금융 연구 재검증 집계

Coincident, Not Leading · 2026-10-18

Table 8의 US57 중앙 마진 6개와 Phase34 30개 모델의 재학습·수치 검산 결과입니다. MiDM 범용 벤치마크 또는 RT별 적중률과 구분하는 별도 연구 자료입니다.

계산 검증은 완료됐지만 원고 최종 반영과 정보시점·가격 인증은 남아 있습니다. 검증 기준 선택 모델의 95% 구간은 0을 포함하므로 안정적인 수익 우위를 확정하지 않습니다.

Table 8 — US57 median dominance margins

Signal	Market return	Down magnitude	Up magnitude
raw_original	0.01738634	0.03029602	-0.00850577
factor_residual	0.01307724	0.02792290	-0.03853288

57 models per cell; original Phase34 excluded. Median of model-specific near-band minus outer-band maximum absolute correlation. No new correlations or p-values.

Table 8 CSV

Phase34 — 30/30 complete

Test signals: 2025-01-02–2026-07-10; last exit: 2026-07-17. Three stocks per day; next-open entry; signal+5-session close exit; round-trip cost 0.30%.

Across-model mean next-open net trade EV: 0.794%; median: 0.881%. These are correlated descriptive estimates.

Validation-selected model: 34-3-US_base_plus_kr_market_mlp__mlp
Next-open net trade EV: 0.478%
95% date-block interval: -0.422% to 1.345%; AUC 0.9319.

Model (validation-score order)	Selected	Test AUC	Next-open net trade EV
34-3-US_base_plus_kr_market_mlp__mlp	✓	0.9319	0.478%
34-4-US_base_plus_kr_flow_ft__ft		0.9353	1.134%
34-2-US_base_plus_us_macro_ft__ft		0.9343	0.450%
34-6-US_macro_kr_market_ft__ft		0.9337	0.860%
34-10-US_macro_kr_full_wide_resnet__resnet		0.9308	1.221%
34-8-US_macro_kr_short_mlp__mlp		0.9300	0.891%
34-1-US_base_phase32_arch_mlp__mlp		0.9371	0.981%
34-8-US_macro_kr_short_resnet__resnet		0.9283	0.880%
34-4-US_base_plus_kr_flow_mlp__mlp		0.9343	0.552%
34-5-US_base_plus_kr_short_mlp__mlp		0.9308	0.760%
34-9-US_kr_full_cross_market_mlp__mlp		0.9326	0.961%
34-6-US_macro_kr_market_mlp__mlp		0.9345	0.812%
34-7-US_macro_kr_flow_mlp__mlp		0.9317	0.952%
34-2-US_base_plus_us_macro_resnet__resnet		0.9315	0.974%
34-10-US_macro_kr_full_wide_mlp__mlp		0.9341	0.916%
34-2-US_base_plus_us_macro_mlp__mlp		0.9298	0.968%
34-7-US_macro_kr_flow_resnet__resnet		0.9301	0.980%
34-6-US_macro_kr_market_resnet__resnet		0.9298	0.882%
34-1-US_base_phase32_arch_resnet__resnet		0.9257	0.604%
34-5-US_base_plus_kr_short_resnet__resnet		0.9326	1.141%
34-4-US_base_plus_kr_flow_resnet__resnet		0.9282	0.631%
34-9-US_kr_full_cross_market_resnet__resnet		0.9272	0.967%
34-8-US_macro_kr_short_ft__ft		0.9345	0.867%
34-3-US_base_plus_kr_market_resnet__resnet		0.9098	0.885%
34-1-US_base_phase32_arch_ft__ft		0.9338	0.774%
34-7-US_macro_kr_flow_ft__ft		0.9319	0.116%
34-9-US_kr_full_cross_market_ft__ft		0.9311	0.456%
34-5-US_base_plus_kr_short_ft__ft		0.9355	0.948%
34-3-US_base_plus_kr_market_ft__ft		0.9315	0.503%
34-10-US_macro_kr_full_wide_ft__ft		0.9320	0.291%

Audit scope / 검증 범위

30 × 189,766 saved held-out predictions were rechecked. Classification metrics, daily top3 selections, direct-price EVs and the close/open/overnight identity were reconciled. Five-session date-block intervals use 10,000 replicates; no p-value is inferred from bootstrap positivity.

EV is not portfolio NAV or benchmark alpha. Frozen features/eligibility, provider/corporate-action certification and identical-purged-row original-label counterparts remain unresolved. The original US57 cohort is not pooled with the new 30-model arm. This is an aggregate release; private underlying files and scripts are not uploaded.

30-model CSV · Full-precision summary · Audit checks · Method/results note · GitHub documentation