Sentiment - ELEKTRON Analytics

Apathy ↔ Hype

Apathy ↔ Hype Index 0 = apathy · 100 = hype

How much the public cares about Bitcoin, on one 0–100 scale. Three equally weighted pillars: interest (Google searches for “bitcoin” and “buy bitcoin” worldwide, Wikipedia views of the Bitcoin article), activity (active addresses, spot trading volume, and Coinbase's App Store rank once it has a year of history) and mood (Fear & Greed). Pieces that grow with adoption are first measured against their own four-year growth trend, then every piece is ranked against its own last four years, so a reading says how hot today is for where Bitcoin is in its adoption, not against all of history. High readings have coincided with cycle tops and low readings with bottoms, but that is hindsight: whether the index tells you anything about the next 7, 30 or 90 days is tested in the Analysis section (series “Apathy ↔ Hype Index”).

—/100

loading…

ApathyNeutralHype

30 days —1 year —

Interest

—

Activity

—

Mood

—

BTC close · log

Apathy ↔ Hype Index · 0–100

Hype beyond price (index − price-momentum percentile)

Sources: Google Trends · Wikimedia pageviews (CC BY-SA) · Coin Metrics Community Data (CC BY-NC 4.0) · BTC spot volume (CryptoCompare) · Apple App Store RSS · Fear & Greed by alternative.me

Overview

Price and one sentiment series Same x-range, stacked

BTC close on a log scale above; the selected series below, on its own units. The Fear & Greed zones are the index's own classifications, read from the stored labels. Dotted verticals are curated events (app-store #1 days, BofA FMS "most crowded" months, IPOs, policy). Standardised overlay puts both on one z-scored axis instead, which is the only honest way to lay them on top of each other.

Series Standardised overlay Event lines

BTC close · log

Fear & Greed index: alternative.me · Price: btc_ohlcv

Retail

Retail interest composite

Equal-weight mean of z-scored components (each winsorised at the 5th/95th percentile): worldwide search interest, Wikipedia views, Fear & Greed, active addresses and the Coinbase Finance-chart rank. A component is dropped while it has under a year of data or has gone stale. Components are shown faint; the composite is the bold line.

The pieces Six retail proxies, latest value at right

Each panel has its own scale. Google Trends is a 0–100 index renormalised on every pull, so compare shape, not level, between panels. The Coinbase rank axis is inverted: up means closer to #1 in the US Finance chart; 101 means outside the top 100.

Trends · bitcoin · US

Trends · bitcoin · world

Trends · buy bitcoin · US

Wikipedia · Bitcoin · daily views

Active addresses · daily

Coinbase · US Finance rank (inverted)

Source: Google Trends · Wikimedia pageviews (CC BY-SA) · Coin Metrics Community Data (CC BY-NC 4.0) · Apple App Store RSS

Retail by the quarter From shareholder letters and 10-Qs

Coinbase consumer share of spot volume and consumer volume, Robinhood app crypto notional (quarter sum of the monthly figure) and Block's bitcoin revenue. Curated from filings; a missing quarter is left blank rather than estimated.

Institutional

Spot ETF net flows Daily bars, cumulative line

Net creations minus redemptions across the US spot Bitcoin ETFs, in USD. The cumulative line is the same unit on the right axis.

CME futures positioning CFTC Traders in Financial Futures

Net long minus short by trader group, in BTC notional (5 BTC per standard contract, 0.1 per micro), stacked around zero. Asset managers are mostly the ETFs' hedges and long-only funds; leveraged funds carry the basis trade and are structurally short. Price in grey for context only.

13F filers holding a spot ETF feed: loading…

Distinct institutional managers reporting at least one spot Bitcoin ETF position, and the subset that are pensions, endowments, foundations or sovereign funds, parsed filing by filing from EDGAR. A quarter is closed once its 13F deadline (45 days after quarter end) has passed; until then filings are still coming in, so a manager not yet seen is pending, not exited. Dotted verticals are curated allocator disclosures.

—

Filers · last closed quarter

—

Pensions, endowments, foundations, sovereigns

—

Reported value, USD

Corporate treasuries bitcointreasuries.net snapshot, daily

Public companies holding bitcoin, and how much. Collected as a daily snapshot, so the history starts when the nightly job did.

Public companies

BTC held

Capital by the quarter Venture, IPOs, institutional spot volume

Galaxy Research crypto VC dollars and deal counts, crypto IPOs counted from the curated event list, and Coinbase institutional spot volume.

Sources: Farside (ETF flows) · CFTC · SEC EDGAR 13F filings · bitcointreasuries.net · Galaxy Research · company filings

Attention — Bitcoin vs AI

Attention: bitcoin ÷ AI Wikipedia pageviews, 28-day sums, weekly

English Wikipedia pageviews of the Bitcoin article divided by those of Artificial intelligence (human traffic, 28-day rolling sums). Pageviews are absolute counts, so the level is meaningful: 2× means bitcoin drew twice the readers. When Google Trends data is available its bitcoin ÷ AI ratio is drawn dashed, rebased to the Wikipedia ratio on their first common week, since a ratio of two separately normalised 0–100 indices has no meaningful level of its own. The ChatGPT launch (30 Nov 2022) is marked; dotted verticals mark BofA Fund Manager Survey months where "long bitcoin" or "long semiconductors" was the most crowded trade.

Venture share AI share of US VC $ · crypto share of global VC $

Where the venture dollars went. AI's share of US deal value from PitchBook-NVCA (yearly, then half-yearly); AI's share of global funding from Crunchbase (quarterly); crypto's share of global funding is Galaxy's quarterly crypto VC divided by Crunchbase's global total, so the two sources' definitions differ and the level is approximate while the trend is not.

Hashrate vs its two-year trend EH/s, log-linear fit

Theoretical hashrate from difficulty against a trailing 730-day log-linear trend. Hashrate below trend is capital leaving mining; the tooltip gives the deviation in percent.

Beta to the Nasdaq-100 90-day rolling, up- and down-days split

BTC daily return regressed on the NDX return over a trailing 90-day window, and separately on up and down NDX days. A down-beta above the up-beta is the "risk asset in a sell-off" pattern.

How long search interest took to come back Worldwide monthly "bitcoin", all history

For each peak in worldwide monthly search interest: months until interest first regained a fixed share of the peak, and months until the next peak. Computed from the series, not typed in.

Source: Wikimedia pageviews (CC BY-SA) · Google Trends · PitchBook-NVCA · Galaxy Research · Crunchbase · BofA Global Research FMS (as reported) · yfinance (^NDX)

Analysis — does it lead price?

Setup

Pick a series and how it is made stationary. Lags, windows and horizons count observations of the aligned series (days at daily frequency, weeks at weekly). The prior from the literature is that price leads attention, not the other way round; the page is built to show that honestly.

Series Transform Range Frequency

Observations

—

ADF p / KPSS p

—

both should agree the transformed series is stationary

Out-of-sample R²

—

Campbell–Thompson, expanding window

Hypothesis tests run

—

every CCF lag, both Granger directions at each lag, ADF/KPSS and OOS — the more you run, the less one p-value means

Lag correlation Transformed series vs log return

Correlation between the transformed sentiment series and the price log return at each lag. Bars outside the grey band are beyond ±1.96/√n, before any correction for the number of lags shown.

◄ price leadssentiment leads ►

Forward returns by bucket Mean log return %, centred on the unconditional mean

Rows are buckets of the series (Fear & Greed uses the index's own classifications; other series use quintiles of the transformed value). Columns are forward horizons. Each cell shows the mean forward log return in percent and n / episodes; the colour is the difference from the unconditional mean for that horizon, gold below, purple above. Hover for the bootstrap 90% interval.

Rolling Spearman correlation

Rank correlation between the transformed series and the contemporaneous return over a trailing window. The band is ±1.96/√window: inside it, the window's correlation is indistinguishable from zero.

Granger causality Both directions, BIC lag and fixed lags

F-test that lagged values of one series help predict the other, on the stationary transforms. The BIC-chosen lag is reported beside fixed lags so the result is not the best of many. Read the two directions side by side.

Here be dragons Skip this if you trust the page. It is here so you can check it.

Waiting for the analysis response…

Levels lie. Price and every attention series here are near unit-root; correlating levels manufactures a relationship (Granger–Newbold). Price enters as log returns, sentiment as a difference, log-difference, rolling z-score or deviation from a trailing median, and ADF and KPSS are reported so you can see whether the transform worked.

Alignment. Daily series are joined on the UTC date; the Fear & Greed value stamped at midnight UTC describes the previous day, so contemporaneous tests shift it back one day. Weekly Google Trends points are labelled by the Sunday that starts the week and paired with the Saturday close, six days later; CFTC positions are as of Tuesday and paired with Tuesday's close. The exact rule used is printed above.

Fear & Greed is half price. By alternative.me's own description, volatility (25%) and momentum/volume (25%) are computed from price and volume, dominance from market caps, and the survey component is paused. A test of the index against returns is partly a test of returns against themselves.

Google Trends is renormalised on every pull. Each request is scaled so its own maximum is 100; values are comparable within one pull only. Every refresh re-pulls the whole window and the page always reads the newest complete pull. A ratio of two such indices is a ratio of two arbitrary scales.

Apathy ↔ Hype Index. Each component is first put on a daily calendar: the monthly worldwide Trends values enter on the last day of their month (when the month is complete) and carry forward, with Google's integer 0 read as 0.5; Wikipedia views are a 28-day sum; active addresses, spot volume, the Coinbase rank (negated, so #1 is high) and Fear & Greed are 28-day means. The components that grow with adoption (both searches, Wikipedia views, active addresses, volume) are then detrended: each day a straight line is fitted to the log of the last four years (1,461 days, today included) and the component becomes today's distance above or below that line. Fear & Greed and the rank are bounded, so they are used as they are. Each is then ranked against its own trailing four years; a component joins after two years of history. A pillar is the mean of its components, the index is 100 × the mean of the pillars, published only where at least two pillars exist and smoothed with a trailing 7-day mean. Every step looks backwards only, so a past reading never changes when new data arrive. Why detrend: ranked as raw levels, adoption grew faster than any four-year window can absorb — active addresses rose twentyfold from 2011 to 2015 — so the January 2015 bottom read as hype. Hype beyond price subtracts the trailing four-year percentile of BTC's 90-day log return.

Multiple testing. The lag chart alone runs dozens of correlations; the heat map runs buckets × horizons; Granger runs two directions × several lags. The count is printed above. A single p just under 0.05 among that many is what noise looks like.

Attributions: Fear & Greed index by alternative.me · Source: Google Trends · Wikipedia pageviews from the Wikimedia REST API, CC BY-SA · On-chain counts from Coin Metrics Community Data, CC BY-NC 4.0 · CFTC Traders in Financial Futures · SEC EDGAR 13F filings · blockchain.info · bitcointreasuries.net · Farside Investors. Informational and educational only. Not investment advice.