Abstract
Background: Illicit online marketplaces for drugs on the Tor network have been studied so far from a descriptive perspective and the only data available on consumer choice behavior stems from interviews with small participant samples. This article seeks to increase external validity of previous research by providing a whole-market analysis of choice behavior on a third generation online drug marketplace.
Methods: Data was collected from the Silk Road Reloaded website using webcrawling software. Consumer choice behavior, approximated by the number of views of items, was analyzed using regression methods.
Results: Attractiveness of items on the marketplace in influenced by a variety of factors, which are: delivery time, items in stock, negative seller ratings, number of reviews, and time since the last log-in of a seller.
Conclusion: This study is the first to provide statistical evidence on consumer choice behavior on a Tor network drug marketplace. The findings confirm the qualitative results found in previous research and extend the results by systematically investigating the bits of information provided by the website. Various theories on consumer choice behavior are provided.
Introduction
There is in the current literature a raise in interest in drug trafficking on the Tor network, also referred to as the so-called dark net or hidden web (e.g. Corazza et al. 2012; Martin, 2014; Van Hout & Bingham, 2013a, 2014). The Tor network is, broadly speaking, an anonymization network within the internet, containing its own websites and guaranteeing its users full anonymity when following a couple of simple precautions. Within this network, a few websites recently made it to the front page of the news, when the FBI took down some of the major drug trafficking websites of the Tor network (e.g. The Guardian, 2014). Without doubt, “Silk Road” was the most prominent of these websites (shut down on October 3rd 2013). Shortly after the end of Silk Road, the market places for drugs re-emerged, for example in the form of “Silk Road 2” (Dolliver, 2015). In a second major coup, the FBI took down a large number of these websites as well on November 6th 2014, just to see shortly afterwards new marketplaces (oten termed “cryptomarkets” (Martin, 2014)) emerge. The aim of this paper is to analyze one of these third generation websites (“Silk Road Reloaded”) from a new perspective. Up till now, the literature has focused on either (a) describing these marketplaces (Aldridge & Décary-Hétu, 2014; Christin, 2013; Dolliver, 2015) or (b) providing case studies on users of the marketplaces (Van Hout & Bingham, 2013a,b). In the current paper, we intend to provide statistical evidence on consumer decision making in such a marketplace. More precisely, we will extract publicly available data from the marketplace and analyze which factors lead to a higher interest addressed to a specific item on the marketplace. As a matter of fact, the Silk Road cryptomarket (and all other marketplaces on the dark net), because of the strict anonymity, strongly rely on a system of trust and trustworthiness in order to operate (Van Hout & Bingham, 2013b). To be more precise, given that a consumer (i.e. a buyer) does only know that (s)he could buy a good from a seller with some random username that does not reveal anything on his/her identity, and that there is no concise system of liability if a good has a different quality than expected or if the seller just cashes the money without providing any good, (s)he needs to rely on every piece of information on the reliability of a seller (s)he can retrieve. Also, goods in online marketplaces are highly substitutable (Dewally & Ederington, 2006; Rysman, 2009): many sellers usually provide similar goods of similar quality for similar prices, and it is up to the consumer to decide where to buy. This paper will be the first to statistically analyze the cues that consumers use in anonymous drug markets in order to choose where to buy.
The remainder of this article will be structured as follows. We first will provide a short description of the Silk Road Reloaded marketplace, focusing only on unique features on this website as compared to the websites that have been studied so far. Diana Dolliver (2015) provides a comprehensive overview over Silk Roads 1 and 2, reviewing the current literature – we will instead focus on singular patterns of Silk Road Reloaded and how these can inform us about consumer choice behavior. We then will formulate our hypotheses and explain the automatized data collection techniques used. The analysis and results section provides the results on which the discussion will be based.
Silk Road Reloaded
While the current market leader for Tor network drug sales is, according to the “hidden wiki” the “Green Road”, we decided to retrieve data from “Silk Road Reloaded” (hereafter “SRR”) instead, for three reasons: first, SRR claims to be the successor of the original Silk Road Marketplaces. Second, access to the SRR website is much more stable. Third, in contrast to Green Road, SRR does not require re-entering a CAPTCHA when losing access to the website, which happens often with an unstable connection. Stable website access is an important asset when automatically collecting data, and automatized CAPTCHA resolution is still under research (e.g. Nachar et al. 2015). Therefore SRR was the site of our choice to collect data. The SRR marketplace is on many dimensions highly similar to the Silk Road 2 marketplace studied by Dolliver (2015). Notably, the number of active items on the marketplace ranged around 1200 (Dolliver lists 1834), being considerably smaller than the original Silk Road with roughly 13,000 items (Christin, 2012). The distribution of countries of origin also is headed by the United States (and the Philippines), followed by mostly EU countries (figure 1 provides a world map with the percentage of sellers of each origin). Yet, 66% of all items on SRR are drugs, whereas drugs represented only around 19% of all items on Silk Road 2. As compared to the original Silk Road, there is no stealth mode option where sellers and buyers can interact in hidden web shops within the site and buyer feedback is not compulsory.
Research question and hypotheses
Relying on the data publicly available on the website, we want to explore which are the factors that influence the attractiveness of a product. To be more precise, we want to use the number of views received by a given item on the marketplace as a proxy for attractiveness of that item and hence as a crude measure of likelihood that the item will be sold (i.e. we assume that the views correlate with the amounts sold). Although we cannot access purchasing data of the website, a broad literature makes us confident in believing that the number of views, at least in clearnet markets, correlates with the purchased amounts (e.g. Bockstedt and Goh, 2011; Kossecki, 2009; Chen et al. 2009; Morales-Arroyo and Pandey, 2009; Nalchigar and Weber, 2012; Sandvig, Mobasher abd Burke, 2008). We do not claim that there is a 1:1 correspondence between viewing behavior and purchases, yet we assume a significant and positive correlation and hence interpret more views as an indicator for a higher purchasing probability. We will correlate the information provided by the website with the number of views received by individual items. We expect that various types of information (information on items, on sellers and on the bitcoin market) will have an impact on the number of views (the information will be described in more detail in the following sections). Based on Van Hout and Bingham (2013b), who have shown that transaction time plays a central role in buyer decisions, we expect that items with a shorter delivery time will receive more views. Similarly, we expect buyers to take time since last seller login as a cue for estimating actual delivery time and we therefore hypothesize that the delay since the last login of a seller also influences buyer decisions. Based on classical economic theories on signaling (Spence, 1974), we also expect the stated amount of items in stock and the number of different items a seller offers to significantly influence the number of views, as these numbers can be interpreted as signals of seller professionalism. Seller ratings and reviews are also expected to influence the attractiveness of an item. We expect positive ratings and the number of reviews to correlate positively with the number of views, and negative ratings to correlate negatively. Finally, as the SRR has similarities to “foreign-currency” markets, we also control for currency related variables and expect the bitcoin exchange rate to correlate over time with the number of views.
Methods: Data collection
Data was collected between December 18th 2014 and February 2nd 2015 from the SRR website. Data collection methodology was similar to Christin (2013) and Dolliver (2015) (i.e. using software to repeatedly and automatically download websites in a given time window, also referred to as “crawling”). The main difference was that, instead of using the HTTrack software, we used the build-in wget tool of the whonix Linux distribution (https://www.whonix.org/ – the whonix distribution is configured by default to tunnel all internet traffic through the Tor network). We chose to rely on wget as, although configuration for wget is more complex than configuration of HTTrack, this software allows to strongly narrow down which data is to be collected from websites (e.g. excluding sub-sites and files of no interest). This way, we were able to only download two types of websites: seller pages, containing information about individual sellers, and item pages, containing individual information on the items sold on the website. We were able to condense data collection to between 30 and 90 minutes per day (depending on Tor network accessibility) and total download sizes averaging 12Mb/day, which is a more than five-fold reduction in time and data traffic as compared to previous research (Dolliver, 2015). A first full crawl of the website was performed on December 12, 2014, using the standard methodology: after performing the first crawl, we waited six days to see if the data collection had been observed. Although we cannot exclude that our activities had been noticed, the administrators of the website did not take any measures to interfere with our crawling activity (such as contacting our user account or blocking it). Data of the first crawl was then discarded and data collection started on Dec. 18th, 2014, for 25 days.
After data collection, an automatic tool, written in Java, extracted the relevant data from all webpages. Extracted data for seller sites include user name, positive and negative ratings, last login of the seller, number of different items a seller sells and number of reviews the seller received. Item pages contained information about the title of the item, user name of the seller, price in € and in BTC, country of origin of shipment and possible destinations for shipment, delivery time, sub- and super-category of the item (e.g. the item entitled “0.5 grams Cocaine | Purity 70%” would belong to the sub-category cocaine which belongs to the super-category stimulants), items in stock, time since the item was added to the marketplace and our variable of interest: number of views the item received. After extraction of this data, item- and seller-data was matched and merged, resulting in 14,454 data-points. Item categorization was cross-checked to ensure that all items were contained in the correct category. The website listed nine super-categories, viz. opioids, prescription drugs, benzodiazepines, stimulants, ecstasy, psychedelics, cannabis, dissociatives and “others” (containing all non-drug items, such as counterfeit, e-books, etc.), and 38 sub-categories.
Results
Data analysys is still in progress – I therefore recommend that you check back in a while (yet results are very promising). Thill then, you can find a descriptive overview of the distribution of drug merchands here and refer to the preliminary results below.
Item related factors
We have shown that the longer it takes to deliver a good, the less views it receives. This result is in line with Van Hout and Bingham (2013b) who provide interview evidence that speed of transaction is a major driving force in item selection. Given that most humans recur to temporal discounting when making decisions (Frederick, Loewenstein & O’Donoghue, 2002), this result is not surprising. Temporal or delay discounting is a well-researched effect which assumes that the longer it takes for an individual to receive a good providing utility, the lower the current utility of this good. It is important to note that temporal discounting is attenuated in drug users (Bickel, Miller, Yi, Kowal, Lindquist & Pitcock, 2007; Reynolds, 2006), which might exacerbate the effect in our data sample, where most of the items are drugs. We also have shown that the amount of items in stock positively influences the number of views. Although the coefficient might seem negligibly small, some sellers state that they have a stock of multiple million items. It probably is not reasonable to assume that a seller indeed has a stock of e.g. 999,999,000 pills of clonazepam; however, sellers might send a signal of reliability or professionalism when stating that they dispose of a large amount of a given good.
Seller related factors
It might at first sight seem surprising to observe that positive seller ratings taken alone do not influence the number of views. One possible explanation might be that there is generally a strong positivity bias in virtual markets (Jøsang, Ismail & Boyd, 2007; Zacharia, Moukas & Maes, 2000) leading to a skewed distribution that strongly favors positive over negative ratings. Yet, when looking at the interaction of positive and negative ratings, a small but positive effect can be shown. Thus it appears that a large number of positive ratings can compensate for a small number of negative ratings.
Negative ratings on the other hand have a strong, negative impact on the number of views. This result could be expected, given that especially a market for clandestine goods strongly relies on trust and reputation (Van Hout & Bingham, 2013a,b). A few negative ratings therefore can be interpreted as signal of relatively low trustworthiness or bad reputation of a seller.
Surprisingly, the number of reviews a seller received also correlates negatively with the number of views. One speculative explanation for this effect could be that we only controlled for the number of reviews, but not for their valence. It could be that some critical fraction of reviews have a negative content, leading to the negative aggregate impact of the number of reviews on the number of views.
A further signal of reliability of a seller appears to be the time that has elapsed since his last log-in. In a market where only few signals on the quality of the trade partner are available, it seems that the time since the last log-in might be used as an additional bit of information to measure reliability. Alternatively, returning to the idea of temporal discounting, a seller that logs in frequently also can be expected to deal more frequently with the orders of his clients.
Market related factors
We controlled for the impact of the BTC market on views by measuring the impact of various variables linked to that market. It appears that Silk Road users are not economically sophisticated buyers in the sense that they do not condition their viewing behavior (which might also be seen as a proxy for buying behavior) on BTC exchange rates. In this sense, the clients on Silk Road behave differently from stock market traders (or of other markets where monetary exchanges has to take place before acquiring goods) who condition their buying behavior on exchange rates (e.g. Ma & Kao, 1990). The explanation for this might lie again in temporal discounting (though this is only one explanation among others): drug users do not want to (or cannot) wait.
Conclusion
This study is the first to try to quantify the cues used by clients on a hidden web drug marketplace to select the items that are likely to be bought. Although we are not able to look into the minds of buyers, we largely confirm the results that have been found in previous works on consumer preferences in Tor network drug marketplaces (Van Hout & Bingham, 2013b). We thereby strengthen the evidence that clandestine online marketplaces are highly trust-based and provide some starting points for future research. Notably, the explanatory theories on buyer behavior discussed, no matter how well founded in the “legal world”, remain to be backed in a Tor market environment with different methods than those applied in this article.
Acknowledgments
The author would like to thank Lena Kittel for her hard work throughout the project. The author also would like to thank Patrick Schleizer and the whonix community for providing the whonix Linux distribution, which made it safe to browse the Tor network. No additional funding was received for this project.