I’d appreciate an advice about the following data mining scenario. A bank institution has customers and keeps tracks of the customer demographics data (e.g. income, age, etc). A customer could have purchased one or more products (checking, savings, CD, and other accounts). My task is to predict the likelihood for a customer to purchase a new product. For example, if the custom has checking and savings accounts, what’s the likelihood for the customer to purchase a CD account so we can offer this product do the customer.
My questions:
- What data mining task (segmentation, classification, etc) does this requirement represent?
- What algorithm(s) is best suited for this task?
- For a given customer, will I be able to predict the buy probability for all products or do I need to query the mining model to predict one product at the time?
1. It makes sense to first utilize MS Clustering see learn more about what customers have in common. This will help determine what key drivers (e.g. geography, demographic, socioeconomic, tenure, etc.) are more important that others in recommending products.
2. MS Decision Trees and Neural Nets will do a great job of identifying what "x" product to offer to "y" customer. Each alogirithm has its own benefits.
3. If you properly set up the model you can set the output to deliver a scored list of product recommendations for each customer record instead of one product at a time. Remember the point of a product recommendation model is to provide multiple products that a customer may be likely to sign up for (think Amazon.com they offer multiple recommendations not just one at a time). For example, the models we have developed for banks will predict a cross-sell product (Mortgage), an up-sell product (gold membership), an add-on (checking plus) to each scored customer record.
Hope this helps.
Jeff
Please visit Apollo Data Technologies http://apollodatatech.com/|||
Jeff,
Thank you so much. I will definately try out your recommendations. If you don't mind, I would appreciate some additional pointers. Our system captures the customer products (checking, savings, CD, etc), account balances, profitability factors, and lmited demographics data (address and age). A big concern I have is that that input parameters may not be sufficient for a good cross-selling prediction.
1. Based on your experience with the banking industry, what could be the most important factors that may influence the customer (individual or business) to buy a new product?
2. Are there any companies that could provide tha missing important inputs (demographics and financial) if our system and the bank cannot supply them? Any idea about how much such services can cost?
I appreciate your help.
|||Shrek - good questions...1. Your first question on key drivers is the holy grail. The key drivers really vary based on many factors - is it a national or regional bank, full-service, strong online presence, how many branches, etc? What I can say confidently without knowing the above answers and looking at the data is that income level and geo/demo are often key drivers. The best way to accurately determine this is to build a clustering model utilizing the MS Clustering algorithm.
2. Yes - there are many companies that are more than happy to sell you geo/demographic and socioeconomic data (at the individual level, household, business) you can overlay ontop of your customer records to improve the richness of the data set. The price ranges based on how many records and overlay attributes you determine you need. We have built a service around identifying which 3rd party data will be the best predictor for identifying product recommendations (these vendors will provide a free sample for evaluation). As a heads up this service pays for itself b/c often times the 3rd party data sold is of poor quality; however, our process identifies the best data sets available and the minimal amount of data needed to purchase so you are not over-buying data that is of little use.
Feel free to contact me if you need some additional pointers info@.apollodatatech.com
for the inner input, meaning