Having grown up in the offline data and predictive modeling world, the focus has always been on finding ways to improve traditional direct marketing solutions for our clients.  This normally centered around appending our consumer data to historical marketing files, developing custom analytics for targeting, and either scoring external lists or extracting marketing lists from our national consumer database.   While this is still a very viable way to sell data and analytics, the online data delivery world is a whole different story.

The online world needs real-time access to consumer data for proactive and reactive marketing and ad-serving.   This is a great new frontier for offering data and custom models for our clients.  Through registration sites, data can be accessed in real-time in order to deliver the best offer/product/service through pop-ups and redirects.  By teaming with ad networks, customized banners can be delivered to members of the ad networks while they are online. 

The integration of our data into the online world is extremely exciting and there is a lot to learn as we bring all of our data to the online marketing community.

Having just returned from DM Days in NYC, I thought I would share some observations:

1). Attendance appears to be way down. I’d hate to venture a guess at how far down it has gone, but it was obvious that the exhibit hall was maybe half as full as last year.

2). There were still some quality speakers (and many self promoting vendors in speaking roles).

3). A relevant crowd still attends the show. I definitely saw a number of senior marketing types at the show.

I understand that the show is moving to the Hilton + Towers next year. Perhaps returning to its former status as a niche show?

We have worked with a number of financial services firms over the years on risk and marketing projects, and are very familiar with the interaction between these competing dynamics. The interesting thing about these influences (and the challenge in dealing with them) is that they are inversely correlated. So, individuals with high credit risk tend to respond well to marketing offers, while those with low credit risk respond poorly to solicitations.

How do you deal with this dilemma? We have used several techniques over the years that have proven to be effective:

1). Develop a custom risk model. In addition to have the benefit of being more powerful than a ‘generic’ risk score (such as FICO), a custom model will carve out a slightly different universe of credit worthy people. It is those individuals that your custom model deems to be creditworthy, while scoring on the margin with FICO, that will respond to marketing solicitations. The FICO score is an excellent indicator of saturation.
2). Develop custom response models only on the relevant credit universe. If your product is aimed at 640-720 FICO, make sure that only individuals in that range are included in a modeling initiative.
3). Segment by risk. In addition to developing on the relevant credit universe, you should consider using FICO as a segmentation tool. For example, you may want to build one response model for individuals scoring 640-680 and another for individuals scoring 680-720. Why? The drivers of response can be quite different by credit strata.

I would not include the FICO score in a response model. Due to correlation issues, doing this would often remove the benefits of segmenting by FICO.

In summary, use the FICO score (which is a highly accurate gauge of the type of credit offers being sent to each individual) as a tool to segment the universe. This will remove the ‘credit bias’ that will otherwise dominate a response model. Additionally, utilize a custom risk score to assess the true credit risk of your market.

While the recent volatility in the economy has, for many companies, put a damper on marketing results so far this year, there are some emerging opportunities.

My firm has done a lot of analysis around utilizing econometric data to improve marketing models. The result: we have found that this data, when used correctly, can add 5% to 15% lift in predictive power at a minimal cost.

In performing this test, we collected data from a number of government/pseudo-government agencies at the Metropolitan Area (CBSA) level. Some examples of the type of data collected include Unemployment Filing Claims, Retail Sales, Bank Deposits (among many others). We selected certain attributes (driver elements) and developed time-series models to capture the trend line. These time-series models were then used as incremental inputs to existing marketing models to gauge the lift provided to a hold-out validation dataset. In a number of trials, we always achieved a K-S lift of 5%, and occasionally achieved a K-S lift of greater than 10%.

Does this make business sense? We think that, everything else being equal, individuals that appear to be homogeneous using ‘traditional’ marketing data will perform differently based on real-time trends in their neighborhood. For example, assume that Person1 and Person2 have identical marketing demographics (income, age, etc.), but that Person1 lives in Seattle and Person2 lives in Atlanta. As it happens, Person1 is lucky enough to own a home that is steadily increasing in value, and neighbors that are planning trips abroad and major purchases. Alas, poor Person2 is well aware that his home is decreasing in value (as evidenced by the sea of ‘For Sale’ signs on his street), and happens to have several neighbors that have recently lost their jobs. Do they have the same purchase behavior?

We believe that these results do make sense, and potentially offer a ‘new frontier’ for predicting response/purchase behavior.

Here are the leading ways to NOT differentiate yourself in direct marketing:
1) Build a massive prospect database. I have had a front-row seat for a number of DB builds over the years, and have come away convinced that large-scale prospect databases are only for the largest DMcompanies. I think that most DM companies need a prospect datamart(i.e. consisting of a representative sample of the prospect universe) combined with a well-defined campaign execution process.
2) Use standard compiled data. The vast majority of DM companies use the exact same compiled data to produce extract lists. These names are tired, and the data is highly inaccurate. To gain an advantage, seek out vendors that claim to have unique data, along with the standard data.
3) Overuse vertical lists. While this can vary greatly by product, I tend to think that large DM marketers should carefully test any vertical lists before integrating them into an overall strategy. In particular, it is important to understand whether the vertical names (which are usually far more expensive than compiled names) would have been sourced anyway via another channel.
4) Use unbenchmarked modeling. I believe that 60% of firms claiming to offer ‘analytics and modeling’ do not know what they are doing. By this, I mean that they produce over-fit (or otherwise flawed) models that do not work well in practice — although they may look great on the development data. All serious DM companies should get an independent assessment of their modeling from time to time.
5) Look forward, not backward. I frequently come across DM companies that spend little, or no, time analyzing what worked and did not work on past campaigns. Important items such as creative, message, offer, etc. can be optimized for the future based on past results.
Best in class DM companies focus on an analytical approach to campaign management.

As an analytics vendor for the past fifteen years, including the last ten years as the head of two different modeling/analytics companies, I have some definite thoughts regarding best practices in selecting a provider.
1) Competency — I believe that only about a third of companies that sell analytical services are consistently competent. By this I mean that these companies always (or almost always) produce analytic solutions that are competitive with what the best vendors would produce using the same data. This impression comes from numerous validations, assessments and benchmarking projects I have been involved with over the years. How do you gauge competency? Here are a few thoughts:
a). The best way is to have the vendor develop a test model, where some of the data is held out as a validation sample. I would expect that the results would be consistent between the development and validation samples.
b). The product should, in most cases, make business sense. In line with this, each attribute should be logical and the ‘signs’ should match — meaning that an attribute that has a positive correlation with what you are trying to predict should not have negative points.
c). I would make sure that the vendor does NOT use a purely automated modeling system. With the current state of the technology, I firmly believe that an experienced modeler will produce a better solution. The ‘art of modeling’ has yet to be automated…..
2) People — While many/most vendors have some experienced people on their staff, it is important to assess the experience and reputation of the people that will actually work on the project. I would recommend working with a smaller vendor that can provide an experienced team, rather than selecting a larger company that will assign a relatively inexperienced team.
3) Interest — Based on my experience, you are always better off with a vendor that views your account as a ‘big deal’. If you are not a big deal to your vendor, getting relevant attention is going to be a problem.
4) Reputation — While word of mouth can sometimes be misleading, it is at least directionally correct. If a vendor has a spotty reputation, it is important to find out why.
Of course, all of the above assumes that there is a strong and synergistic relationship in place between you and the modeling vendor that you select.