Last full session today is Lisa Kart and Jean Zoch of Fair Isaac talking about optimal pricing – balancing profitability with competitiveness. Lisa’s focus, she says, is on what makes price optimization work. Price optimization is a very broad topic, even in financial services, but their focus is on being able to target price for different consumers.
A recent survey showed that over half have adopted some form of price optimization and more than 3/4 plan to use if by 2012. Looking for profitability in difficult circumstances. Another survey shows that larger banks are using optimization more than mid-size or smaller ones but that the rate is increasing in every category.
Price optimization in financial services is somewhat unique – you have a lot of data about consumers, many objectives and constraints and different strategies to assign products to different consumers. The objective is to find the optimal price for different consumers and consumer segments to maximize profit given these boundaries. For instance, in credit cards one might be optimizing initial rates, long term rate, balance transfer deal and other factors. In account management might be optimizing rate increases balancing attrition, risk and revenue. In loans or hybrid products, might be optimizing discounts, terms and more. All of these are price optimization projects that Fair Isaac has completed. These projects have huge potential payoffs – 7% profit increase despite strong constraints, 35% profit increase while maintaining market share or 45% profit increase with losses held constant. Very impressive numbers, considering most lenders are already pretty sophisticated.
Price optimization is hard for various reasons:
- Hard to know how consumers will respond to different prices
- Balancing trade-offs like risk v profit or take-up v profit is complex at both the consumer and portfolio level
- Pricing decisions are constrained by the overall portfolio – can’t just optimize for the individual consumers, must optimize the portfolio too.
- Data holes and biases
- Integrating pricing decisions with other decisions like loan amount or type
You really need to make true estimates of customer price sensitivity and elasticity, anticipate changes to the environment and still carve out the most value from your decisions while meeting your constaints. Jean come up to talk about how to address some of the challenges.
There are multiple steps to optimized pricing strategies:
What data do you have? How good is it?
Identify objectives and constraints and really design a decision model to suit the situation. Optimization is a mathematical process but the objective and the function to be maximized must be known, well described and correctly constrained.
Action-effect models and a decision model need to be developed. Action effect models are key as it is no longer enough to model probability, must model actions and their effects. Validating and diagnosing these models is different.
Set up the constraints, run scenarios and generate efficient frontiers. This is where the optimization engine, the solver, runs.
Financial projections of multiple what-if scenarios, understanding trade-offs and drivers
- Deployment of the optimal strategy
Build a decision tree, a ruleset, that implements the optimal approach
Two of the biggest challenges are developing a consumer price sensitivity model and balancing tradeoffs within the constraints of the overall portfolio. To build a model of consumer price sensitivity you need to identify the key interaction variables with price. For instance, a profile variable is modeled against the price to create a two-dimensional model of price sensitivity and this is built into the action effect models. Choosing the right interaction models is hard as you are looking for variables with a strong interaction – those where a change in the interaction variable really impacts how the consumer reacts to price. The data for this can be hard to come by as historical data is biased towards the historical approach but over time many different prices may have been offered to similar customers, creating useful data. However, randomized testing can be critical to getting a robust set of data – experimental design makes a huge difference to the quality of data. Similarly, expertise and experience can be applied to the models where data is missing or of poor quality – what Ian Ayres would call human support of an analytic process. Smoothing of the curves is the last step.
The decision model or influence diagram, takes the various action-effect models and the various decisions that could be taken to see how each influences the other. Each influence can be described mathematically, leading to a chain of influences that impact objectives – a map of the decision model. Many different elements have an effect. Lower prices increase take-off but decrease profit AND increase chance of pre-payment and when they might pre-pay. Thinking through all these influences is both hard and necessary – what will, in fact, influence the price sensitivity and profit.
Efficient frontiers are generated from the solver to see how changes in constraints, what-if scenarios, would impact the overall profitability or objective. The impact of changed constraints on the overall, optimal scenarios can be seen and decisions taken about changing constraints to boost results.
Lisa wrapped with the two key takeways:
- Spend time getting the price sensitivity relationships right despite the historical biases, holes in data etc
- Explore the tradeoffs between objectives and constraints to get the most value, making sure you should not change the constraints before optimizing
There was a fair amount of detail in this show that I wasn’t able to capture but it is clear that the combination of decision models (influence diagrams), action-effect models and a good solver can really make a difference when it comes to pricing. That’s it for today, more tomorrow.