PEGA Decision strategy manager (DSM)
- 1. Precursor to PEGA – Decision Strategy Manager
- 2. Contents • Overview • Major Components of DSM • Introduction to Predictive & Adaptive Analytics • Differences between Predictive & Adaptive Analytics • Decisioning Components • Segmentation • Data Import • Data Enrichment • Arbitration • Selection • Aggregation • Decisioning Rules
- 3. Overview • To be able to make the right decisions, become more effective, and deliver more relevant actions for the customer, we need to be able to determine customer interest in a product and their likely behaviour. • Decision Management capabilities help businesses optimize every customer relationship and interaction to satisfy both customer needs and business objectives at the same time. With Decision Management, business users can implement a management strategy that is personalized for each customer, and that guides every customer interaction and decision. • The Decision Strategy Manager (DSM) allows marketing managers to participate more directly in the evolution of the applications by including facilities that support cross-sell, upsell, retention, and risk management. • The Decision Strategy Manager (DSM) delivers: • Proposition management • Strategy development • Driving process flows using interaction, scorecard, and predictive model rules • Using third party models • Multilevel decisioning • Single and distributed batch execution of strategies • Capturing interaction results using Interaction Services (IS), and associating interaction records with work objects • Visualization, monitoring, and forecasting using Visual Business Director (VBD) • Advanced adaptive analytics using Adaptive Decision Manager (ADM)
- 4. Decision Strategy Manager (DSM) • The core decision management component, DSM allows business users to design customer interaction strategies and propositions based on decisions and rules that reflect customer behaviour, preferences, legislation, corporate policies and desired business outcomes. • The decision model is highly visual, and consists of a rich set of decision components that allows business users to create strategies directly on a canvas. Adaptive Conversation Advisor (ACA) • The user interface for agent-driven channels, leveraging the decision management capabilities to make Next-Best- Action recommendations. Different conversation styles can be nominated such as assessment, negotiation of offers, bundling, top-n offers, Q&A, etc. • ACA assesses everything known and said by the customer in current and previous interactions and recommends the Next-Best-Action to be taken. MAJORCOMPONENTSOFDECISIONMANAGEMENT
- 5. Adaptive Decision Manager (ADM) • Adaptive Decision Manger provides self-learning models that are able to learn from real-time customer behaviour and adapt on-the-fly. ADM is especially helpful in situations where there is not enough data to create a robust model, there are too many different products to create predictive models for, or where behaviour is very volatile. • Once initiated, ADM will begin learning and indicate when it has gathered enough data to begin making accurate predictions on its own. ADM also includes a monitoring capability to track the performance of the models. Predictive Analytics Director (PAD) • Is a predictive tool that offers a streamlined business-oriented process to quickly and safely develop accurate and reliable models that predict customer expectations, propensities, and behaviours. • PAD focuses on the complete process of data analysis, model development, model analysis, and deployment of predictive models. Visual Business Director (VBD) • Enables planning, monitoring, simulation, and control of the customer experience across all customer segments and product lines in all channels. • It allows business to determine certain control parameters – effectively making tactical changes to the decision model. VBD offers a highly visual 3D user interface. MAJOR COMPONENTS OF DECISION MANAGEMENT
- 6. • Predictive analytics make predictions about future events by analysing current and historical data and applying it to propensity models to make and display calculated predictions of the likelihood that a certain event will occur. • The models that are created are based on a snapshot of data and are refreshed or rebuilt at regular intervals. Predictive models could be used in a number of operational or customer service related contexts. For example, a customer service organization that is using predictive models might be collecting purchase behaviour data, competitive information, demographic data, and other data used to proactively provide recommendations on what the customer might be willing to purchase based on their history. • Adaptive analytics use an algorithm that learns about customer behaviour in real-time, instead of using a snapshot based on a predictive model. After each response to a proposition the model adapts, resulting in increasingly accurate decisions. • Adaptive decisioning can calculate who is likely to accept/reject a proposition without little prior experience. It also captures and analyses data to deliver predictions where customer behaviour data is volatile. For instance, if a customer is offered a product and accepts it, the likelihood score of customers with a similar profile will increase slightly. • Predictive and adaptive models can be used independently or together in a champion-challenger approach. Pega has capability (Predictive Analytics Director and Adaptive Decision Manager), but also provides import of predictive models developed with 3rd party tools such as SAS or SPSS. PREDICTIVE&ADAPTIVE ANALYTICS
- 7. When to use Predictive Vs. Adaptive Predictive Adaptive Predictive Analytics requires historical data containing the behaviour we want to predict Adaptive Analytics can start from a blank sheet of paper and learn from customer interactions. Greater Control as we are directly involved in the development process which gives us more influence over the model. Adaptive Analytics has greater automation capabilities Predictive Modelling uses a range of powerful techniques and as a user we can select the best model for our needs. One straight forward modelling technique Predictive Models should be used when predictability and compliance is important. For example, risk related behaviour such as credit risk, claims risk or fraud, etc. Adaptive models learn constantly and therefore their performance will change over time, which makes the outcome less predictable than that of predictive models. For example: Product Acceptance, Routing, Fulfilment. Tens of Models. Don’t get a very quick response for some behaviour types. It can take months or even years to accumulate data. Predictive modelling should be used in these situations. Possible to implement hundreds of adaptive models in a solution because of the high degree of automation Longer delay for outcome Shorter delay for Outcome
- 8. Decisioning Components • The core function of Decision Management is to create and maintain Decision Strategies. Business users can define Decision Strategies using a palette of Decisioning Components. • The Decisioning Components are the building blocks of every Decision Strategy and each Decisioning component has a business purpose. • There are 16 Decisioning Components in total, which are grouped into the following 6 categories Segmentation Components: These are used to define sub- groups of customers that need to be treated differently. Typically the segmentation is based on the customer’s likelihood of interest in a proposition and the relevancy of a proposition to that customer. Data Import Components: These allow us to bring data into the context of our Decision Strategy, for instance information regarding our proposition or Recommendations that come from other Decision Strategies. Data Enrichment Components: These can be used to add additional data to our Decision Strategies, allowing recommendations to be even more personalized and context sensitive. Arbitration Components: These enable us to arbitrate between propositions by filtering out irrelevant propositions on the basis of, for instance, eligibility rules, and then prioritizing the remaining relevant propositions. Selection Components: These allow us to choose between different options. For instance to determine what business issue is more important than other business issues and under what circumstances. Or more randomly to test new propositions. Aggregation Components: These give us the ability to calculate aggregate values, for instance counting the numbers in a list, or calculating an average for a range of values.
- 9. Decisioning Components – SegmentationThe “Segmentation” Decisioning Components determine in which group, or segment, a customer falls. A different strategy, and therefore, a different treatment, can be applied to each segment. There are 5 segmenting components; each employs its own approach to creating segments. We can add one of each to the canvas by right clicking on the canvas and selecting the one we want from the Add tab. • Decision Table: The Decision table is used to segment customers by properties. For example, say we are creating rules for approval of a mortgage loan. Customers in a Decision Table would be evaluated based on properties such as: Salary, Age, and the Amount of the Mortgage. If the customer is not accepted, the same properties are used to determine if the customer will be referred to a specialist, or rejected. • Decision Tree: Decision trees offer a bit more flexibility because different properties can be used for different evaluations. For instance, if the customer’s mortgage application is not accepted based on the properties that were used in the decision table, other criteria, such as arrears history, can be used to determine whether their application will be referred to a specialist. • Predictive model: Used to segment customers based on Predicted Behaviour. These segments are also known as behavioural segments. Being able to predict the customer’s likelihood of interest in a certain proposition helps determine if and how we want to recommend that proposition. The predictive models used in this component are created in and imported from the Predictive Analytics Director tool. • Scorecards: These are broadly used and most commonly known in credit scoring. The scorecard component allows segments to be defined based on ranges of credit scores. Scorecards provided by third parties or created with Predictive Analytics tools can be implemented in the Decision Strategy using this component. • Adaptive model: This is similar to the Predictive model in that it allows segmentation based on predicted behaviour. The difference is that adaptive models learn from data gathered during real-time interactions. And Adaptive models can be defined in the Decision Strategy.
- 10. Decisioning Components – Data Import Decision Strategies use data as input, and make decisions based on this data. The majority of the data items that are used can be found in the Data Model. Data Import decisioning components provide access to additional data that lives outside our Data Model that can be mapped to our strategy properties. There are three types of Data Import Components. • Type one is “Sub Strategy”: The sub strategy data import component is used when we want to leverage the output from other Decision Strategies as input for the current Decision Strategy. For instance, the Next Best Action Strategy can have references to multiple strategies, such as cross-sell, retention, education, etc. The Next Best Action decision strategy uses the output from those strategies to determine the Next Best Action. • Type two is “Proposition”: A proposition is a potential action considered by the decision strategy. Which proposition the decision strategy ultimately recommends is determined by a myriad inputs, like the margin made by the organization, the cost to the organization, and the benefit to the customer. • The “Proposition” component leverages the data as defined in the Issue-Group-Proposition hierarchy on our Strategies Landing Page. • Type 3 is “Data Import” : This component is used to access data that is not directly available in the data model from our Apply to Class. With Data Import we can map to other Pages in the system For instance an invoice data coming from a billing system that is available on a Page elsewhere in the system.
- 11. Decisioning Components – Data Enrichment Components in the “Data Enrichment” category are used to add extra information to our decision strategies. • Strategy Set: Strategy set components enrich data by adding information to the components they are connected to. Using strategy set components, you can define personalized data to be delivered when issuing a decision. Personalized data often depends on segmentation components , and includes definitions used in the process of creating and controlling a personalized interaction, such as: Instructions for the channel system or product/service propositions to be offered including customized scripts, incentives, bonus, channel, revenue, and cost information. Probabilities of subsequent behaviour, or other variable element. • Data Join: Data join components import data from an embedded or named page using a key to match data, and map its contents to properties from the imported data to strategy properties.
- 12. Decisioning Components – Arbitration & SelectionArbitration Decisioning components filter propositions based on priority and relevance. There are two types of Arbitration Components: Prioritization Filtering. • Prioritization: We can rank propositions and then select the best or most relevant propositions for a customer or group of customers. In reverse, we can rank a group of customers based on their predicted likelihood of interest in a set of propositions and match them up that way. • Filter component : We can filter out propositions that are not relevant for the situation or that we don’t want to offer, such as credit cards, which we will only offer to people 18 years and older. Selection group of Decisioning components. allow us to choose between options, such as one proposition or a group of propositions. • Switch component : Used to select between options based on business rules. A Next Best Action decision uses the Switch component when it selects between business issues such as Sales, Retention, Risk Mitigation, etc. • Champion Challenger: Component enables a random selection. Champion Challenger is typically used to test different variations of propositions. Where the Champion is the mainstream proposition and the Challengers are alternatives that we want to test to hopefully select a new Champion.
- 13. Decisioning Components – Aggregation Aggregation Decision Components give us the ability to make calculations from a list of values. • Aggregation: We can use the aggregation component to set the value of a strategy property based on an aggregation of values from a source component. We can for instance calculate the total number of payments and the average value of those payments. • Financial: The “Financial” component can perform one of the following functions: • Net Present Value • Internal Rate of Return • Modified Internal Rate of Return
- 14. DSM : Rule Types • Scorecard • Predictive Model • Adaptive Model • Strategy • Interaction • Decision Management introduces five new rule types that enable us to implement sophisticated Decision Strategies in PegaRULES Process Commander (PRPC). • Decision strategies are defined in the Strategy rule type. The strategy rule type allows us to arbitrate between business issues and prioritize across a variety of propositions, recommendations and actions. • The strategy rule type can leverage predictions derived form three new rule types – Predictive, Adaptive models and Scorecards • The scorecard, predictive and adaptive models can also be directly invoked from the decision shape in a process flow. • To invoke a decision strategy from a process flow we use the interaction rule. The same interaction rule is also used to capture customer responses, which are used to monitor the effectiveness of the decision strategies and learn and adapt where needed.
- 15. Scorecard • Scorecards are a well-known method to predict or score the likelihood of a certain behavior. They are widely used to determine customer credit scores • The first column shows the properties that are included in the scorecard • In the second and third column we can see the “Condition” and the “Score” • Final Score = Weight * Score • The combination of all these scores leads to one score for the whole scorecard. This score is calculated by the “Combiner Function” – SUM, MIN, MAX and AVERAGE
- 16. Scorecard • We specify the cutoff values on the results tab, which is similar to what we would do on a standard decision table • Score ranges : The minimum and maximum scores are calculated based on the Combiner Function selected on the Scorecard tab • Result : Enter a name for a decision result • Cut Off value : Enter a numeric cut off score value for the result based on the minimum and maximum scores displayed at the top of the tab • Audit Notes : Check this box if you want to capture scorecard details in the case history
- 17. Predictive Model • Use a predictive model rule to import a model produced by the Predictive Analytics Director (PAD)product and relate its outputs to properties in your application. • Predictive models can support your strategies for customer retention, risk management and maximization of customer lifetime value..etc • Predictive Models are developed on the basis of historical customer behavior data • Upload .OXL file i.e. exported from PAD / any third party tool.
- 18. Predictive Model • On the “Input Mapping” tab we can see the input fields, also known as predictors • The “Field Name” and “Field Type” come from the predictive model we uploaded and are therefore defined by the data set that was used to develop the model.
- 19. Pedictive Model • The “Results” tab shows that the predictive model classifies a customer into different behavioral classes. • Each class has been detected by Predictive Analytics Director as having statistically different when it comes to, in this case, their probability of Churn behavior. • Churn Rate: (Values between 0 and 1) Probability of the customer being loyal to the product he / she are purchasing. FORMULA: churned count / (churned count + loyal count) • Lift: Lift is a measure of the effectiveness of a predictive model calculated as the ratio between the results obtained with and without the predictive model. FORMULA: (churned behavior of classification * 100) / total churned behavior The “Edit” button gives us the ability to combine classes.
- 20. Predictive Model • Additional meta data about the predictive model and the development process is collected by Predictive Analytics Director. This information is shown on the “Statistics” tab. • The “Attributes” section contains information about the configuration of the predictive model, including information such as who created the model, when the model was created, What kind of behavior it predicts, etc.
- 21. Adaptive Model • Adaptive Decisioning is an algorithm that learns about customer behavior in real-time from interactions with customers. For instance, whether they accept an offer presented to them by an agent in the call center. • The adaptive model will learn from each response that is identified as positive or negative. When a customer responds positively to a proposition, the model changes so that customers with a similar profile will get a higher likelihood of interest. • In this example of the “Sales Offers Model” adaptive model we see on the “Configuration” tab which of the properties can be used as predictors
- 22. Adaptive Model • The “Settings” tab is used to configure the Adaptive Decision Manager node • The “memory” property specifies what number of responses from recent interactions will be used to calculate the model. All evidence from interactions will be used if the value is left as Zero, otherwise the specified number of responses will be used to create the memory. A memory of 1000 responses is a very short memory, meaning the model will adapt quickly when behavior changes. A memory of 100,000 responses will result in a less volatile model and will therefore learn more slowly from changes in behavior.
- 23. Adaptive Model
- 24. Strategy • This rule allows a strategy designer to model sophisticated business decision strategies. • The “Strategy” tab is a canvas that allows us to graphically compose the decision strategy • Strategy rules are used in interaction rules, and in other strategy rules through the Strategy component • The Decision Strategy rule has a very powerful capability which enables us to arbitrate between different directions and prioritize between many options
- 25. Strategy • Right click on Work area and add components to your strategy • A strategy is defined by the relationships of the components that are used in the interaction that delivers the decision
- 26. Strategy - Data Import Components • Proposition components import propositions defined in the proposition hierarchy • Sub strategy components reference other strategy rules. They define the way two strategies are related to each other, access the public components in the strategy they refer to, and define how to run the strategy if the strategy is in another class • Embedded page components import data in an embedded page • Named page components import data in a named page
- 27. Strategy – Segmentation • The output of a scorecard rule is a score. • The output of a predictive model rule is statistics generated by the PAD model that provides the prediction. • The output of an adaptive model rule is a partial list of adaptive statistics (evidence, propensity, and performance).
- 28. Strategy - Data Enrichment • Strategy set components enrich data to define personalized data to be delivered when issuing a decision. Personalized data often depends on segmentation components, and includes definitions used in the process of creating and controlling a personalized interaction, such as: Instructions for the channel system or product/service propositions be offered including customized scripts, incentives, bonus, channel, revenue, and cost information. • Data join components import data from an embedded or named page using a key to match data, and map its contents to properties from the imported data to strategy properties
- 29. Strategy - Arbitration • Filter components apply a filter condition to the outputs of the source components • Prioritization components rank the components that connect to it based on the value of a strategy property, or a combination of strategy properties. These components can be used to determine the service/product offer predicted to have the highest level of interest, or profit
- 30. Strategy - Selection • Switch components express component selection through the Switch tab. Add as many rows as alternative paths for the decision as necessary, use the Select drop down to select the component, and enter the selection criteria as an expression in the If field. The component selected through the Otherwise drop down is always selected when the condition expressed in the If field is not met. • Champion challenger components express component selection through the Champion Challenger tab. Add as many rows as alternative paths for the decision as necessary, and define the percentage of cases for each decision path. All alternative decision paths need to add up to 100%
- 31. Interaction • Interaction rules define parameters for running a strategy, how to prepare the interaction history and how to save the interaction results • Interaction rules are used in flows through the Run Strategy and Capture Response shapes • On the “Interaction History” tab we define how the customer is identified. This information is important when determining Next Best Actions, including applying contact rules such as not offering the same proposition twice within a certain period of time or when the customer has already accepted or declined the proposition
- 32. Interaction • On the “Run Strategy” tab we reference the decision strategy, in this example the “NextBestOffer” strategy. • There must be at least one strategy but there can also be multiple public components. Public components are defined in the Decision Strategy.
- 33. Interaction • On the “Capture Response” tab we configure the information we want to capture from each interaction. • First of all we define the behavior and response. • We define the segment and potential sub segments the customer can be assigned to. • In the “Customer Response” section we define the proposition the response refers to. In this example it is the “SelectedProposition”.