Bitcoin Market Update 5/16

Hi Everyone,

This week markets have been trading between $430/BTC and $460/BTC.  Continuing this pattern for the third week in a row, vol is at an all-time low this year (exchange volume is also very low).  This feels similar to the quiet, low vol periods of 2012.

This week in bitcoin news:
  • Bitpay raises $30m in a round led by Index Ventures.
  • Ebay and Paypal are actively considering bitcoin integration.
  • Circle releases a demo of their consumer product:
Circle is offering an interesting consumer product.  They will be offering instant deposits/withdrawals from credit/debit, instant buy/sell, insured BTC storage (multi-sig, encrypted, geographically distributed) and no fees for any of that.  Instant deposit/withdrawals from credit/debit will be a very nice feature.  Chargebacks will likely be a hassle for them but I'm sure they will take precautionary measures to ensure that chargeback risk is a minimum and then, when it does happen, they will probably just take the loss as a cost of doing business.  Instant buy/sell is also a good feature.  Most likely they will be outlaying capital with each of the exchanges they use for routing customer orders and front the customer's order with their own capital while the customer's funds are in transit (they will likely batch customer orders to minimize frictional costs like wire fees).  Their storage solution follows industry standard practices and will likely be good.  Insurance on the vault, of course, begs the question, who is the underwriter? (I wrote extensively about the challenges of vault insurance in one of my earlier market updates).  If only the hot wallet is insured, they can likely underwrite it themselves given the capital they have raised.  If they have an external party underwriter, the expected value of having insurance will likely be negative and this is most likely a PR move.  On the point of no fees, I think one of two things is happening: (1) either they are willing to hemorrhage VC money to get user growth and will worry about monetization later (not a bad strategy; many tech companies in the growth phase do this) or (2) they are actually charging their customers up on the spread when they buy/sell (arguably this isn't literally a fee but semantically it's the same as a fee).  All-in-all I'm interested to see how it plays out for Circle.

In world news, official Treasury International Capital (TIC) numbers on major foreign holders of US treasuries came out yesterday:  It shows Russia dumping about $20bn in March but more surprisingly Belgium (actually some private account located in Belgium) has been buying up a total of $200bn since October of last year making them the third largest holder of US treasuries as of March 2014.  So in December, the Fed tapered $10bn from monthly asset purchases ($85bn->$75bn), another $10bn in Jan ($75bn->$65bn) and another $10bn ($65bn->$55bn) in Mar all the while some private buyer in Belgium is buying up $200bn in treasuries?  The timing seems strange.  It seems to me that if this secret Belgian did not buy up treasuries, markets might have reacted more negatively to the taper which, at this point, looks artificially successful.

Last week I touched briefly on applying a support vector machine to bitcoin market data.  This week I was able to refine my algorithm and can now consistently achieving an out-of-sample predictive accuracy of about 62%-63%.  I've run a series of batch tests on  different blocks of data, different discretizations of time, different features and different kernels (See: For certain feature sets, frequencies, and kernels work better than others so I've settled on a combination I believe to be the best.  There is strong consistency in results from different blocks of training data so this gives me confidence that there is something real and valuable here.  I should mention that even though 62%-63% accuracy seems very good, the true effectiveness of improved accuracy scales down exponentially the higher your accuracy gets.  That means that the improvement from 50% to 55% accuracy gives greater benefits than the improvement from 55% to 60% accuracy.  This is largely due to there being fewer large up and down moves and more small up and down moves, assuming larger magnitude moves are easier to classify by the SVM (which they should be).  Next week I'll be looking into multi-class SVM and also figuring out a heuristic for magnitude estimation given that the SVM is correctly predicting direction.

Kevin & Team Buttercoin
Bitcoin Trading Made Easy | 

Buttercoin currently in Private Beta, if you'd like access right now apply here