172 data points across 12 months. Three independent tests. One consistent result.

The market doesn't read Google.

FinBuzz is a narrative-testing engine. It took the most popular retail trading belief of 2025-2026 - that search attention moves stock prices - and tested it against a year of verified data. NVIDIA gained 18.8%. Google searches for "AI stocks" peaked, crashed, and peaked again. The two had nothing to do with each other. But the reason why turned out to be more interesting than the finding itself.

-0.151
Window 1 (2025)
NVDA +5.9%. Not significant.
-0.179
Window 2 (2026)
NVDA -0.4%. Not significant.
-0.062
Full year
NVDA +18.8%. Not significant.
0 of 16
Predictive lags
Zero consistent signals.
172
Data points
119 daily + 53 weekly.
Search attention vs. NVDA price (weekly, 53 weeks)"AI stocks"NVDA ($)
WEEKLY Jul 2025 - Jul 2026 | r = -0.062, p = 0.661 | NVDA $171 to $203 (+18.8%) | SHA 37227 / 05176 VERIFIED
Window 2: daily, Apr-Jul 2026 (60 trading days)"AI stocks"NVDA ($)
DAILY Apr-Jul 2026 | r = -0.179, p = 0.175 | NVDA $208 to $207 (-0.4%) | SHA 5de00 / 7f9fc VERIFIED
Window 1: daily, Aug-Nov 2025 (59 trading days)"AI stocks"NVDA ($)
DAILY Aug-Nov 2025 | r = -0.151, p = 0.258 | NVDA $129 to $136 (+5.9%) | SHA 0ab6c / 81440 VERIFIED
00 Contextwhat actually moved the stock

Every price move traces to a real event, not a search trend

The timeline below maps every major NVDA price move against the earnings reports, Fed decisions, regulatory shifts, and competitive threats that actually caused them. When search attention appeared to correlate with price, a news event was always the shared cause of both.

EarningsFed / MacroRegulationIndustry
AUG 27, 2025 earnings
NVDA Q2 FY2026: beat estimates, revenue +56% YoY
EPS $1.05 vs $1.01 expected. Revenue $46.74B. Stock fell 3.6% the next day despite the beat. Search interest was at 73 and didn't react. Source: SEC 8-K filing
SEP - DEC 2025 fed
Fed cuts rates three times, ending 2025 at 3.50-3.75%
The easing cycle supported growth stocks. NVDA rose from $107 to $136. Search interest declined from 76 to 18 over the same period. Price and attention moved in opposite directions. Source: federalreserve.gov FOMC statements
NOV 19, 2025 earnings
NVDA Q3 FY2026: beat again ($1.30 vs $1.25)
Falls in Window 1's final days. Stock dropped 2.2% post-earnings. Search interest was at its lowest (18). The earnings reaction happened independently of attention. Source: SEC 8-K filing
DEC 11, 2025 regulation
Trump signs Executive Order: "National Policy Framework for AI"
EO 14365 directed federal agencies to consolidate AI oversight and challenge state AI laws. Pro-innovation stance, broadly positive for AI companies. Source: Federal Register, EO 14365
FEB 25, 2026 earnings
NVDA Q4 FY2026: EPS $1.62 vs $1.54 expected
Captured by the full-year weekly view but between our daily windows. NVDA near $191. Source: SEC 8-K filing
MAR 18, 2026 fed
Fed holds at 3.50-3.75%, citing Iran conflict
First pause after five cuts. 11-1 vote. Powell cited elevated inflation and geopolitical uncertainty. NVDA near $182. Source: FOMC statement, March 2026
MAR 2026 regulation
White House proposes federal preemption of state AI laws
National policy framework recommended one federal standard replacing state-by-state AI regulations. Reduced compliance risk for AI companies. Source: White House policy brief
MAY 27, 2026 earnings
NVDA Q1 FY2027 earnings
Near the start of Window 2. Search interest was climbing toward 100. NVDA hit $235. Earnings drove the price; the search spike was a side effect of news coverage. Source: SEC 8-K filing
JUN 1, 2026 industry
NVIDIA announces Vera CPU, open-source AI agent tools
First in-house data center CPU. Customers include OpenAI, Anthropic, SpaceX. Search hit 100 the next day. The product launch caused both the attention spike and the stock move. Source: nvidianews.nvidia.com
JUN 11, 2026 fed
FOMC meeting (mid-Window 2)
Falls exactly when Window 2's rolling correlation crossed from positive to negative. The Fed decision moved the broader market. Source: FOMC calendar
JUN 22, 2026 regulation
EU delays key AI Act obligations
Provisional agreement to defer high-risk AI system deadlines. Reduced near-term compliance pressure. Source: EU Parliament provisional agreement
LATE JUN 2026 industry
OpenAI delays IPO to 2027
Signaled sector caution. NVDA fell 10.7% in June. Searches dropped from 85 to 51. Both reacted to the same news. Neither caused the other. Source: Reuters
JUL 20, 2026 industry
Chip stocks' worst week in a year; AMD launches rival AI system
AMD signed Microsoft and Meta. NVDA near $203. Search interest had collapsed to 25. The competitive threat drove the selloff. Source: CNBC
AUG 26, 2026 earnings
NVDA Q2 FY2027: $96.22B revenue, more than doubled YoY
EPS $2.22 vs $2.09 expected. Data center $89B (+117% YoY). Stock initially fell 1.8%, then rallied sharply the next day. Five consecutive beats, stock dropped after three of them. The earnings, not search attention, determined price direction. Source: SEC 8-K, NVIDIA Newsroom, CNBC
THE CONFOUNDING VARIABLESearch interest and price sometimes moved together in these windows. But every instance traces to a shared cause: a news event that drove both the searches and the price. The searches are an echo of the news. The price is a reaction to the news. Neither causes the other. Five consecutive earnings beats, and the stock dropped after three of them. If search attention predicted price, every beat should produce a rally. It didn't. A naive sentiment tracker reports the echo as a signal. FinBuzz identifies it as a confound.
01 Findingsreplicated three ways

Five things I can defend in an interview

-0.15 / -0.18 / -0.06Three correlations

Buzz and price moved independently, three times

All three correlations are negative, weak, and not statistically significant. The null result held across a rising market (+5.9%), a flat market (-0.4%), and a strong rally (+18.8%).

0 reliable16 lag tests

Attention never predicted tomorrow's price

Sixteen lag tests. Window 1 produced two with p < .05 but opposite signs (noise fingerprint). Everything else: not significant. Zero consistent predictive signals across any window.

+0.84 to -0.85Rolling correlation

The relationship reversed itself in every window

The rolling correlation swung from strongly positive to strongly negative in all three windows. A real signal holds steady. This one flipped constantly, the signature of two unrelated series.

Replicated3 market regimes

The null result survived bull, flat, and rally markets

Three different market conditions, same finding. Most junior projects test once. This one tested three times across two data granularities.

ConfoundIdentified

When they did move together, a news event caused both

Every co-movement between attention and price traces to an earnings report, Fed decision, or competitive threat. Five consecutive NVDA earnings beats across the analysis period - stock dropped after three of them. The searches didn't cause the price move. The news caused both.

02 Dashboardall windows side by side

The instability is the finding

If search attention drove price, the rolling correlation would stay positive. Instead it wanders from +0.84 to -0.85, crossing zero repeatedly in every window.

Rolling corr - full year
8-week rolling. -0.62 to +0.71.
Rolling corr - 2026
14-day rolling. +0.82 to -0.85.
Rolling corr - 2025
14-day rolling. -0.74 to +0.84.
Lag tests - all windows
16 tests. Two crossed p < .05 with opposite signs.
NORMALIZATION NOTEGoogle Trends scores are relative to each window's own peak (100 = highest point in that window). Scores are not comparable across windows. The full-year view shows the narrative arc. The daily windows are where the rigorous statistics live.
03 Reportcombined analysis
FinBuzz combined insightJul 2025 - Aug 2026

Tested once, it's a finding. Tested three times, it's evidence.

The full year. From July 2025 to July 2026, "AI stocks" search interest sat in the 27-50 range for eight months, then exploded to 100 in late May 2026 before crashing to 27 by mid-July. NVDA climbed from $171 to $203 (+18.8%). The weekly change correlation was r = -0.062, essentially zero.

Window 1 (Aug-Nov 2025). Attention peaked at 100 in October, then faded to 18 by November. NVDA climbed from $129 to $136 (+5.9%). r = -0.151, not significant.

Window 2 (Apr-Jul 2026). Attention spiked to 100 twice, then collapsed to the low 20s. NVDA ended flat: $208 to $207. r = -0.179, not significant.

Three windows, three attention stories, three market conditions, one result: retail search buzz does not move this stock.

The confounding variable. Every apparent co-movement between search interest and price traces to a shared cause: an earnings report, a Fed decision, or a competitive event. On June 1, 2026, NVIDIA announced the Vera CPU. Search interest hit 100 the next day. The stock was near $224. Both the search spike and the stock reaction were caused by the product launch. Remove the news, and the correlation disappears.

Five beats, three drops. Across five consecutive quarterly earnings beats spanning August 2025 to August 2026, NVDA's post-earnings price direction was inconsistent. The stock fell after three of those five beats despite massive revenue growth (from $46.7B to $96.2B in revenue over one year). If search attention predicted price direction, every earnings-driven search spike should produce a rally. It did not. The earnings drove the searches and the price. The searches were bystanders.

What this is not. Not investment advice. Not a trading system. Not a universal claim. One keyword, one stock, three windows. The finding is narrow, honest, and reproducible.

04 Methodhow to check the work

Every number is reproducible

  1. Sources. Google Trends exports (US, web search). Yahoo Finance and yfinance adjusted close. Events sourced from SEC EDGAR 8-K filings, Federal Reserve FOMC statements, Federal Register executive orders, NVIDIA newsroom, Reuters, CNBC.
  2. Alignment. Daily: inner join on date, weekends and holidays removed. Weekly: Sunday trend scores matched to nearest trading day close.
  3. Statistics. Correlations on daily percentage changes, not raw levels. Each computed two independent ways (scipy + pandas). Pipeline refuses to publish if they disagree.
  4. Replication. Three non-overlapping analyses across rising, flat, and rallying market regimes.
  5. Confounding analysis. Each co-movement cross-referenced against the sourced event timeline. Five earnings events verified against SEC 8-K filings.
  6. Normalization. Google Trends scores are relative per window. Not comparable across windows.
  7. Sampling variance. Two pulls of the same query disagreed on 31 of 59 dates in Window 1.
  8. Multiple comparisons. 16 lag tests. Two crossed p < .05 with opposite signs = noise.
05 Aboutwhy this exists

A self-assigned analyst internship

I built FinBuzz to do the work before anyone hired me to do it: collect real data, verify it, analyse it honestly, and publish the result.

Then I re-ran the analysis on fresh data, identified the confounding variable most sentiment trackers never notice, and checked whether the finding held. It did, three times.

What this demonstrates

Data discipline - provenance logs, file hashing, integrity gates.

Statistical judgment - percentage changes over levels, lag testing, calling noise "noise."

Confounding analysis - identifying the shared cause behind apparent correlations.

Replication mindset - three tests, two granularities, three market regimes.

Communication - findings for non-technical readers, caveats for technical ones.