11.5 When do bitcoin traders sleep?

0. Introduction
At the moment I'm trying to relate bitcoin trading to other forex trades, to find statistical similarities and differences. One of the things I need to do is compare the difference in the log of the daily USD / BTC VWAP to the log daily return, ie. log(close / open).

BTC exchanges run continuously; how do you define the "open" and "close"? I wanted to use a time of day with the least variance and I suspected that would be the time when least trades occurred. Since the trades I'm interested in are for USD I thought that early morning east coast US would probably be the time of fewest trades and least variance.

Instead of using a method with provable results, I decided to "do it by eye". This post contains those charts, and some other interesting items about the way trades change during the day, and how this has also changed over time.

1. MTGOX traders don't sleep at all anymore.
The chart below is a 2d histogram of the number of trades for a particular time of day (UTC) from April 2011 until now (I chose that start date as the first day BTC traded for more than US$1.10, a rough indicator that trading had become consistent). The deeper / redder the shade, the fewer trades occurred.

It's clear that until the start of this year, the fewest trades occurred around 8:00 to 12:00 UTC. After the start of this year, trades are fairly constant throughout the day - no dip in trading at any time. I speculate that this is either an indicator of the increase in unsupervised tradebots, geographically wider use of MTGOX than just US citizens, or both.


Even though it's clear that the 8:00 to 12:00 dip disappeared at the start of the year, I've included all the data in the next plot. It shows the density of the percentage of trades per day,  per hour of the day (UTC).  It dips significantly at about 10:00 UTC, peaks from 18:00 to 3:00 UTC.

For an even clearer indication of the daily reduction in trading, the chart below points to 10:00 UTC as the mean minimum. I was expecting to find two mean minimums due to changes to and from daylight saving time, but this seems not to be the case. It also confirmed my speculation that the time of least trading would also be the time of least variance.

So, I'm done, yes? Well as far as my original aim goes, yes. However having a bunch of data (thanks to bitcoincharts.com) I decided to look at a few other USD BTC exchanges.

2. Neither do BTC traders on other exchanges, apparently (sleep, I mean).
In this section, I'm reproducing the above visual analysis - initially for all USD BTC exchanges with recent trades, and then for the most consistently long term trading USD BTC exchanges.

The chart below is equivalent to the first chart in the last section. I've changed the colours to a heat map since there's a wide variation in the difference between trade densities.

MTGOX (based in Japan) and CampBX (based in the US) have minimum trades at similar times, as do BTCe (Russia) and Bitstamp (UK). The other exchanges have too few trades to show a clear indication of times of least trading. Interestingly, while BTCe and Bitstamp also show a reduction in the daily minimum since the start of the year, CampBX does not show such a significant change - there's still a time of day when the least number of trades consistently occur.

The next chart is equivalent to the second chart of the first section, showing the density of the percentage of trades per day, per hour of the day (UTC) for the four larger and more consistently trading exchanges, except using heat maps colours for the same reason as previously. The exchanges that seem to have the same time of trading minimum have been lined up vertically for easy comparison.

The trading times do look quite similar for the members of the two groups, and quite different in comparison. There is definitely a preponderance of traders from specific locations.

This shows the minima and maxima more clearly:

It's interesting that Bitstamp has a different cycle to the others, and has a much higher standard deviation too. This variance is probably part of the reason that the trading maxima times are so flat. The variance at BTCe doesn't trend the way it does for other exchanges, so I'd expect random spurts of high intensity trading there at any time, although it would be less likely at 3:00 UTC.

In summary:

MTGOX minimum at 10:00 UTC, maximum at 20:00 UTC
CampBX minimum at 11:00 UTC, maximum at 21:00 UTC
Bitstamp minimum at 5:00 UTC, maximum at 10:00 to 20:00 UTC
BTCe  minimum at 2:00 UTC, maximum at 16:00 UTC

Moscow is UTC +4:00, California is UTC - 8:00, and New York is UTC - 5:00, so if you assume that most MTGOX traders do not live in Japan, BTC traders around the world sleep (local time) at:

MTGOX: 2:00 CA / 5:00 NY
CampBX: 3:00 CA / 6:00 NY
Bitstamp:  5:00 UTC
BTCe:  6:00 Moscow (what's the acceptable abbreviation for "Moscow"?  Edit: A reader tells me 'MSK' is acceptable, so I'm going with that)

and are most active at:

MTGOX: 12:00 CA / 16:00 NY
CampBX: 13:00 CA / 17:00 NY
Bitstamp:  10:00 to 20:00 UTC
BTCe:  20:00 MSK

Summary of the summary: BTC traders used to keep odd hours. Since the start of the year though, they just don't bother sleeping anymore.

3. Do traders take the weekend off?
The bitcointalk forums are full of posts claiming that there are fewer trades on the weekend, but I haven't seen any significant evidence of this - just lower volume. This does not necessarily equate to fewer trades (although it probably does).

So as a bit of a bonus, I'm reprising the analysis from the last section, applied to weekdays instead of hours of the day.

The 2D density plot really doesn't show much variation in number of trades per weekday - except for Saturday and Sunday  - note that the upper most two bands are darker than the rest, indicating fewer trades on these days. However with such variance, the significance of the difference is not clear.

In the mean / median / standard deviation chart, however,  the difference is quite clear - for every exchange in the group:

1. There are fewer trades on Saturday and Sunday compared to weekdays;
2. The mean - median difference is always greater on Sunday than on Saturday, as is the standard deviation. So while Saturday is nearly always a quiet day, Sunday is only quiet usually - and every now and then trading goes crazy. 

4. Is this a picture of traders surfing, or what?
The final chart is also the piece de resistance. It's a chart of the mean and median percentage of weekly trades occurring on a particular hour of the week. I've smoothed it a little to make the cycles a bit more clear, and the shaded areas show the standard error.

Since there are 168 hours in the week, the expected percentage of trades to occur in any given hour would be about 0.6% - use that figure as a guide for interpreting the chart.

Beautiful, huh? It neatly encapsulates all the information in the previous sections:

• There are (or have been) peaks and troughs in trading at certain times of the day, which seem to correlate with (very late) sleeping times for the geographical area in which the exchange is located, rather than being determined by the location of the currency being traded for BTC.
• MTGOX is the exception to this rule, probably due to its position as the first and best known BTC USD exchange.
• There are also cycles of greater and fewer trades during the week - weekends see fewer trades on average, although Sundays can see much more trading than the previous day.

I realise there are no hypotheses being tested in this post, but the visualisations are quite clear and not at all an excuse to post chartporn.

Thanks to bitcoincharts.com for supplying the data on which this post was based:

1kqHKEYYC8CQPxyV53nCju4Lk2ufpQqA2



organofcorti.blogspot.com is a reader supported blog:

12QxPHEuxDrs7mCyGSx1iVSozTwtquDB3r


<elevenpoint five>