Predictive Modeling and Programmed Trading in SEM

There’s been a lot of buzz lately about predictive modeling as a better way to deal with the volatile, data-intensive world of the auction search marketplaces. Many of the latest bid management and campaign management technologies have predictive modeling components built in. Before you marvel at the whiz-bang buzzwords, understand “predictive modeling” is just a fancy term for using prior data and the results of that data to make a decision, either for now or for some discrete period of time. There are several levels of predictive modeling. If appropriate, you can use predictive modeling within a campaign regardless of which type of campaign management or bid management system you employ.

I’ll help you understand when and if some level of manual or automated predictive modeling makes sense as either a tactical or strategic campaign management tool. To do that, let’s dig below the buzz to the practical application of modeling within the paid search auction marketplaces.

Predictive modeling has been used in marketing for many years. In email marketing, if you get your best response when mailing on Fridays and decide to mail more often on Fridays, that’s a simple form of predictive modeling. Similarly, in traditional media and marketing, planning and executing campaigns or promotions based on prior data might be called predictive modeling. For example, knowing that adding a coupon for orders over $50 increases your average shopping cart size by $22 on Mondays, $27 on Tuesdays and Fridays, $19 on Wednesdays, and $33 on Thursdays might cause you to plan your coupon promotions to minimize the promotion’s net cost while maximizing return.

Database marketers have also been successfully using predictive modeling for more than 40 years, including recency, frequency, and monetary (RFM) analysis. Use of predictive models in non-search marketing was covered last year in ClickZ by Brian Teasley.

Predictive modeling doesn’t guarantee magical, killer results unless the model is highly predictive and accurate and delivers actionable data to the person or system managing the campaign and making decisions about bids or traffic levels. It uses a combination of recent and older historical data to make decisions. Both new and older data may be useful predictors, but generally more recent data is a better predictor.

In baseball, the batter uses a predictive model to guess what pitch the pitcher will throw next. He uses both historical data (after a slider he always throws a curve ball) and real-time data (the pitcher’s stance, how many runners on base). In the PPC (define) world, search historical data may be almost meaningless if the current marketplace provides us with clear indicators. Often, when models rely too heavily on old data in volatile, competitive marketplaces, the models predict incorrectly and therefore suggest an improper tactic.

There are many drivers to a complete model, including:

  • Current market state
  • Position, to understand upside potential (Note: position used to factor in more heavily, but the engines are moving toward opaque auctions)
  • Market reactivity (or elasticity)
  • Competitive density
  • Daypart
  • Day of week
  • Keyword
  • Engine
  • Ad copy
  • Landing page

In a real-time marketplace, the most recent data can be extremely important, just as it is for the batter. Before you think you’re dealing with data overload, remember the best models are both proactive and reactive. Be sure to test each hypothesis the model suggests, then correct that hypothesis (change the model) as soon as it’s likely a better solution exists. The best predictive model for your business may be completely different from one for another industry.

When building custom predictive models for clients, we often find data supplied at conversion has at least as much impact on the optimal campaign strategy as marketplace predictors (bid-related data). Get to know your customers, particularly the best ones. This information may result in straightforward predictive models that yield significant efficiency improvements.

Now that MSN and Google allow demographic targeting, for example, a simple model that improves targeting may yield double-digit efficiency gains. The more data you have about customers and conversions, and the more data you have about the PPC marketplace and how it responds to changes, the better the models you can build.

Predictive modeling, used correctly, can be both sizzle and steak. When poorly applied, it just sizzles out. Discuss your existing data availability and business objectives with any SEM (define) or technology provider before buying the sizzle.

Meet Kevin at Search Engine Strategies in Toronto, April 25-26, 2006.

Want more search information? ClickZ SEM Archives contain all our search columns, organized by topic.

Related reading