作者：DARIUS KAZEMI

有些新闻网站发表了一篇由Louis Bedigian编写的文章，他引用了分析师Arvind Bhatia的观点，称“每获得一名新付费用户都会让Zynga损失150美元”。我看了这篇文章，认为只是宣传的噱头而已，便猜想不会有人转载这篇文章。我的想法是错误的，现在这篇文章到处都是。

后来Andrew VandenBossche提醒我看一篇Dylan Collins编写的文章，他引用了行业首席执行官Torsten Reil的说法，后者的回应是：“你的方法论是错误的！Zynga可以从每个付费用户上赚得30美元！”所以，此刻我的想法是：这到底是什么情况（游戏邦注：本文作者在全职制作HTML5内容前，在游戏工作室担任了6年的数据分析师，分析对象包括MMO和Facebook游戏）。

attrition-rates(from genengnews.com)

表面分析

Collins和Reil认为原分析方法过于简单，这是正确的。原分析结果是基于完全未考虑用户流失率的模式而得出的。他们声称Zynga完全通过营销资金来获得逾40万的所有付费用户，在这个前提下，他们的结论是正确的。

不幸的是，Collins和Reil考虑到20%的流失率，因此估算出Zynga在每个付费用户上花费的金钱是120美元，而每个付费用户能够贡献150美元，由此得出每个付费用户30美元的净盈利。我之所以说这很不幸，是因为如果流失率是10%，那么根据Reil的指标，他们就会在每个付费用户上损失21美元。如果数值是30%，那么他们可以从每个付费用户上赚到57美元。盈利的多寡取决于流失率，而这项数据的确切数值我们并不知晓！有些游戏是10%，有些游戏甚至高达90%。20%似乎是综合考虑成功和非成功游戏的平均值，但是坦诚地说，我们根本不知道具体的数值，因为我们不在Zynga工作。然而，奇怪的是，Zynga也不知道他们自己的流失率。

模式和黑盒

所有这样的数值都是构建于分析师设立的模型之上，而模式是构建于猜想上。举个简单的例子，当我们谈论用户流失的时候，我们提到的是什么现象？通常情况下，我们指的是“有人不再是游戏玩家的时候”。但是，在社交游戏的背景下，你要如何来确定这种时刻呢？Facebook用户往往不会卸载应用，他们只是停止使用应用而已。所以，你必须选择主观分界点。玩家不活跃时间达到1周应当被归入“流失”类别？2周？还是1个月？记住，这个数值是主观的，所以你可以根据自己的偏好调整数值（游戏邦注：当然，应当选择合理的数值，比如不能以100年作为分界点），直到你获得满足自己标准的流失率。至于这些标准是“呈现游戏实际表现情况”还是“让股东感到满意”，这就另当别论了！

但是无论如何，这个损耗率会影响到所有其他计算的数值。现在，理想情况下，一旦你选择了数值，就会保持内在的一致性。但是秘密在于，即便你完美地维持内在的一致性，比如总是使用“2周不活跃”作为流失分界点，依然存在其他可以修改的间接因素。而且，在某些情况下你必须修改这些数值！否则，你可能会发现自己使用的模式无法反映游戏的现实状况。

换句话说，游戏工作室内部分析师的工作就是组建黑盒，也就是“流失”的概念，呈现给CEO、CFO、股东、外部分析师和专家，这个概念似乎相当直观：离开游戏的人群。黑盒的工作是报告介于0%和100%间的数值，如果数值接近100%的话，或许会让你感到恐慌。

但是，工作室内部分析师组建这个概念可以借助多种信息。他们或许会询问游戏设计师，看看设计师认为玩家离开游戏多长时间是“自然”的情况。如果游戏设计的目标是玩家在工作日体验游戏，那么即便有人在周末或圣诞休假期间不玩游戏，你也无需感到担心。他们可以查看游戏的历史数据，发现不活跃天数达到12天的玩家中有80%不再回到游戏中，不活跃天数达到15天的玩家中有90%不再回到游戏中。所以，或许我们可以选择将15天作为临界点。但是，我们查看的是当前游戏的历史数据，游戏现在的状况与之前并不相同，所以这不是完美的方法。所以，或许我们可以参考过去30天的数据，那时的游戏和现在最为相似。但是如果这样，我们对“不再回到游戏中”的定义实际上就等同于“在过去30天时间里不活跃天数达到15天且未回到游戏中的玩家”。当然，因为我们以30天为期限，所以将分界点设为15天是合理的。这样，我们的历史数据就更能够代表游戏的当前状态。

猜想的无尽分解

总之，游戏是非常复杂的系统，围绕此媒体的数值都是以黑盒式的猜想为基础构建而成的。这些黑盒总是可以被分解成诸多成分，而这些成分可以被再次分解成子成分，如此不断重复下去。尽管这看起来有点古怪，但事实情况就是如此。在某些层次上，你需要暂停这种无尽的挖掘，得出有关游戏运转情况的猜想，以此作为模式的基础。从原则上来说，这样做并没有错：这种情况经常在科学研究中发生，而且能够成功得出某些描述世界的优秀模式。但是，科学研究和社交游戏指标分析间有很大的差异。科学家无需在他们的大学或企业的空间中工作。科学家处理的不是“专有数据”，他们无需担心因为同其他科学家分享结果甚至方法论而被解雇。游戏工作室内部数据分析师是在一定的空间中工作，如果同外部分析师分享结果就有可能被解雇。这意味着，很可能我们公布的只是最好的猜测结果。而且这意味着，不同公司公布的数值间根本无法进行对比。尽管“每用户平均盈利”听起来很直观，但可能是基于完全不同的基础性猜想而得出的结论，这种差异取决于公司和游戏的不同。

这是我放弃游戏数据分析工作的主要原因之一。我不喜欢进行那些从某种程度上来说完全无用的猜想。最后，对于游戏公司能够赚多少钱的问题，唯一的关键数值是：公司每个月的收入是多少钱？公司每个月的支出是多少钱？其他所有的数值都只能视为猜想而已。

游戏邦注：本文发稿于2012年2月6日，所涉时间、事件和数据均以此为准。（本文为游戏邦/gamerboom.com编译，拒绝任何不保留版权的转载，如需转载请联系：游戏邦）

What is Zynga making per paying user? Nobody, not even Zynga, will ever know.

DARIUS KAZEMI

There was a post on some news site last week with an article by Louis Bedigian quoting an analyst (Arvind Bhatia) claiming that “Zynga loses $150 on every new paying customer.” I read it, thought to myself, “That’s absurd linkbait,” and then assumed that nobody would take the bait. I was wrong: it got picked up everywhere. Sigh.

This morning, Andrew VandenBossche alerted me to an article by Dylan Collins, quoting industry CEO Torsten Reil, that responds, no, you idiots! Your methodology is wrong! “Zynga is probably MAKING $30 on every paying user!” So… here’s what I think: nobody knows what the fuck is going on. (For those of you wondering why I’m writing about this, before I did HTML5 stuff full time, I spent 6 years as a data analyst for game studios, both MMO and Facebook games.)

Surface analysis

Collins/Reil are absolutely right to call the original analysis oversimplified. It was based on a model that completely failed to account for attrition — they’re correct when they state that Zynga certainly gained far more than 400k paying users for their marketing money.

Unfortunately, Collins/Reil pick a number out of thin air (20% attrition rate) which results in a rough estimate where Zynga spends $120 per paying user, and makes $150 per paying user, resulting in a net profit of $30 per paying user. I say unfortunately because if the number is 10%, then by Reil’s metric they’re losing $21 on every paying user. If it’s 30% they’re earning $57 per acquired paying user. It all hinges on their attrition rate, which we don’t know! Some games see 10%. Some games see 90%. 20% seems like a roughly correct ballpark for a mix of successful and unsuccessful games, but honestly we have no idea what it is because we’re on the outside looking in. But the truly weird thing to consider is: Zynga doesn’t know what their attrition number is either.

Models and black boxes

All numbers like this are built on models that analysts put together, and models are built on assumptions. Simple example: when we talk about attrition, what phenomenon do we refer to? Typically we mean “the moment when someone is no longer a player of the game.” Yet in the context of a social game, how do you define that? Facebook users don’t typically uninstall an app — they usually just stop using it. So you have to pick an arbitrary cutoff point. Does someone fall into an “attrition” bucket after 1 week of inactivity? 2 weeks? A month? Remember, this number is arbitrary, so you can adjust that number all you like (within reason, you’re not going to pick 100 years) until you come up with an attrition percentage that meets your criteria. Whether those criteria are “seems more realistic” or “would appease our shareholders” is another question!

But regardless, this attrition percentage then affects all of your other calculations. Now, ideally you want to remain internally consistent once you pick this number, but a dirty secret is that even if you maintain perfect internal consistency in always using “2 weeks of inactivity” as your cutoff for attrition, there will always be dozens of other fiddly and less directly consequential definitions that you can tweak. And the thing is, on some level you have to tweak these numbers! Otherwise you might find yourself stuck with a model that doesn’t reflect what looks like the reality of your game.

To put it another way: the internal game studio analyst’s job is to assemble a black box known as the concept of “attrition” — to the CEOs and CFOs and shareholders and external analysts and pundits at home, this concept seems pretty straightforward: it’s the people who leave your game. End of story. The black box behaves and does its job, reporting a number between 0% and 100%, and presumably you panic if the number is closer to 100%.

But the internal studio analyst needs to assemble this concept from a variety of sources. They might ask the game designers what they see as a “normal” or “natural” amount of time away from the game — if a game is designed to be played during the work week, then you shouldn’t sweat it when someone isn’t playing over the weekends or on Christmas. They might look at historical data for the game and notice that 80% of players who are inactive for 12 days never come back. And 90% of players inactive for 15 days never come back. So maybe we pick 90% and say 15 days is our cutoff. But of course we’re looking at historical data for the current game, which is different today than it was back then, so it’s not a perfect analogy! So maybe we want to rely on data from the last 30 days, when the game was most similar — but now our definition of “never come back” really means “people who were inactive for 15 of the last 30 days and haven’t been back.” But of course, those people “haven’t been back” for a maximum of 15 days since we’re looking at a 30 day window. So now that our historical data is more representative of the current state of the game, our very definition of “never” comes into question!

An infinite regress of assumptions

In summary: games are very complex systems, and the numbers that get thrown around in the media are built on black-box-style assumptions. These black boxes can always be broken down into components, and those components into subcomponents, forever and ever into an infinite regress. If this seems mind-bogglingly weird, well: it is. On some level you need to stop digging into the infinite and come up with assumptions about the way the game works that become the foundation for your models. There’s nothing wrong about that in principle: science does this all the time, and manages to come up with some great models to describe the world. But there’s a huge difference between science and analyzing the metrics for social games. Scientists do not work in a vacuum within their universities or corporations. Scientists do not work with “proprietary data” and they do not run the risk of getting fired for sharing their results and even their methodologies with other scientists. An internal game studio data analyst does in fact work in a vacuum, and will get fired for sharing with outside analysts. This means that the chances that our assumptions are off-base are pretty good. And it means that the numbers that different companies throw around can’t even be compared. “Average revenue per user,” which sounds straightforward, can be based on entirely different foundational assumptions at different companies and on different games.

This whole mess is one of the main reasons I stopped being a data analyst for games. I did not feel comfortable coming up with assumptions that weren’t, on some level, complete bullshit. Now, the level on which these assumptions operated was often very low-level, fiddly stuff. But it was an art, not a science. Which, again, nothing wrong with that — except that the black boxes that I generated were being treated as science rather than as art.

In the end, for the purposes of arguments about how much money a company is making, the only numbers that matter are: how much money is coming into the company each month? How much money is leaving the company each month? Everything else should be viewed with utmost suspicion. (Source: Tiny Subversions)