GME attention surges (news-count spikes) vs forward 5-day returns
The heaviest-coverage sessions for GME over the roughly three-year window tended to mark local exhaustion rather than sustained momentum: the five top-attention days averaged about a -6.9% forward 5-day return versus +0.8% on non-event days. I examined minute bars resampled to daily closes, matched calendar news counts to trading sessions, and compared forward 5-day returns for the top-attention sessions (N=5) against the rest.
The gap is economically large (roughly -7.7 percentage points) but rests on a very small sample and a few extreme moves — surge-day volatility was actually lower than non-events, and one surge session dropped about 19%. The Welch t gives p ≈ 0.11, so the signal is suggestive, not definitive. Full methodology, charts, and the session-level results follow below.
For GME over the past ~3 years, do days with an abnormal spike in news-article count (attention surges) mark local exhaustion rather than momentum, with forward 5-day returns after the highest-attention days running below the baseline? Thesis: GME's heaviest-coverage sessions cluster at local tops, so big-buzz days are followed by weaker-than-average forward returns, debunking the retail reflex that more buzz means more upside.
How this was measured
Resampled GME minute bars to daily closes and computed forward 5-trading-day returns: close[t+5]/close[t]-1. Aggregated GME_news by calendar day (count of articles), then mapped each calendar date to the first trading session on or after that date; summed counts when multiple calendar days (e.g., weekend) map to the same session. Defined 'attention surges' as the highest-attention sessions by article count (top decile of non-zero news days, with a minimum of 5 sessions when available). Compared the forward 5-day return distribution on surge days to both non-event days and the unconditional all-days baseline via Welch's t-test (unequal variance).
The key numbers
Reading the numbers
Across 752 trading days there were 5 top-decile "attention-surge" sessions. Those five averaged -6.89% over the next 5 trading days versus +0.81% for non-event days, a -7.70 percentage-point gap, but the Welch t-test p≈0.109 means this difference is not clearly significant at 5%.
The charts
This histogram displays the five individual forward 5-day returns that followed attention-surge sessions (n=5). Four of the five observations sit on the negative side — the worst was -19.34% and the best was +4.22%, producing a mean of -6.89%. Because there are only five points, each bar really represents a single session, so the left-side concentration is suggestive of local exhaustion but fragile given the tiny sample.
The bar chart lines up mean 5-day returns: attention-surge days at -6.89% versus non-event days at +0.81% (baseline computed over N=742 non-event sessions) and all days at +0.76%. Visually the surge bar flips sign from a modest positive baseline to a sizable negative, a -7.70 percentage-point gap that supports the 'surges mark tops' narrative in plain economic terms. That said, the statistical test yields p≈0.109, so the gap could plausibly be sampling noise rather than a definitive effect.
Highest-attention sessions (top by news count)
| date | news_count | fwd_5d_return |
|---|---|---|
| 2025-06-11 | 5 | -0.0773 |
| 2025-03-26 | 3 | -0.1934 |
| 2025-05-28 | 3 | -0.0527 |
| 2025-12-10 | 3 | 0.0422 |
| 2025-12-17 | 3 | -0.0631 |
| 2025-10-16 | 3 | 0.0248 |
| 2025-12-09 | 3 | 0.0199 |
| 2026-01-21 | 3 | 0.0469 |
| 2026-01-26 | 2 | 0.0524 |
| 2026-01-30 | 2 | 0.0059 |
| 2025-08-11 | 2 | 0.033 |
| 2025-03-27 | 2 | -0.0543 |
| 2025-06-12 | 2 | 0.0482 |
| 2025-06-25 | 2 | 0.0188 |
| 2025-08-26 | 1 | 0.0236 |
| 2025-08-18 | 1 | -0.0199 |
| 2025-07-02 | 1 | -0.0273 |
| 2025-06-16 | 1 | 0.0175 |
| 2025-05-29 | 1 | 0.0055 |
| 2025-03-31 | 1 | 0.1046 |
| 2025-06-10 | 1 | -0.2115 |
| 2025-10-23 | 1 | -0.0373 |
| 2025-12-04 | 1 | -0.0489 |
| 2025-12-02 | 1 | -0.0455 |
| 2025-11-24 | 1 | 0.1163 |
| 2025-12-11 | 1 | 0.0349 |
| 2025-12-12 | 1 | 0.0617 |
| 2025-08-28 | 1 | -0.0031 |
| 2025-09-02 | 1 | 0.0622 |
| 2025-09-23 | 1 | 0.0179 |
| 2026-01-09 | 1 | -0.008 |
| 2025-12-24 | 1 | -0.0399 |
| 2025-12-15 | 1 | -0.0122 |
| 2026-01-13 | 1 | 0.0737 |
| 2026-01-12 | 1 | 0.0424 |
| 2026-01-27 | 1 | 0.0012 |
| 2026-01-28 | 1 | 0.0529 |
| 2026-01-29 | 1 | 0.0417 |
| 2026-02-02 | 1 | -0.0511 |
| 2026-02-03 | 1 | 0.0262 |
The takeaway
Short answer: yes — the heaviest-coverage GME sessions tended to precede weak, not strong, 5-day returns. The top-attention group (N=5) averaged a -6.89% forward 5-day return versus +0.81% on non-event days and +0.76% across all days, an edge of about -7.70 percentage points versus non-events. That gap shows up even though the surge-day volatility was smaller (surge std = 0.0840 vs non-event std = 0.1561), but it rests on only five surge sessions. The formal test gives Welch t ≈ -2.03 with p ≈ 0.11 — about an 11-in-100 chance this difference is luck — so this is suggestive but not conventionally significant. There are extreme single-day moves in the surge list (for example one session fell ≈19.34%), so one or two outliers could swing the mean. Practical takeaway: heavy-news GME days look more like local exhaustion than reliable momentum, but the evidence is a lean, not a slam dunk.
The fine print
- Tiny sample: surge group has only 5 sessions — thin evidence and sensitive to single-day outliers.
- Surge definition is arbitrary (top-decile with min 5); different cutoffs may change the result.
- Article counts include syndicated duplicates and do not measure unique narratives or sentiment.
- Forward-5D windows overlap across days, so test degrees of freedom are inflated and p-values are approximate.