5 steps to building a predictive analytics project
The benefits of predictive analytics were highlighted by a survey of 833 business professionals by The Data Warehousing Institute (TDWI) in Feb 2014. These included a bank increasing it's campaign return on investment by 100 per cent and an airline boosting customer satisfaction by deducing the number of passengers who would miss flights.
It is no wonder that so many companies across the world are talking about applying predictive analysis to marketing information. Don't be too quick to bite, though. Aggregating customer data to predict audience behaviour isn't as simple as pressing a few buttons. Although the technology behind forecasting is improving, marketers can't let the software do all the legwork.
Before jumping head first into a predictive analysis project, there are five things marketers need to do:
1. Define your dataset
We're about to get a tad mathematical here, so bear with us. A recent article from computing giant IBM described how to develop predictive models. The first step in doing so is to create a dataset that has a specific identity. This dataset describes a customer, and may contain information such as gender, age, frequently purchased products, value of items bought and so forth.
Once you organise data on multiple patrons (Note: pulling info from your customer relationship management (CRM) application will be a big help, here) you'll have the raw material required to estimate outcomes.
2. Qualify your information
This is where contact validation solutions can help. You want to make sure that the information you're about to analyse is accurate and thorough. Dean Abbott, president of Abbott Analytics, outlined a procedure to help professionals 'clean' their data:
- If you pull data from external sources, ensure those parties are reputable. In addition, clarify the manner in which an outside organisation gathered and structured the material.
- Highlight the attributes within certain datasets and create subgroups. Document these changes and note why you classified them the way you did.
- Format your data. This involves identifying their 'type' (Boolean, integer and real, for example).
3. Know what you want to predict
Maybe you want to anticipate how many customers will purchase a particular product or service. Perhaps you desire insights into the number of churned clients for the next quarter?
Pick and choose relevant datasets from your data warehouse. So, if you want to know the answer to the first question introduced in this section, select datasets that describe patrons who have bought a specific item. If you get stuck at this step, you can employ audience profiling practices to help you find the appropriate datasets.
4. Test your algorithm
Now is the chance to take your algorithm and 'train' it - to use TDWI's jargon - to scrutinise the information you've selected. If you don't know what an algorithm is, it's just a set of instructions for completing a process or solving a problem.
Remember, when TDWI says 'train', it's not talking about testing a predictive model to see if it produces the results you want to see. Tailoring your algorithms or omitting data to produce favourable conclusions is not conducive to identifying an accurate prediction.
5. Apply your model
Once you have a strong idea of how your algorithm will respond to the datasets you've selected for analysis, you can position your predictive model against them. Don't be discouraged if you don't receive an outcome that doesn't appear overtly valuable.
TWDI provided a good example of a strange result: when a grocery store discovered a correlation between beer and diaper sales. From the surface, there's not a whole lot a business could do with such intelligence. However, maybe the grocery store places beer and diapers at opposite ends of the outlet, exposing customers to as many products as possible?
In other words, don't be discouraged by the predictive conclusions. Being afraid of creativity or taking chances isn't going to help you act on this intelligence, either. Bottom line: don't hesitate to apply knowledgeable changes, no matter how bizarre they may seem.
