Science Killed My Marketing Job

This is a marketing analytics scenario.
*****
Jeremy traded US stocks for a few years, but is now thinking about broadening his trading to oil, gold and international stocks.

On wsj.com, he reads about the virtues of diversification through the use of international stocks. Next to the article, a BankCorp banner shows a picture of a cool, confident investor who trades stocks in Deutsche Bank and China Telecom from his cell phone. Jeremy does not click on the banner, but a cookie is placed on Jeremy's computer and he makes a subconscious association between BankCorp and trading.

A few days later, Jeremy hears that Bet24, the large bookmaker traded on the Austrian stock exchange, is undervalued and decides to act on the information. Now he needs a financial services provider right away. Jeremy does a Google search using the keywords “stock trading” and “austrian stocks”. Next to the organic search hits, the Google Adwords ads shows a BankCorp ad as the third advertisement, specifically mentioning the possibility of trading Austrian stocks. Jeremy clicks on BankCorp's ad because he has seen the name before.

On BankCorp's landing page, their trading platform's international stock trading capability is emphasized at the center of the page. The website registers what AdWords ad led to the website, along with Jeremy's search keywords. Jeremy is from New York City, a fact that is revealed and registered through his IP address. Finally, it is noted that Jeremy was once exposed to a wsj.com international stock ad.

Jeremy is curious about the price of real-time price feeds, the range of available stock exchanges and types of tradable commodities, but it also occurs to him to search for what kind of guarantee he has that his money will be paid back in the case of default. More than 4 minutes have passed.

Jeremy decides to sign up for a trial trading account, answers a question about trading experience and investable assets, hands out his name and email address. While playing around on the trading platform, he carries out a few simulated trades on Austrian stocks.

Inside BankCorp's systems, an algorithm is triggered by the arrival of a new sales lead. The algorithm can access records of past lead profiles and correlate them to the rate of lead to client conversion and to the average revenue generated by client-lead profile pairs.

Based on Jeremy's search keywords, his New York City IP address, the time spent on the bank's website, the pages he visited, the survey responses and the type of simulated trades carried out on the trading platform – from this data – it is inferred that he is likely to become a valuable customer. The algorithm decides that an experienced sales representative with special knowledge about stock trading would do well in contacting him.

Three months later it is determined for sure that Jeremy indeed did become a valuable customer. He conducted a high number of trades and generated a lot of revenue for the bank over this period of time.

A second algorithm is now triggered in BankCorp. The algorithm is interested in the relation between the wsj.com ad, the Google AdWords ad on Austrian stock trading and the value of clients such as Jeremy. The marketing database shows that he and other valuable clients were exposed to a combination of the wsj.com and Austrian stock trading ad. The algorithm set a series of decision in motion. Firstly, it decides to increase the bids on Google keywords containing the terms “trading” and “Austrian stocks”. Secondly, it decides to increase relative spending on the wsj.com marketing channel, while spending less on smartmoney.com. Thirdly, it decides to increase relative spending on the cool-confident investor rich banner ad, while spending on other banner ads are decreased slightly.

BankCorp's CEO is pleased to know, that the days of knowing what half of the marketing budget was wasted, are finally over.



Credits

Picture by Tony

Sony Reader and Mac

The Sony Reader is not designed to run on the mac, but there is a simple workaround.

###[October 30, 2009] Update: The Sony Library has finally been released for Mac. So the text below might not be relevant anymore###

It is possible to download DRM-protected EPUB and Adobe Digital Editions formatted books to a mac and then transfer it to Sony Reader via usb.

Just once, one must go through the following process.

The two pieces of software, Sony Library and Adobe Digital Editions, must be installed on a Windows PC computer. Connect the Reader to the PC. From within Adobe Digital Editions, authorize the reader.

Now install Adobe Digital Editions on the Mac. Buy an EPUB or Digital Editions book from, say, www.booksonboard.com. The book can now be read on the mac through the Adobe Digital Editions software, while the file itself is found in Documents/Digital Editions/.

Connect the Sony Reader to the mac. Drag and drop the .epub or .pdf file from the mac folder to the Sony Reader. Done.

Do I think that Amazon, Sony and the book publishers are doing a good job of making it as painless as possible to buy and read ebooks? No. Please give us one open standard for ebooks, get rid of digital rights management and make all books available to everybody all the time through a new revenue model. 60 years from now, I do not want to discover that my precious ebooks are lost, since the rights management server has perished and no current technology can read the proprietary format.

Business Management Through Customer Lifetime Value

A company's customer base can be thought of as a collection of stochastic future cash flows, not unlike a portfolio of securities. Managers then evaluate the effects of business decisions against expected changes in the present value of customers' cash flows. This is customer lifetime value analysis.

The figure above shows the event histories of a random sample of newspaper customers. Yellow marks a period of rebate subscription, while green marks a full price subscription. Red denotes a period of no subscriptions - the customer has churned. Read more in "Analyzing Customer Lifetime Value Using Tree Ensembles" (link).

Analysis based on the customer lifetime value concept must typically account for two main aspects of customer behavior: The lifetime of a customer and the monetary value generated while active.

Given a set of customer attributes, such as demographic data and past transactions, a model of customer lifetime value can be created. The model seeks to relate customer attributes to lifetime and value.

An example of a basic quantitative approach is to model value through a linear regression model, while lifetime is estimated based on, say, an exponential distribution.

A customer lifetime value model has several exciting applications.

• Resource Prioritization. Sensible lifetime value estimates can serve as a basis for organizational resource prioritization: Customer service level design, marketing spending, sales force pay and so on.
• Novelty Detection. Detect and react to changes in the marketplace by monitoring the formation of new patterns in customer lifetime value. Are former valuable clients all of a sudden churning at higher rates? Are new groups of high value clients forming?

Businesses can differentiate valuable customers at an early point in customers' life cycle, monitor the customer base evolution and act on their data on a systematic basis. Such businesses stand to harvest great rewards.

God, Tax and AI

Today, while walking to a client meeting, I was approached by a man.

Man: Are you interested in talking about God and such?

Me: No, not really. I am more interested in making money.

Man: Is it not strange that the Large Hadron Collider recreates circumstances similar to those during the Big Bang, the creation of our universe. The Hadron Collider was built by intelligent people, then, why shouldn't there be intelligence behind the Big Bang?

Me: I am interested in artificial intelligence and hope to one day build a machine with an intelligence surpassing that of humans. However, I would not want the machine to care about its creator. I would want it to solve problems.

Man: Surely a machine would have no conscience, as man, as is witnessed by his enactment of the laws of society.

Me: Who says that man has conscience, that it is not an illusion? Anyway, I think a machine could be programmed to have conscience. ...And when I look at the tax legislation in this country, I really cannot see that as an expression of conscience.

Reinforcement Learning - Tic Tac Toe

Based on Sutton and Barto's "Reinforcement Learning", I did a simple python implementation of tic tac toe.

The algorithm has no knowledge of tic tac toe, except it knows when it has won or lost. By playing against you and itself it gradually learns how to play well.

This particular implementation does not take "exploratory" steps, but it would be simple to implement.

Python code