Bitcoin Price Bubbles and Socio-economic Feedback Cycles

Hi Everyone,

This week bitcoin traded between $575/BTC and $600/BTC, ending the week at $590/BTC.  Volatility continues to be low for many weeks in a row.

News
  • Gavin Andreson made a recent post on Bitcoin's scalability: http://bit.ly/1kuO7iY.  By using invertible bloom lookup tables, allows the decoupling of block size from block propagation time by allowing nodes to share only the block headers instead of the full block.  This erases the incentive for mining pools to include fewer transactions with the intent of propagating the block through the network faster because the block is smaller.  This would matter if two pools find a block at roughly the same time.
  • Monacoin crashes; BitcoinDark on the rise, Cloakcoin falls after a jump last week; Stellar starts trading on exchanges and has seen a rise in market cap albeit not much movement in price.
    • On a side note, the way Cloakcoin works is by having transactions which can be conditional on future states of the blockchain.  For example Bob creates a transaction (to be released 4 blocks from now) which sends coins to Mixing Node A which releases only if Mixing Node A sends some similar amount of coins to Joe in one of the next 4 blocks.  Combining this with one-use stealth addresses and two mixing steps, supposedly anonymity is achieved.  I'm not yet sure if this works or not.
  • There are rumors that AQR and DE Shaw are looking at bitcoin markets.
A Paper on bitcoin price bubbles and socio-economic feedback cycles was published week: http://bit.ly/XGhjJE.  I found it very informative and it confirms some ideas I've had  about demand-side leading indicators to price movements like search volume and client downloads.  It was also very readable compared to some of that heavy crypto stuff and the results gave immediate real world insight.

Without getting too technical, it looks at the feedback loops between bitcoin price, bitcoin search volume (Google/Wikipedia), bitcoin word-of-mouth spreading (Facebook/Twitter), and bitcoin user adoption (client downloads/blockchain address analysis) by running a series of statistical/econometric tests.

The results are shown in this diagram:


Interpretation
  • Social Cycle: search volume goes up => word of mouth goes up => price goes up => search volume goes up
    • Hype begets hype (bubble behavior).
    • "Bitcoin's growing popularity leads to higher search volumes, which in turn result in increased social media activity on the topic of Bitcoin.  More interest encourages the purchase of bitcoins by individual users, driving the prices up, which eventually feeds back on search volumes."
    • This is mediated by the media reporting on price increases, which drives curiosity and greed, triggering their search activity.
  • Adoption Cycle: price goes up => search volume goes up => user adoption goes up => price goes up
    • Understanding begets adoption (organic growth).
    • "New bitcoin users download the client and join the transaction network after acquiring information about the technology.
    • "This growth in the user base translates into a price increase, as the number of bitcoins available for trade does not depend on demand, but rather grows linearly (more or less) with time."
  • Search-Price Dyad: negative external event happens => search volume goes up => price goes down
    • Negative news (e.g. goxings, bans, hacks) dissemination precedes market crashes.
    • Search activity responds faster to negative events than price drops.
    • People must figure out what's wrong before they can determine whether or not to sell.
    • 3 of the 4 largest price drops were preceded by the first, fourth, and eighth largest increases in Google search volume.
Other Findings
  • An increase of 10k client downloads leads an increase of $3.80 in price.
  • An increase of .01% in bitcoin tweets as a proportion of total tweets leads to an increase in price of $2.70.
  • Search levels experience sharp increases 2-4 days after price increases.
  • Word of mouth (through social media) increases 1-2 days after search volumes increase.
  • User growth (client downloads) also increases 1-2 days after search volumes increase.
  • There is little to no relationship between each of the variables beyond 4 days.
  • Bitcoin prices significantly deviated away from fundamental value (based on electricity costs during mining and mining efficiency) during the June 2011 bubble and the April 2013 bubble.
  • Price never dropped significantly below fundamental value.
    • This suggests the cost of producing bitcoins as a lower bound to their value.
    • This also suggests that at the current supply of bitcoins, the market always prefers them to the electricity used in their production.
  • Their model correctly identifies the sign of the 10 largest price increases and 9 of the 10 largest price drops.
***Technical Stuff***

Data is grabbed and bucketed into days.  Daily Google search volume was simulated out of weekly data and 3 months of the most recent daily data.  User adoption was, in one method, proxied by blockchain change address analysis and, in another method, by wallet client downloads.

From the data they take first differences of each time series to guarantee stationarity (Augmented Dickey-Fuller test, KPSS test).

Next a vector autoregression of lag 1 (VAR(1)) is done and yields significant p-values thus passing the Granger causality test (Note: This is predictive causality, not "true" causality in the narrative sense, but it is good enough for trading strategies).

Pairwise correlation analysis was also done and while the results show interesting insights, more complex 3-cycle relationships are hidden so the authors prefer VAR(1).

Both normalized and non-normalized results of VAR(1) are given.

Impulse response functions are created to show the impact of 1 standard deviation shocks of each variable to the other variables.

The Bayesian Information Criterion achieves a minimum for the model with lag 4 (Note: This is a measure of the tradeoff between model complexity versus predictive power.  It penalizes having more variables more so than the Akaike Information Criterion does.  Given that AIC is arguably superior to BIC (http://bit.ly/1nGao8C), using 5 or 6 lags might make for a more comprehensive model.

Possible refinements to the model:
  • Use hourly buckets instead of daily buckets
  • Use VAR(4) or VAR(5) instead of VAR(1)
  • Model fundamental value based on changes in mining technology (GPU=>FPGA=>ASIC) instead of based on a constant energy efficiency value since later technologies are more efficient per kilowatt.
  • Add supply side modeling with hashing rate vs difficulty and the block reward halving point.
Cheers,
Kevin & Team Buttercoin
Bitcoin Trading Made Easy | Buttercoin.com 

Buttercoin currently in Private Beta, if you'd like access right now apply here: http://bit.ly/1kSgHZg.