Hi Everyone,
This week bitcoin markets have been relatively stable trading between $420/BTC and $460/BTC. Currently, the price sits at $455/BTC. In the middle of the week, some bad news came out of China with BTC China no longer accepting CNY deposits and it didn't drop the price very much. By the next day, the price had already recovered from the small downturn. Bad news from China seems to no longer have an effect on the markets so barring any drastic developments, I see the markets consolidating for a rally soon. General market sentiment seems positive.
Today I'd like to talk about a measure of "order flow toxicity" called volume-synchronized probability of informed trading (VPIN).
A good explanation and application of VPIN to the bitcoin markets can be found here: http://bit.ly/1l0aEhG.
To give some background, order flow is considered "toxic" (from the perspective of the market maker) when it is largely comprised of informed directional trading as opposed to noisy, non-directional uninformed trading. In the face of a lot of informed trading, market makers lose money as the price moves against their inventory after they are adversely selected against by traders with better information. To compensate for this adverse selection (moral hazard), market makers go wide on their quotes resulting in larger spreads. As market makers recede into the order book, flow becomes even more toxic (since to trade against a progressively wider spread, you must be progressively more informed) and this feedback loop continues until market makers exit the market entirely and liquidity dries up resulting in a steep crash (in theory, by symmetry, the price could sharply skyrocket too but in practice this doesn't happen). What I've just described is a concept called probability of informed trading (PIN). VPIN adds on top of this concept by measuring time in volume instead of clock-time. What that means is that the time axis is discretized into equal volume-weighted ticks (e.g. 100 share volume) instead equal time ticks (e.g. 1 second). The authors argue that this is a better way to measure time because it gives equal weight to time periods with the same amount of information density even if the periods are different in terms of clock-time. VPIN is credited with predicting/forecasting the flash crash of May 6, 2010.
In any case, it would be interesting to see more extensive work done on this area as applied to bitcoin markets especially as more market makers enter the space. Already, there is, at least, one very distinct market making bot on Bitstamp which quotes $5 of size every 3 cents up and down the bids and asks with its best bid and best ask being a little over 1% away from each other (All the small .0111, .0109, etc. orders are from this bot):
Currently, the algo does not seem very sophisticated but as more complex and professional algos show up, a VPIN model might serve as a good early indicator of near-term crashes.
Unrelated to VPIN, I've been looking into applying a machine learning algorithm called support vector machine (SVM) to bitcoin market data. SVM is essentially a binary classification algorithm which, in this case, tries to determine a pattern in the feature space (this feature space can consist of any features which you think might explain bitcoin returns like lagged returns, volume, rolling-vol, day of the week, the number of times a positive emotional word is mentioned in the same tweet as the word "bitcoin", or whatever) for classifying upticks versus downticks. Preliminary results show a predictive accuracy of between 51% and 57% in classifying out-of-sample bitcoin returns. These results are very cursory due to dependence on how time is discretized (e.g. 1 min, 2 min, 5 min etc.), the technical caveat that there are really three classes (uptick, even, downtick) instead of just two so I've had to do a 'hacky' tweak on the training data, and the fact that the feature space is relatively small and simple for now. Also, I should mention that these types of algorithms are susceptible to black swan risk since they seek out patterns in historic data and would not be prepared for a unique and impactful event which has never happened before (e.g. Germany criminalizes the use of bitcoin, Amazon accepts bitcoin, etc..). Given all of that, I think there is still value to be had here as long as we keep in mind its assumptions and limitations.
Cheers,
Kevin & Team Buttercoin
Bitcoin Trading Made Easy | Buttercoin.com
Buttercoin currently in Private Beta, if you'd like access right now
apply here
.