11.3 MTGOX trade post Dwolla

16th June 2013

0. Introduction

A couple of weeks ago bitcointalk.org forum member ldrgn sent me a tip (thank you sir) and asked me to analyse the global bitcoin exchange volume after the seizure of MTGOX's Dwolla account. Some information about this began to circulate from 14th May 2013, but was not confirmed by MTGOX until 17th May 2013. One might reasonably expect this to have an effect on the MTGOX volume of trade, but there are a number of interesting questions yet to be answered:

• Will most traders use other methods to get fiat currency to MTGOX? If this is the case we should see a recovery in trade volume.
• Will most traders move to other exchanges? In this case other exchanges will have an increase in trade, and the MTGOX trade volume will not quickly recover.
• Will this cause a decrease in BTC trades, due to a loss of confidence in bitcoin? If so, all currency pair trade volume will decrease.

Although financial analysis is not something with which I have a great deal of experience  I agreed to have a look at the data when I had time and see if I could come to some conclusions. This post presents some of the different ways analysed the available data.

The first part of the post presents the methods I used to analyse the data and some general conclusions about the last six months of fiat currency / bitcoin trading, and the second part of the post answers ldrgn's question more directly. I will be analysing various trading statistics with the exception of changes in price.

1. An overview of the data
When I'm initially presented with a data analysis problem I don't know how to approach, I start with a general overview of all the data I am able to obtain. I downloaded all the available exchange data from bitcoincharts.com, with the exception of the MTGOX data which I pulled directly from MTGOX using the MTGOX API.  After converting raw data to weekly and daily totals I charted volumes for all the exchanges in the data. The dates are for the last day of the week (Friday to Thursday), or the last hour of the day, UTC. The dotted line shows the May 17th date, although BTC trade will have started to reduce a few days before that on 14th May. I have not provided a legend since this was a simple overview, however it is obvious that the red line are MTGOX USD BTC trades.

This chart was sufficient to make interesting points.

• Most of the larger exchanges follow the MTGOX USD BTC weekly and daily volume changes.
• The most useful data the daily trends showed was that for MTGOX USD BTC trades volume actually increased from the 14th to the 16th of May, only decreasing from the 17th May onward. The loss of volume due to the inability of traders to top up their MTGOX accounts using Dwolla did not have any significant effect for several days. Apart from this, the daily averages are too volatile to show trends. 
• A large drop in volume had started more than a week before the Dwolla forfeiture, after the BTC price stabilised. This makes detecting a drop in volume after the MTGOX Dwolla account forfeiture much harder - a continuing drop in volume could be part of the general stabilisation of the USD BTC price.

and a few ideas about how to proceed:

• In order to compare exchange currency pair trade volumes, the data needs to be rescaled.
• Exchange currency pairs with small volumes have too much variance to be useful and need to be removed from the data.
• The size of the dataset needs to be restricted to make an initial analysis more tractable. Data that are probably dependant on similar variables will be most useful.

2. Rescaling and reducing dataset

For the initial phase of analysis I decided to reduce the dataset to USD BTC trades at larger exchanges. MTGOX USD BTC trades would be most affected by the MTGOX Dwolla account forfeiture, and I expected this to be a point of difference to other exchanges.

Rescaling data can be done many different ways. I chose to rescale the weekly trade volume to a percentage of the average trade volume since the January first this year. The method is simple - for each exchange pair take the average of the weekly volumes since the start of the year, and divide each week's volume by this amount. I expressed the result as a percentage to make it a little simpler to follow, and plotted on a log scale.

The chart below shows the rescaled weekly volumes for MTGOX, Bitstamp, BTC-E and CampBX. The dotted white line is for May 17th.

Although the data can be compared generally, it can't be compared directly with any degree of confidence. For example the post 17th May MTGOX trade volumes are much lower than those for the other exchanges, except for BTC-E. Does this mean BTC-E was negatively affected by the MTGOX problems and that the CampBX and Bitstamp volume increases reflected a general increase in BTC trades? Or conversely that the increase in CampBX and Bitstamp volumes were caused by the MTGOX problems and that BTC-E trade volumes reflect the wider trends of BTC trading. Or is there little significant difference between them?

Before I figured out how to approach that question, I thought having a look at the weekly number of trades and weekly average BTC per trade might also be rewarding.

• The number of USD BTC trades per week is trending upward at all exchanges in the group, most clearly for Bitstamp and CampBX.
• As might be expected with an increasing BTC price,  the average BTC per trade has been decreasing. However the average BTC per trade has been increasing at CampBX since the start of May. I'm not sure why - perhaps only smaller traders had used it until then, or perhaps they changed some trading rule.

I mentioned a problem with the above approach - with a small dataset some general comparisons can be made and trends identified, but it's impossible to determine the significance of any observable differences. This could be done simply if assumptions (for example normality) could be made about the nature of the variables, but this is not possible. Instead, data for many high volume exchange currency pairs can be combined and percentiles for each date calculated.

2. Rescaled trade variable percentiles

First, all weekly trade volumes / number of trades / average trade per btc for each exchange currency pair is rescaled to the average since January first. I selected 33 currency pairs with a median weekly volume greater than 350 btc (in alphabetical order):
bit2cILS
bitboxUSD
bitcurexEUR
bitcurexPLN
bitnzNZD
bitstampUSD
btcdeEUR
btceEUR
btceRUR
btceUSD
btcnCNY  
cbxUSD
fybsgSGD
icbitUSD
intrsngEUR
intrsngGBP
intrsngPLN
kptnSEK
mrcdBRL
mtgoxAUD
mtgoxCAD
mtgoxEUR
mtgoxGBP
mtgoxJPY
mtgoxPLN
mtgoxUSD
rippleUSD
rippleXRP
rockEUR
vcxEUR
vcxUSD
virtexCAD
virwoxSLL

Next, percentiles at each date have to be calculated. It can become easy to get confused at this point since both the percentiles and the rescaled percentages are measured in the same units.

The percentiles for each date indicate the number of exchange currency pairs at a particular value.The 10% (tenth percentile) of trade volume, for example, indicates a rescaled percentage below which are ten percent of exchange currency pair rescaled trade volumes for a given date.

The MTGOX rescaled trade volume percentages are provided as a comparison.

The path of the MTGOX rescaled trade volumes moves between the 30th and 70th percentile bands until the price drop in April when it moves to the 90th percentile band. This means that until mid April, the MTGOX rescaled trade volume was greater than that of 30% to 70% of exchange currency pairs, and then is more than 90% of exchange currency pairs.

The rescaling is done with respect to the yearly average, so this could mean either that trade volumes in the first quarter divided by the average until now were similar to most other exchange currency pairs, or that the peak divided by the average until now was higher than most other exchange currency pairs, or a bit of both.

The MTGOX path after May 17th drops through several percentile bands and then levels out - meaning a much larger drop in the rescaled trade volume than other exchange currency pairs.

I also did the same for the weekly number of trades and the weekly average BTC per trade.

• The number of trades per week has been increasing generally since the start of the year.
• The average BTC per trade has been steadily decreasing since the start of the year, although it has been increasing over the past month.
• The reduction rescaled number of trades per week at MTGOX after 17th May is in line with that for most other exchange currency pairs - although volume was affected, the number of trades was not.
• As expected from the above point, the average rescaled BTC per trade at MTGOX after May 17th reduced significantly from the 70th to the 40th percentile bad and has remained there.

3. Rescaled trade variable percentiles, alternate visualisation.

These section 2 charts are important since they provide a good overall indication of BTC trade statistics generally.

However following the path of any one exchange currency pair is difficult, especially when the percentile bands are close together. Comparing more than one exchange currency pair on this sort of chart is impossible. A chart that "straightens" the percentile bands (a chart with a y axis in percentiles rather than in rescaled trades per week) will be more useful for this purpose.

The chart below shows the MTGOX weekly trade volume percentiles from the weekly BTC volume percentile chart (first chart in section 2), except with percentile bands "straightened". 

This makes the changes much easier to see, and the large post 17th May dip (and subsequent recovery) more obvious. There's also plenty of space to compare to other exchange currency pairs.

The next set of charts are the same as the above, without the colour legend (since the y axis serves that purpose) and compare weekly rescaled volume, number of trades and average BTC per trade percentiles for various BTC currency pairs and various exchanges.

The USD BTC comparisons show just how misleading relying only on the section 2 charts would be.

• The MTGOX rescaled weekly volume did indeed drop significantly in the week immediately after 17th May, from about the 80th to 25th percentile. However BTC-E had also dropped by a similar amount by the following week, and then recovered more sharply than MTGOX. There seems to be a possible correlation there.
• The rescaled weekly volume for CampBX and Bitstamp did not increase percentiles after 17th May. It's not possible to say that this means they were unaffected by the MTGOX problems, but it seems suggestive.
• The drop in the number of trades per week is more obviously significant than it was in the section 1 charts, but it has subsequently recovered. The sharp increase in number of trades as MTGOX's weekly number of trades declines is interesting.
• In terms of BTC per trade, BTC-E seems to follow MTGOX's lead again.

I will leave further analyses (and those relating to non USD BTC trade statistics) to the reader.

These charts have a limitation - the rescaling method. Rescaling to the average is fine for medium term trends, but when the point of interest is a four week period other events in the preceding six months may have had an impact.

4. More rescaling

Instead of rescaling to the average since January 1st, rescaling can be done with respect to the weekly trade data of the week ending May 16th. This couldn't be done from the start: if MTGOX USD BTC trades for the week ending 16th May were abnormal in some way, then a comparison to the abnormal data is useless. However section 1 and section 3 show that MTGOX USD BTC weekly trade statistics for the week ending May 16th is not unusual, both compared to the year average and the percentile for all 33 exchange pairs, so a comparison to this date is valid.

I produced rescaled (section 2) and percentile charts (section 3) using this data:

Finally, a robust result.

• MTGOX USD BTC trade volume: in the two weeks after 17th May dropped to 50% of the volume of the week ending 16th May and after four weeks has almost reached that volume again.
• Comparison to other exchange currency pairs: the percentage of volume of the previous week was lower than for any other exchange currency pairs: MTGOX USD BTC was the 0th percentile. After four weeks, 75% of exchange currency pairs have a higher percentage of their week ending 16th May trade volumes than MTGOX USD BTC.
• MTGOX USD BTC weekly number of trades: dropped to about 75% compared to the number of trades in the week ending 16th May, but then recovered to above this level.
• Comparison to other exchange currency pairs: the rescaled number of trades did not change significantly for very long - 70% to 50% of  exchange currency pairs had higher percentages of the comparison week number of trades.
• MTGOX USD BTC weekly average BTC per trade: also dropped to about 75% compared to the number of trades in the week ending 16th May, and then recovered to almost 100%.
• Comparison to other exchange currency pairs: the rescaled BTC per trade immediately dropped to the 20th percentile (80% of exchange currency pairs had higher rescaled BTC per trade) and does not recover. 

5. Summary

I have really only scratched the surface of the sorts of analyses this method allows, but I'm not really familiar enough with the exchanges or the intricacies of BTC trading generally to make many inferences based on the analyses. For example, why does the weekly volume and weekly average BTC per trade for BTC-E USD BTC follow that for MTGOX USD BTC so closely? Coincidence, or were BTC-E indirectly affected by MTGOX's Dwolla problems?

Anyway, here is a summary of the inferences in which I have a degree of confidence, and some answers to my initial questions.

1. MTGOX USD BTC daily trade volume was not affected from the 14th May (the date Dwolla announced it would no longer be a MTGOX partner) and the 17th May. 
2. MTGOX USD BTC weekly trade volume and average BTC per trade reduced significantly after 17th May, and have not recovered. Interestingly, the weekly number of trades has not been significantly affected. Why? I don't know. Suggestions are welcome.
3. CampBX USD and Bitstamp USD both had a significant increase in both trade volume and average BTC per trade compared to the week ending 16th May. The weekly numbers of trades do not show a significant change.
4. MTGOX currency pairs that may have been affected by the Dwolla account closure (in order of confidence): USD, GBP, PLN.
5. Non MTGOX currency pairs that may have been affected by the Dwolla account closure: BTC-E USD.

To answer my initial questions directly:

• Will most traders use other methods to get fiat currency to MTGOX? If this is the case we should see a recovery in trade volume.

From point 2, this does not seem to be the case. Rescaled trade volume compared to 33 other exchange currency pairs was still quite low - lower than 80% of exchange currency pairs.

• Will most traders move to other exchanges? In this case other exchanges will have an increase in trade, and the MTGOX trade volume will not quickly recover.

From point 3, this is quite possible, Bitstamp and CampBX being the most likely recipients of the volume.

• Will this cause a decrease in BTC trades, due to a loss of confidence in bitcoin? If so, all currency pair trade volumes will decrease.

No. With the exception of BTC-e USD BTC and possibly MTGOX GBP BTC and PLN BTC, no other exchange currency pairs showed any correlation to MTGOX's trade statistics.

I'd like to hear some suggestions from readers (who managed to persevere to the end of the post) about why these somewhat mysterious changes may have occurred, and also any other conclusions that may be obtained from the data - or if you think I've made an egregious error somewhere.

Thanks to the following for use of their data:

bitcoincharts.com: 1kqHKEYYC8CQPxyV53nCju4Lk2ufpQqA2

MTGOX

organofcorti.blogspot.com is a reader supported blog: 12QxPHEuxDrs7mCyGSx1iVSozTwtquDB3r