# LATE_NEWS - Do late events move results beyond what polls show?

Short answer: **there is some evidence late events move opinion a point or two in specific cases, but no evidence we can use to build a predictive late-news term for 2026.** The effect is small relative to the shared polling error (about 2.4 pts SD between cycles), the events are few, and attribution is confounded by trends already under way.

## 1. Evidence reviewed

- **2016 Comey letter (Oct 28).** FiveThirtyEight's popular-vote projection of Clinton lead fell from 5.9 to 2.9 pts over the following week, and in swing states from 4.5 to 1.7 by Nov 4; Silver judged the data fit an effect of about 1-2 or 3-4 pts [WEB, secondary: https://www.npr.org/2017/05/05/526936636/pollsters-find-at-best-mixed-evidence-comey-letter-swayed-election and reporting summarised there]. AAPOR's 2016 committee: "at best mixed evidence" that the announcement tipped the race (same NPR link). A political-science paper using national polls found a small, statistically non-significant drop (same coverage; paper not located, so UNVERIFIED as a citation). Trump was also already gaining before the letter, and some field periods ended before it was public. Conclusion: **not identifiable from polls; plausibly 0-3 pts, one event, one cycle.**
- **Late deciders (2016).** AAPOR 2016: Wisconsin voters who decided in the last week broke for Trump nearly 2:1 (exit polls) [WEB, secondary: https://www.pewresearch.org/short-reads/2017/06/16/qa-pew-research-centers-president-on-key-issues-in-u-s-polling/]. But AAPOR 2020 found late deciders did not explain 2020 [WEB: https://www.pewresearch.org/short-reads/2021/07/21/a-conversation-about-u-s-election-polling-problems-in-2020/]. Late-decider direction is not stable.
- **Debates, scandals, other October surprises.** [MEM] The political-science literature generally finds debates and conventions produce short-lived movements that mostly fade; I did not retrieve a specific citation this session, so treat as UNVERIFIED. I am not going to cite specific papers I have not opened.

## 2. A test we could run now (poll-only proxy) [RUN]

Script `late_shift.py`, output `out/late_shift.txt`. For 807 races over 14 cycles (1998-2024), I took the poll average 15-28 days out ("early") and 0-14 days out ("late"), and regressed (result minus late average) on (late minus early).

- Pooled slope **-0.31** (cluster-bootstrap 95% CI -0.40 to -0.25); negative in 13 of 14 cycles.
- Reading: **late movement partly reverses**, not continues. This does NOT show that late news is irrelevant, and it should not be read as a news effect. It is mostly mechanical: the late average's sampling noise enters (late minus early) positively and (result minus late) negatively. It argues for averaging over a longer window and against chasing the latest poll.
- Leave-one-cycle-out, adding a momentum term gives MAE 4.10 vs 4.25 for the bare late average, which says "use more data", consistent with just averaging 0-28 days.
- Net: poll trajectories contain no usable evidence for a late-news term.

## 3. Proposed GDELT features (design only; **not tested, no GDELT data pulled**)

GDELT DOC 2.0 API supports `timelinevol` (share of monitored coverage), `timelinevolraw`, and `timelinetone` (average tone), no authentication required, but it exposes only a rolling window of about 3 months [WEB, secondary: https://pypi.org/project/gdeltdoc/ and https://blog.gdeltproject.org/doc-2-0-updates-1-5-year-searching-and-updated-mobile-interface/ ; I did not read GDELT's own spec page]. So **historical backtesting needs the raw GDELT file archives or BigQuery**, whose coverage start dates I did not verify (UNVERIFIED; GDELT 2.0 reportedly begins in 2015 [MEM]). That would give at most 5 comparable even-year cycles (2016, 2018, 2020, 2022, 2024), versus our 14 poll cycles.

Candidate features, per race r, computed over days -14..-1:

1. `vol_shock_R = z-score of (candidate/race coverage volume in last 14 days) vs its own 60-day baseline`.
2. `tone_shock_R = z-score of mean tone in last 14 days vs baseline` for each major candidate; feature = difference between D and R candidates (antisymmetric by construction).
3. `topic_shock`: share of coverage matching scandal/investigation/indictment terms (e.g. keyword lists), same z-scoring.
4. National analogues for the shared factor.
   Constraints that must be respected: features must be antisymmetric in party (flip D/R and the sign flips), computed only from data available at forecast time, and pre-registered before looking at outcomes.

**Power check.** With ~5 cycles, ~50-100 competitive polled races each, and an effect probably under 1 pt per SD of shock, the SE on a pooled coefficient will be larger than any plausible effect. Tone is also endogenous: media tone follows polls and events, so a negative tone shock often just reflects a candidate already falling. I expect an unidentifiable result and recommend the feature be shown, if at all, as an annotation ("news volume spike detected"), with **zero weight in the headline number** for 2026.

## 4. Verdict on late news

- Evidence for a measurable, predictable late-news effect beyond polls: **weak to none**.
- Evidence for the existence of occasional 1-3 pt late movements: **suggestive** (Comey; late deciders in 2016), but unrepeatable.
- Correct treatment in the model: widen the variance for days-to-election > 0 (a random-walk drift term), not a signed predictive term. See MODEL.md section 6.