How should Amazon handle bad sellers?

In the current issue of the Harvard Business Review, there is an article discussing some of the results from this working paper by Feng Zhu and Qihong Liu. They look at products on Amazon which are available only through a third-party seller. The authors try to understand how Amazon decides which new products to add. Specifically, for objects previously available from a third-party, you have a bunch of data which may be able to help distinguish one item from another.

In the past there have been two possible motivations proposed for platform owners: They might identify items which the third-party market is not handling well, or they might instead try to improve their profits as much as possible. The article makes an empirical case for the latter objective, which makes sense. The point of this post is to explore if entering markets as a customer service move also makes sense.

First we have to wonder, does the platform owner entering a market necessarily improve the customer experience? There are some cases I can think of where it would, but also cases where it would not. If the product is good, presumably the right distribution would lead to satisfied customers. In Amazon's case, they need to identify if the manufacturer (e.g., packaging, understaffed distribution office) or the third-party seller (only ships on Tuesdays) is to blame for the poor distribution. Working with the manufacturer directly could improve the customer service experience for those cases.

Bad products are a somewhat different situation. Should the platform owner allow the bad product to stay on their platform? It depends why the product is bad. Just yesterday Apple banned hundreds of apps which violated users privacy. Similarly, you can find lots of stories online of Amazon sellers being banned without recourse for alleged violations. But what about the Fizz Saver which is just a terrible product? Does it help or hurt Amazon to let this product continue to be sold on their platform?

One final question: How can we use algorithms to distinguish from a bad product and merely bad distribution? For example, in my hunt for an exceptionally bad product I found many things which got either good or bad reviews as a joke. Does the existence of these kind of products on the platform have any effect on Amazon?

Book Review: Poorly Made in China

I recently read Poorly Made in China by Paul Midler on the advice of a friend. It was selected as the best book of 2009 by The Economist, and with good reason. Not only is it informative, but also a highly entertaining read.

When this book came out, I was living in Hong Kong for a semester abroad. When I first got there, I was very reluctant to buy from the street markets. I was both learning how to negotiate prices for the first time, and having to be much more careful about quality. As the semester progressed I got better at negotiating and somewhat better at checking quality. But the key trick I learned was to buy things that were easy to assess the quality of (tea, toys with non-movable parts, etc.).

The main lesson I learned, which is reiterated in the book, is that appearance is put first in China. One of my favorite examples of this was a watch which I bought without examining it closely enough. I thought I was getting a 3-function watch (which showed the time, date, and if it was night or day) which I expected to break within a few months. Instead, I found after I bought it that only the time actually changed and it always said it was half day-half night and the 17th, and also broke within a few months. Not all of my purchases went this way (I got an excellent price on a hiking backpack which I still use from time to time), but over time I got more careful about which sorts of things I bought off the street.

Paul Midler talks about this culture as it relates to working with Chinese manufacturers making goods to be sold in the rest of the world. My strategy worked for two reasons: I was only committing to buy the object one time, and I bought things that were more robust to the culture. Importers can be strategic about which products they contract to manufacturers in China, but quality fade over time is a serious issue for any product (as is exemplified many times in the book). Hopefully the fact that more people are aware of the culture has helped companies make better sourcing decisions.

Models are not real.

My department has had two speakers this semester who said some variant of "models are not real." From my physics training, that statement is obvious. There are tons of jokes about physicists stating their ridiculous assumptions before they start to solve a problem. But the reason the assumptions seem outlandish is that they explicitly state the difference between their model and reality. As the complexity of the system grows beyond a hydrogen atom, even articulating those differences quickly becomes impossible. And that is ok. The point of a model is not to be real, it is to be useful.

So how is all this a problem? In engineering we really cannot state all the assumptions we make in our model. Instead we focus on stating assumptions we make that may be different than those other people in our field make. But if an entire field has been making the same invalid assumption, there is no obvious opportunity to identify it.

This is where empirical work and diversity can help. If we find that some observed phenomenon is not captured by our model, we can change the model. If someone who is new to the field asks questions about something we didn't even realize we were assuming, we can change the model. As long as we try to focus on what we want our model to do, we can try to include the important features of reality to get meaningful results.

Ranking for simulation

This week I visited Georgia Tech and met with several faculty there. One of those people was Seong-Hee Kim who does work on a variety of topics including ranking and selection strategies for simulation. The idea behind selection via simulation is that we have a (manageable) set of alternatives and want to choose the best option. However, identifying some options which are not likely to be the best is much easier than actually identifying the best. Therefore, the goal is to get enough information about each alternative (in her example that was a few simulated outcomes) in order to identify some bad solutions. You then keep repeating the process (more simulated outcomes, eliminate more bad solutions) until you are left with only one option.

This method seems to be similar to satisficing (mentioned here) as a decision making strategy. With satisficing, we pick any alternative which matches our criteria. Said another way, we eliminate unacceptable alternatives until anything left is acceptable. There are certainly differences in the goals between the two problems... But if you paused either algorithm in the middle, you would probably have a similar looking set of alternatives. 

Bidding strategically on eBay

A while ago I discovered that eBay makes the last three months of sold as well as completed listings available (just click on the little "Advanced" tab on the right of the search button). Learning about this feature hugely changed how I shopped on eBay. While you can do a search of the current listings to try to get a sense of your options, you can get a much more accurate picture of what the market is doing by looking at what people have historically paid. Here are a few ways I use the historical data to decide how to bid:

• If I see a "buy it now" item that is around the lowest price the auctions typically end at, I can simply use that option.
• I can figure out what the market rate is for the item and set my proxy bidding maximum at that amount. If I get outbid, I just try the same price on a similar auction.
• I can figure out how often the item is available via eBay and if I should try to find a good deal there versus looking elsewhere.

If you aren't familiar with this feature, I highly recommend giving it a try. Also, if you have any other ideas for how to use this tool, feel free to share them in the comments.

Logistical planning

As I learned when searching Airbnb for somewhere to stay during the INFORMS Annual Meeting, the pope is visiting Philly this year. He will be there September 26-27 for the "World Meeting of Families Congress," and there has been an incredible amount of logistical planning to both ensure security and allow for the large numbers of people they expect to attend the pope's events.

This link covers a lot of the changes that have been planned which include: limited or no driving and parking (some of which started today!), reduced subway / trolley / bus routes, and extra security to get close to downtown at all. They also have planned for a lot of extra services including porta-poties, free wifi, and even 500 "teenage athletes" to help push around people in wheelchairs. And that doesn't even get into the "logistical nightmare" that would arise if a thunderstorm rolls in.

Everything should be long back to normal by the time I get there... But I am happy to take advantage of the extra capacity on Airbnb as a result of the visit.

Second-order bias in models

You may have heard about the controversy surrounding the SAT over the years since people of color have always done worse on the test. The test-writers have attempted more and more to avoid racial bias in the results. It is clear that the test used to have a significant racial bias. It is less clear if that this is still the case today.

This kind of question is studied in the field of "disparate impact." Salon.com just published this article talking about that field and the risks of having disparate impact especially when you trust algorithms to make the decisions. I would call the old SAT an example of first-order bias, and the remaining issues in the test mostly second-order bias.

The problem is clearly not simple to solve. However, hopefully with time the very data analysis techniques that currently lead to biased decisions can ultimately be used to avoid bias in outcomes.