Advanced Analytics

A new class of nonlinear analytic techniques now promises to deliver step-function improvements over traditional linear modeling approaches.  Adaptive, flexible, and able to “learn,” these next-generation models are already far outperforming the status quo. 

These are analytic tools for the era of data explosion, built to leverage and continuously learn from massive amounts of information. They combine computers’ tireless computational power and vast capacity with the elastic, connective, intuitive, and learning capabilities of the human brain, revealing patterns, correlations, and segmentations hidden deep within enormous data pools.

As our performance in the Netflix Prize illustrates*, Opera Solutions is among the world’s most accomplished consulting firms at applying these advanced analytic approaches to solve business problems.  We have over 70 scientists, machine learning experts, mathematicians, and statisticians, resident in three analytic and technical Centers of Excellence, who focused on developing advances in this area. 

As the chart below illustrates, our modeling approach is pragmatic and informed by “real world” requirements and considerations. We commit to providing explainable, adaptable, and interactive models that deliver significantly better performance.



We have developed innovative solutions and approaches to many different classes of problems, relevant to a wide range of businesses and disciplines:



These techniques have wide applicability; we are already using them to deliver superior results in:

• Credit risk behavior scores
• Point-of-compromise for skimming fraud
• Future cash streams from distressed mortgage assets
• Valuation of commercial and residential asset-backed securities
• Collections treatment optimization and settlement offers
• Real-time price optimization for auto auctions
• Loyalty-based offer management for retail customers

For example:

Risk Platforms.  For a Top-5 credit card issuer, we are using a suite of nonlinear approaches to develop new risk platforms.  We are significantly outperforming the client’s existing 20-year model (widely recognized as one of the industry’s best) and are outstripping FICO scores by 20-30%. Click here for case study.

Customer Acquisition and Marketing Channel Optimization.  Traditional linear methods cannot create predictive relationships between specific marketing channels and new customer acquisition; new techniques can.  We have used nonlinear modeling to help one client develop a cross-channel approach that maximized total ROI and improved acquisition performance by 6X.

Customer Management and Cross-Selling.  We are using models to predict each consumer’s sensitivity to all available offers.  Armed with this information, we are developing a custom “curriculum” — that is, a series of offers — for each customer. Even more powerful: these models have a unique capacity to "learn." A feedback loop adds back the actual response to each offer into the analytic model, and the model then adjusts and optimizes. As important, these models can predict who will be neutral to a particular offer, ensuring that the company does not allocate marketing and promotional investments to unlikely prospects. We are also using these techniques to dramatically improve cross-selling effectiveness, improving revenue at one retail client by 15%.
Click here for case study.

Reducing Adverse Actions.  Versus traditional models, next-generation computational techniques significantly improve the accuracy and confidence interval of risk scoring.  For example, a Top-5 credit card issuer tightened credit policies and reduced credit limits among a large group of consumers to reduce the company’s latent risk exposure, but this generated intense consumer and media backlash. Using new analytic techniques, Opera was able to eliminate adverse actions for 20% of the population without increasing risk exposure. Results:  improved loyalty, reduced customer complaints, and an additional $20MM in annual profits. 

Predicting Bust-Out More Accurately.  These techniques are also able to pick up “faint signals,” when traditional linear models cannot. Bust-out fraud (when a seemingly good customer suddenly maxes out credit lines and defaults) is extremely hard to find through linear modeling. Opera’s nonlinear approach has not only identified more bust-out candidates, but has found them days earlier, saving one company over $75MM annually. Click here for case study.

Improving Float Decisioning.  For financial services institutions, float management strategies based on nonlinear analytics are also an opportunity to make better, more profitable, (and sometimes counterintuitive) customer decision. Deciding how much of a payment or deposit to make available to a customer before the check clears is as much a revenue opportunity decision as a risk mitigation strategy. This is a complex evaluation, weighing the benefits of incremental spending and fee revenue versus credit risk. We have found that the optimal result is achieved by crediting more accounts earlier — even higher risk accounts, whose risk is offset by fee revenue.

	The Techniques:  a Quick Primer Techniques: Opera uses the full panoply of new nonlinear and multivariable methods, melding them together through advanced ensemble approaches. These techniques are broadly applicable to many classes of problems in Credit & Risk, Marketing, Portfolio Evaluation, and Collections problems. Below, we provide more detail on these approaches and what they can deliver: TECHNIQUES OVERVIEW Neural Networks Neural Networks uncover hidden, nonlinear relationships between people’s characteristics and their behavior, for example Neural Networks are particularly effective at finding hidden structures and correlations within massive amounts of data, as well as dealing with nonlinearities, diverse data, and interrelated variables. Matrix Factorization (also called Singular Value Decomposition, or SVD) SVD is a latent factor linear model that can, with high accuracy, predict preferences for large numbers of customers and items, including who will buy what or who will be motivated by a specific offer. It creates “behavioral DNA” — a map of each individual that can be used to describe, group, and ultimately predict his or her actions. This approach helps categorize individual customers by characteristics and behavior across large numbers of categories, i.e., purchases, payments, rewards, or delinquency. SVD provides an enhanced ability to predict behavior, develop robust segments, and know customers. It has particular strength in dealing with missing data and developing predictive segmentations. Restricted Boltzmann Machines (RBM) RBM is excellent at finding correlations and predicting outcomes from incomplete data. It "trains" on massive amounts of information, uncovering hidden structures and correlations. Then, it compares new information to the patterns and correlations it has already identified. RBM is a natural framework for incorporating massive amounts of unstructured data. It is fully extensible and flexible, easily adding in new data as they are generated. RBM has particular strength in dealing with nonlinear interactions and missing data and in producing probabilistic output, so it is applicable to problems in fraud, default prediction, and profitability modeling. K-Nearest Neighbor (KNN) KNN works on the theory of “birds of a feather flock together.” It finds important and predictive similarities that improve modelers’ ability to find segments that will act alike, as well as to assign new people or products to segments. KNN models learn from recent events and adapt accordingly, making them highly powerful and allowing them to dramatically improve business results. Ensemble Ensemble knits together the results from different analytic techniques. Its use typically results in a better-performing overall model, because it can draw on each individual technique’s unique strengths.  By combining models, Ensemble yields a final prediction that is more accurate than any one approach (for instance, it doubled the effectiveness of any single algorithm in Opera's Netflix Prize model).

* Opera, as part of The Ensemble, matched the Grand Price Winners' performance, but submitted its model twenty minutes later.