Follows reports/jev_verdict.md and reports/jev_deepdive.md. Two new
studies, both fully PIT-safe on local caches, no promotion claimed.
Study 1 — TimesFM quantiles / uncertainty (validation/timesfm_quantile.py)
The C6 challenger stores only mean_predictions; the model also emits 9
quantiles (0.1..0.9). Rebuilt all 78 equity 1D shadows with the distribution
and scored it (reports/timesfm_quantile.md):
| signal | mean IC | t | p |
|---|---|---|---|
| point forecast (baseline) | +0.0134 | +0.74 | 0.23 |
| P(up) from quantiles | +0.0068 | +0.40 | 0.36 |
| point / width (risk-adj) | +0.0111 | +0.65 | 0.26 |
- No selection lift. The quantiles do not improve cross-sectional IC; baseline is reproduced exactly (+0.0134 == C6).
- Uncertainty is calibration, not direction. corr(width, |realized − predicted|) = +0.28 (n=3900) — the interval width does discriminate the magnitude of forecast error. But the low-width half of the top forecast decile realized worse (+1.54% vs +2.20%, p=0.86), i.e. filtering on low uncertainty is useless/backwards for returns.
- Takeaway: keep the mean for selection. The one live use for quantiles is sizing/risk (scale by forecast width), not stock-picking — a separate, unproven idea.
Study 2 — Jev × TimesFM (validation/jev_timesfm.py)
Same 78 monthly dates × 50 covered names. Two independent cross-sectional
judges on identical PIT data. Jev's state is deliberately echo-free (macro +
earnings + raw trailing returns; no TimesFM forecast, rank, or
conviction — the deep-dive lesson).
| signal | mean IC | t | p |
|---|---|---|---|
Jev p | −0.0082 | −0.27 | 0.61 |
| TimesFM | +0.0134 | +0.74 | 0.23 |
| 126d momentum | −0.0089 | −0.27 | 0.61 |
| Jev residual over momentum | −0.0138 | −0.53 | 0.70 |
| ensemble rank-avg(Jev, TimesFM) | −0.0001 | −0.00 | 0.50 |
- Jev veto on the TimesFM top-8: paired D +0.10% (p=0.45, 44% pos) — nothing. The confirm filter fired too rarely to test.
- No incremental signal, no ensemble lift, no veto edge.
Combined verdict
Consistent with the prior Jev verdict: on strictly historical, news-less states, Jev adds nothing measurable, and TimesFM's discarded distribution adds no selection edge. Neither earns enforcement. This is not a refutation of the live sleeve — Jev's primary input (semantic news/theses) has no historical series, so history cannot test it; the live forward shadow remains the only binding evidence. Both models stay shadow-only.
Follow-ups worth a vote (not adopted)
- Forecast-width sizing (not selection): scale TimesFM-leg positions by inverse uncertainty. Needs its own study.
- Re-run Study 2 once ≥30 live Gemma artifacts exist, with news present — that is the first state where Jev could plausibly differ from the echo.
- If a cheap historical news proxy is ever assembled (price-volume headlines, analyst-surprise texts), it would unblock a real Jev history test; until then, don't spend more Jev backtest effort.