Techniques We use
Bayesian networks: Bayesian networks allow the modeling of interrelated cause-and-effect events with formulas representing relationships between each pair of events. The network is constructed to simulate the relationship between events, including feedback loops, and can then be used to understand effects of changes in the network of events. Such networks are useful in predicting directional outcomes, e.g. revenue growth, disease state, customer segment responses, market share changes, etc.
Clustering: Clustering places customers (or products, stores, etc.) into homogeneous segments. Clustering, sometimes called segmentation, is generally based on obtainable demographics and/or captured behavior. Clustering allows a company to identify needs of a specific segment and treat members in that segment as the same, improving customer need fulfillment while reducing costs. It can also be used to improve predictive models, such as acquisition or cross-sell.
Combinational Optimization: Combinatorial optimization uses a combination of methods to find a solution in a solution space that is generally so large as to prohibit full exploration by brute force. Using a variety of methods, the solution space is parsed into areas likely and less likely to provide a viable solution. The likely area(s) are then explored to test for a viable solution. Classic examples include the traveling salesman which represents many types of routing and logistics problems.
Conjoint analysis: Conjoint analyses are used to understand how consumers balance choices between products, product features, or services. The tradeoffs are measured by utility; an economic concept allowing common value measure for otherwise unrelated items. Conjoint analysis is most often used to learn which product features matter most to a consumer. For example, if a TV is being considered, would a larger screen, higher resolution, or built-in DVD be more important, and how much more?
Differential Market Basket: A differential market basket analysis (d- MBA) compares products purchased by a consumer, but differentially compares the findings across multiple stores, regions, channels, seasons, etc. Co-purchased products indicate important behaviors. Customers who purchased A also bought B. Tactically, d-MBA improves customers experience and increases sales/margins. Strategically, d-MBA describes customer segments, geographic, or time (weekly or seasonal) differences.
Design of Experiments: Design of Experiments (DOE), is a technique first used in science to determine the statistical significance of variables involved in an experiment. However, in business applications, DOE can be used to gather unbiased statistical understanding of variables in as few tests as possible. DOE is particularly helpful in testing responses to multiple inputs. For example, several different ads, each containing several variations in colors, images, and copy may be tested to determine the most effective combination
Monte Carlo Simulation: The Monte Carlo simulation is more of a concept than a specific algorithm. Most algorithms are deterministic and produce the same result each time they are performed. Monte Carlo simulations are probabilistic and therefore useful in studying problems that have uncertain inputs or result states. Monte Carlo methods can be used to predict financial, medical, or event outcomes when the inputs are not known, only known within a range, or highly interdependent.
Multiple Regression Analysis: Regression analysis is a form of statistical analysis that, in its simplest form, helps understand the relationship between two variables. However, in multiple regression, many variables are analyzed to understand their relationship (if any) to the target variable. This technique can be used to understand the cost/value relationship between an action and an outcome. For example, how much effect does price have on customer retention? Estrategy/Decile uses goes beyond standard nonlinear regression by using splines to describe complex relationships.
Neural Network: A neural network simulates the workings of a brain with artificial neurons feeding data between neurons strengthening and weakening their relationships as more data is processed. Because of this, neural networks are useful in learning situations such as fraud detection, and pattern change detection. Since neural networks use very different methods of prediction, Decile often uses neural networks in conjunction with other modeling techniques to improve total performance.