| Method | Weight | Point | 80% lo | 80% hi | 95% lo | 95% hi |
|---|---|---|---|---|---|---|
| linear | 25.0% | +13.52% | +9.84% | +17.18% | +8.27% | +19.19% |
| monte_carlo | 25.0% | +12.78% | -3.46% | +31.11% | -11.37% | +41.34% |
| ar1 | 25.0% | +0.04% | -2.66% | +2.81% | -4.06% | +4.31% |
| random_walk | 25.0% | +0.00% | -13.88% | +16.12% | -20.43% | +25.67% |
| Factor | Loading | Ticker value | Contribution |
|---|---|---|---|
| value_score | +0.0495 | -0.7351 | -0.0364 |
| quality_score | +0.0192 | +3.0000 | +0.0577 |
| momentum_score | +0.0286 | +1.7383 | +0.0497 |
| lowvol_score | +0.0173 | -0.3625 | -0.0063 |
| revisions_score | +0.0236 | +3.0000 | +0.0709 |
| news_activity_score | +0.0402 | +0.0000 | +0.0000 |
| (intercept) | - | - | -0.0005 |
| Factor | Mean IC | ICIR | % positive | Cumulative attribution | n periods |
|---|---|---|---|---|---|
| value_score | +0.2563 | +2.690 | 100.0% | +0.5383 | 252 |
| quality_score | +0.1105 | +1.096 | 86.5% | +0.2401 | 252 |
| momentum_score | +0.1428 | +1.354 | 92.5% | +0.2460 | 252 |
| lowvol_score | +0.0878 | +0.887 | 82.1% | +0.1870 | 252 |
| revisions_score | +0.1375 | +1.319 | 88.9% | +0.2738 | 252 |
| news_activity_score | +0.3540 | +3.832 | 100.0% | +1.0893 | 252 |
| Horizon | n predictions | Directional accuracy | MAE (return) | RMSE | 80% CI hit rate | Pearson(pred, real) |
|---|---|---|---|---|---|---|
| 1d | 691 | 52.2% | 2.24% | 3.10% | 81.3% | -0.029 |
| 30d | 662 | 53.3% | 16.01% | 20.17% | 73.3% | -0.032 |
| 60d | 632 | 57.3% | 27.85% | 35.50% | 64.1% | -0.014 |
intraday_forecaster_v10_trained_lgbm.
horizon_conflict flag when |1h move| > 1%.
meta_labeler_v10_trained_lgbm.
meta_abstain flag is added to the critique. See AFML Ch.3.6 (Lopez de Prado 2018).
| Factor | Loading wi | Ticker z-value | +1 sd impact | -1 sd impact |
|---|---|---|---|---|
| value_score | +0.0495 | -0.735 | +0.0495 | -0.0495 |
| news_activity_score | +0.0402 | +0.000 | +0.0402 | -0.0402 |
| momentum_score | +0.0286 | +1.738 | +0.0286 | -0.0286 |
| revisions_score | +0.0236 | +3.000 | +0.0236 | -0.0236 |
| quality_score | +0.0192 | +3.000 | +0.0192 | -0.0192 |
| lowvol_score | +0.0173 | -0.362 | +0.0173 | -0.0173 |
| Scenario | Description | Stressed point | Delta vs base |
|---|---|---|---|
| rates_+100bps | Yield curve +100bps (lowvol -1sd, value -0.5sd) | +9.32% | -4.20% |
| recession_risk_off | Recession (momentum -1.5sd, lowvol +1sd, news -1sd) | +6.94% | -6.58% |
| quality_flight | Quality flight (quality +1sd, revisions +1sd, momentum -0.5sd) | +16.38% | +2.86% |
Point return formula (linear factor model):
r_hat = alpha + sum_i (w_i × f_i)
where w_i is the cross-sectional loading for factor i
(estimated from the cross-section of US large-cap names) and f_i is the
z-scored factor value for this ticker (Grinold-Kahn winsorization at ±3).
Per-factor contribution shown in the waterfall above is exactly
w_i × f_i (decimal-return units), so the bars sum (with intercept) to
the linear forecast point. The ensemble point combines linear / Monte Carlo /
AR(1) / random-walk per per-method-comparison table.
Bias mapping (composite z-score → LONG/NEUTRAL/SHORT): weighted sum of value (1.0), quality (0.7), momentum (1.0), low-vol (0.5), revisions (1.0), news-activity (0.6); thresholds ±1.0 with macro override (risk-off + LONG → reduce to NEUTRAL; risk-on-growth + SHORT → reduce to NEUTRAL). LONG → Playbook A (long-call / call-debit-spread / covered-call); SHORT → Playbook B (long-put / put-debit-spread / bear-call-spread); NEUTRAL → cash.
Sensitivity (one-factor-at-a-time tornado above): impact = w_i × 1sd
per factor (MSCI / Two-Sigma Venn convention). Stress scenarios are canned multi-factor
shocks (rates +100bps; recession risk-off; quality flight).
Factor decay uses exponential half-life per factor type:
catalysts 7d, news 14d, revisions 30d, momentum 45d, value 90d, quality 90d
(Tetlock 2007; Cohen-Malloy-Pomorski 2012). The heatmap above shows
contribution × exp(-ln(2) × t / half_life) on a 5-day grid.
Factor IC backtest: Spearman rank correlation between each factor's
cross-sectional ranking on day t and the realized h-day-forward
return, aggregated across the panel (Grinold-Kahn 2000 ch.4 fundamental law of
active management). Source for this report: synthetic_panel_252d_100tickers_seed42 (parallel runner).
Placebo audit (v6, GT-Score):
GT-score = 0.52
SPURIOUS — setup likely lacks real signal
.
Computed from 3 seed(s) across placebo kinds:
shuffled, gaussian_iid, garch.
Method: real directional accuracy = 52.2%;
max placebo accuracy = 54.2%;
AMG = -0.019.
Source: placebo_audit_v6_ar1_runner.
Provenance: arXiv 2604.15531 "Spurious Predictability in Financial ML" (Paper #3).
Risk-flag spurious_predictability_audit_failed added to narrative.