123	45,984
4h	har_rv	0.7938	0.6571	0.2725	45,984
4h	har_rv_j	0.7949	0.6593	0.2716	45,984
4h	har_rv_cj	0.7936	0.6541	0.2722	45,984
1d	persistence	0.7307	0.4429	0.289	45,984
1d	har_rv	0.7548	0.5735	0.2533	45,984
1d	har_rv_j	0.753	0.5728	0.2539	45,984
1d	har_rv_cj	0.7529	0.5733	0.2537	45,984

Per-window R²(log) by horizon

1h horizon (21 windows)

window	n	pers_r2	har_rv_r2	har_rv_j_r2	har_rv_cj_r2
2021-01-02 → 2021-04-02	2160	0.4973	0.6054	0.606	0.3879
2021-04-02 → 2021-07-02	2184	0.5301	0.628	0.6325	0.6156
2021-07-02 → 2021-10-02	2208	0.1294	0.3372	0.3379	0.334
2021-10-02 → 2022-01-02	2208	0.1149	0.308	0.3105	0.2874
2022-01-02 → 2022-04-02	2160	0.3312	0.4707	0.4731	0.4523
2022-04-02 → 2022-07-02	2184	0.573	0.6483	0.6499	0.6511
2022-07-02 → 2022-10-02	2208	0.2065	0.3947	0.3949	0.3941
2022-10-02 → 2023-01-02	2208	0.5698	0.6523	0.653	0.652
2023-01-02 → 2023-04-02	2160	0.5591	0.6299	0.6315	0.6307
2023-04-02 → 2023-07-02	2184	0.2926	0.4323	0.4356	0.4403
2023-07-02 → 2023-10-02	2208	0.5133	0.5888	0.5912	0.5891
2023-10-02 → 2024-01-02	2208	0.4765	0.5561	0.5591	0.5576
2024-01-02 → 2024-04-02	2184	0.5543	0.6245	0.6245	0.6244
2024-04-02 → 2024-07-02	2184	0.5624	0.6319	0.6342	0.6329
2024-07-02 → 2024-10-02	2208	0.4395	0.5361	0.5355	0.5367
2024-10-02 → 2025-01-02	2208	0.1796	0.3951	0.388	0.3212
2025-01-02 → 2025-04-02	2160	0.6726	0.6987	0.6989	0.689
2025-04-02 → 2025-07-02	2184	0.5673	0.6249	0.6251	0.6122
2025-07-02 → 2025-10-02	2208	0.4877	0.5662	0.5663	0.5572
2025-10-02 → 2026-01-02	2208	0.5715	0.6216	0.6216	0.6115
2026-01-02 → 2026-04-02	2160	0.6049	0.66	0.6613	0.6585

4h horizon (21 windows)

window	n	pers_r2	har_rv_r2	har_rv_j_r2	har_rv_cj_r2
2021-01-02 → 2021-04-02	2160	0.4437	0.5786	0.5792	0.4676
2021-04-02 → 2021-07-02	2184	0.4803	0.5997	0.5974	0.5851
2021-07-02 → 2021-10-02	2208	-0.0523	0.245	0.2525	0.2528
2021-10-02 → 2022-01-02	2208	-0.1526	0.1749	0.183	0.1736
2022-01-02 → 2022-04-02	2160	0.1401	0.3898	0.3888	0.3788
2022-04-02 → 2022-07-02	2184	0.4439	0.5933	0.5959	0.5966
2022-07-02 → 2022-10-02	2208	0.0069	0.2992	0.3043	0.3037
2022-10-02 → 2023-01-02	2208	0.4601	0.5875	0.5882	0.5891
2023-01-02 → 2023-04-02	2160	0.3349	0.5344	0.5356	0.5372
2023-04-02 → 2023-07-02	2184	-0.0179	0.2894	0.2928	0.2987
2023-07-02 → 2023-10-02	2208	0.3128	0.4717	0.4737	0.4751
2023-10-02 → 2024-01-02	2208	0.1795	0.4118	0.4174	0.4225
2024-01-02 → 2024-04-02	2184	0.3557	0.5255	0.5284	0.5253
2024-04-02 → 2024-07-02	2184	0.4076	0.5511	0.5533	0.5537
2024-07-02 → 2024-10-02	2208	0.2115	0.429	0.4292	0.4325
2024-10-02 → 2025-01-02	2208	0.3493	0.4797	0.4852	0.4236
2025-01-02 → 2025-04-02	2160	0.4689	0.5892	0.5895	0.5815
2025-04-02 → 2025-07-02	2184	0.3336	0.5176	0.5166	0.5083
2025-07-02 → 2025-10-02	2208	0.2716	0.4574	0.4578	0.4548
2025-10-02 → 2026-01-02	2208	0.3016	0.4802	0.4843	0.477
2026-01-02 → 2026-04-02	2160	0.3729	0.5424	0.5444	0.5417

1d horizon (21 windows)

window	n	pers_r2	har_rv_r2	har_rv_j_r2	har_rv_cj_r2
2021-01-02 → 2021-04-02	2160	0.337	0.4802	0.4635	0.4583
2021-04-02 → 2021-07-02	2184	0.3282	0.4753	0.4596	0.4684
2021-07-02 → 2021-10-02	2208	-0.256	0.0136	0.0159	0.0143
2021-10-02 → 2022-01-02	2208	-0.4862	-0.0738	-0.055	-0.056
2022-01-02 → 2022-04-02	2160	-0.0382	0.2321	0.2476	0.2476
2022-04-02 → 2022-07-02	2184	0.3928	0.5227	0.524	0.5231
2022-07-02 → 2022-10-02	2208	-0.2347	0.1715	0.1834	0.1889
2022-10-02 → 2023-01-02	2208	0.3488	0.4781	0.4736	0.4816
2023-01-02 → 2023-04-02	2160	0.362	0.4917	0.493	0.4941
2023-04-02 → 2023-07-02	2184	-0.3244	0.0455	0.0529	0.0604
2023-07-02 → 2023-10-02	2208	-0.221	0.1723	0.1573	0.1529
2023-10-02 → 2024-01-02	2208	-0.4131	0.068	0.0661	0.0828
2024-01-02 → 2024-04-02	2184	0.1763	0.3995	0.3977	0.3953
2024-04-02 → 2024-07-02	2184	0.0604	0.393	0.3914	0.3931
2024-07-02 → 2024-10-02	2208	-0.1023	0.2559	0.2558	0.2599
2024-10-02 → 2025-01-02	2208	-0.0013	0.3516	0.3485	0.3308
2025-01-02 → 2025-04-02	2160	0.0401	0.381	0.3732	0.3722
2025-04-02 → 2025-07-02	2184	0.1347	0.3895	0.3889	0.3835
2025-07-02 → 2025-10-02	2208	-0.3116	0.2209	0.2254	0.2229
2025-10-02 → 2026-01-02	2208	-0.2387	0.223	0.2249	0.225
2026-01-02 → 2026-04-02	2160	0.1702	0.4318	0.4156	0.4155

per-window R² 4h

lift over HAR-RV 4h