Introduction to Optimisation in Collections...2. Core elements of optimisation...4. Objective (Goal Function)...4. Decisions...4. Data...5. Models...

Transcription

1

2

3 Table of contents Introduction to Optimisation in Collections...2 Core elements of optimisation...4 Objective (Goal Function)...4 Decisions...4 Data...5 Models...5 Constraints...5 Opportunities to optimise within this process...6 Pre collections strategies...6 Performance Improvement...7 Self cure strategies...7 Path optimisation (early/mid/late)...8 Debt Collection Agency (DCA) optimisation...10 Recoveries...11 Dialler optimisation...11 Summary...13 Appendix: Optimisation across the Collections Process...14 Pre collections strategies...14 Self cure strategies...14 Path setting...14 Debt Collection Agency (DCA) optimisation...15 Recoveries...15 Customer collections versus account level collections...15 Tools...15 About Experian s Decision Analytics Global Consulting Practice...16 About the author: Adrian Carr...16 An Experian white paper Page 1

4 Introduction to Optimisation in Collections Debt collection is no longer considered a cost centre it has become a front-line profit driver since financial organisations increasingly realised the huge impact its collections operations have on their financial results. But, the real key is to collect in a smarter fashion. Lenders need to know their customers better, segment and target their approach and automate as much as possible to maximize collections. Optimisation is now a well used term in everyday English. It is typically used to imply that something is being carried out in the best way possible. Inherent in the definition is that there are multiple ways of doing things, and the most effective has been identified. With regard to optimisation in collections, this paper focuses attention on the actions that can be carried out on individual customers, and how they can be chosen in such a way as to optimise overall performance in the collections environment. Back office or process management required in the day to day running of a collections operation are not for discussion here. An example of customer level decisions that can be optimised is: If a choice can be made between sending a letter or making a phone call to a customer There are many customers Only a certain number of calls can be made And it is desired to maximise debt collected Optimisation achieves an uplift in performance against traditional techniques by: Considering every single combination of actions Understanding the trade offs between the different actions (which are forced by constraints) Choosing the best set of actions to fit within the constraints Experian has observed KPI improvements in the region of 5-30% when applying optimisation techniques to the collections process. A simple example helps demonstrate this. The table below shows the expected debt collected and cost for three different customers, for three different actions: Action No Action Letter Phone Customer 1 Debt Collected Cost Customer 2 Debt Collected Cost Customer 3 Debt Collected Cost Page 2 Collections optimisation

5 Choosing the best action for each customer results in 510 debt collected, at 12 cost. If however the cost is constrained to be a maximum of 10, then the best combination is not obvious, until all combinations have been calculated, (it s 490 at cost of 9). Similarly, if one can only choose one of each action, then maximising performance is driven more by the differential (or trade off) between actions, rather than the individual customer response to the actions themselves. If one were to use the traditional hierarchical approach of: 1. Choose the best return from the best customer first 2. Then move to the next best action for the next best customer 3. Until a constraint is met This would result in 430 collected at a cost of 7 ( 300 is picked first, then 100, then 30), missing the optimised maximum of 470 if cost were constrained at 7. More mathematically, optimisation is the term used to describe the task of maximising a function, subject to constraints. For example, this could be maximising debt collected, subject to operational constraints such as call capacity, and/or operational cost. These are often described as non linear integer programming problems, and there are a number of tools that help solve these problems. This paper describes: The collections process Core elements of optimisation The opportunities to optimise decisions made at various points within this process Deployment of the solutions One key question that hasn t been answered explicitly is Why do I need to optimise, why can t I just use better models? The key reasons for this are: The combinations of different options The complexity of competing KPIs (e.g. bad debt/resource/cost/roll rates etc) Management of multiple constraints Management of multiple customer level dimensions (e.g. propensity to self cure, and propensity to Right-Party Contact (RPC), propensity to pay, arrears balance, exposure, etc.) This is why a structured approach with the backing of a suitable software tool is essential. An Experian white paper Page 3

6 Core elements of optimisation Optimisation can be used to improve performance for each of the Collection processes and using all of the tools presented in Appendix 1. The challenge is to deliver improved performance, in a simple and rapid fashion. Experian has completed a number of projects across the process, with clients and has identified demonstrable benefits. In each example, there are a number of core requirements to deliver an optimised solution. These are: Objective (Goal Function) What is it we are trying to optimise? Ultimately for most projects, this will be to minimise ultimate expected losses, net of cost, in other words to minimise net impairment. Equally however, there may be subtly different themes of this, that are however correlated e.g. to maximise cure rate, or minimise roll rate. Decisions These are the actions that can be carried out on a customer. It could be a path, or a time to call or a number of grace days, etc. The advice of walk before you run is well heeded in many walks of life, and it is no different with optimisation. When considering decisions, it is best to start optimising decisions that have been known to occur in the past, rather than optimising things that have never been tested before Page 4 Collections optimisation

7 Data Broadly there are three types of data Operational data This is data that is available to be used to make a decision. (E.g. data that exists within the workflow management tool), and exists at a customer level. Analytical data Data that may influence a decision (e.g. outcome data that helps drive a model), but cannot be used in an operational tool. This also exists at a customer level. Portfolio or Action Level data For example cost of a letter, average call duration, impairment ratios, etc. The operational and analytical data will include items such as: Demographic data Behavioural data Scores Bureau data Previous Action data This paper is not designed to assess the merits of these different pieces of data, but just to use them where they are predictive. Models Customer level estimates for particular KPIs that influence the goal function will add a great deal of discrimination, and hence performance improvement to an optimisation problem. These customer level estimates (e.g. propensity to self cure) can either be provided for by: Existing scorecards that are predictive (e.g. behaviour or collections score) Variables that are predictive or directly influence the KPI (e.g. balance) New scoring models (e.g. propensity to self cure, given action A). Note that these new models are typically response style models that are designed to highlight differences in outcomes based on different actions, rather than the traditionally labour intensive credit risk or collections scorecards. As such they could be uni, bi, or multi variate models, and a pay off between time versus benefit can often be estimated. Constraints The constraints are the things that simply must or must not happen. E.g. we must have a minimum of $x collected, or we must not have more than 20,000 outbound calls made this month. Constraints can occur at lower levels of detail too, e.g. time period of activity, (1,000 calls between 10 and 12), and at a descriptive level too (e.g. average debt per DCA must be the same). An Experian white paper Page 5

8 Opportunities to optimise within this process It s attractive to consider optimising across the entire end to end process for every single customer. However, this is an onerous task, both in terms of data requirements, and analytical resource. Equally the robustness of such a solution could be questioned. Experian have found that by taking discreet parts of the process, studying them and optimising them, significant benefits can be obtained at each stage typically 10 to 20 percent improvement in target KPIs. The following examples describe some of these discreet points. Pre collections strategies A recurring theme with all optimisation opportunities is that no two problems are alike, there are so many points at which variation can occur. This includes the obvious ones, such as the definition of a KPI like impairment, or the value of a call volume constraint, but also the shape of the problem itself, which is often determined by the actions. Consider a call barring decision in the Telco environment for example. A decision can be made to fully/partially/specifically bar a customer. This can be made for fraud/risk reasons based on: A percentage/absolute amount Over and above a set limit/previous usage And the customer can be called/sms sent/directed to collections on inbound calls As such flexibility and scalability of any tool used for optimisation is a key requirement. Experian developed a call barring strategy for a Telecommunication company, the key elements seen below: The trade off between risk (propensity to enter collections, driven by behaviour score) and balance (driven by spend) is the key to achieving the objectives and meeting constraints. Customer spend behaviour is particularly interesting in the Telecommunications environment. Different customers present different behaviour, but sometimes for the same causal factor. Consider a taxi driver having a long weekend break, versus a married couple with children taking a long weekend break from the children. The former will probably have a drop in spend, the latter an increase. Page 6 Collections optimisation

9 Thus when setting a limit on monthly spend, it is more important to understand the individual customer s variability of spend, rather than the average, or point estimate of spend. A limit can then be set to only be hit in x percent of cases. This is analogous to the observation that the standard deviation for an individual customer, is far greater than the standard deviation around the mean estimate of a customer. Models are built on thousands of customers, and as such they are predicting the average of those thousands of customers. The prediction for the group will be very accurate but one should not set a limit based on this accurate estimate of the mean, because the individuals within that group will vary, not just amongst the group, but also individually too. For example, consider a customer with a mean spend of $100, and a standard deviation of $20. The probability that their mean will be greater than $120 next year is 5,000 to 1. The temptation is to set a limit of $120, and only expect 1 in 5,000 customers to be barred. However, on average a customer will exceed this twice a year resulting in almost 1,000 bars! Performance Improvement There are a number of swap sets with an optimised approach versus a traditional champion strategy that sets limits by risk. A few examples help illustrate where benefit may be achieved: Low risk, low spend, low variability = low limit (versus previous high limit) Low risk, low spend, high variability = high limit High risk, high spend, low variability = med/high limit (versus previous low limit) A reduction in the volume of bars of over 20 percent can be achieved using optimisation, with no impact on volume or value of debt entering collections. Self cure strategies The key goal in self cure strategies is to avoid spending resource on customers that will pay their arrears, even if no action is carried out on them. As ever, nothing is certain, so there is a pay off between the propensity of a customer to self cure, versus the cost of contacting them, the template structure of this decision can be seen below: An Experian white paper Page 7

10 Modelling again forms a key part of the decisioning process, and consideration should be given to developing separate models for different days past due (or groups of days past due) e.g. Propensity to cure, given no action, given x DPD. This obviously has to be traded off against the benefit that will be achieved on customers that would not have self cured, but will make some payment if contacted. It is also crucial to ensure that the objective is correctly aligned. Maximising debt collected will simply result in customers likely to self cure being contacted. At the very minimum, an incremental debt collected (versus no action being carried out at all) is required. Interestingly in this situation, for the first challenger strategy it may be worth having a minimum number of outbound calls carried out. This ensures that the optimised solution does not leave call centre staff unused. Also it acts as a safety check against trusting the models implicitly. Once a new strategy has been demonstrated to improve upon the champion, and then consideration can be given to reducing staff, or employing those more effectively elsewhere. Path optimisation (early/mid/late) Path optimisation is the holy grail of collections optimisation. Identifying the right communication for each customer, executing it at the correct time, via the right channel, with the right terms is close to this holy grail.as long as it is then followed up with the correct set of communications, all working seamlessly together. The requirements of this perfect path (per customer) are onerous, as not just the outcome of each combination is required, but also the interaction of the actions must be understood which is hard to assess especially as: Relationships between actions can be very strong or very weak Actions are often close together and so cause and effect are hard to be sure of There is typically a lack of testing (high risk customers are rarely receive no action ) There are many possible outcomes and not clearly good/bad Sequences of actions can be long over 90 days A more manageable challenge to optimising paths is to simplify the problem to the point that it can be optimised. There are a few ways of doing this, including: Identify the best action for a customer, and move that action to the front of the path Optimising smaller discreet periods of activity (e.g. week by week) Treating the path as a single action The last of these methods was carried out with a UK Retail Bank, and the methodology described on the next page: Page 8 Collections optimisation

11 In this example, there were three potential paths that could be assigned to a customer. It s easy to think about them as high/medium/low risk paths. The champion approach was taken to assign the high risk path to high risk accounts, and so on. There was also a challenger strategy which tested some accounts of one risk, through a different path. This champion challenger test was useful in that it allowed customer level models to be designed for each path. Thus one could estimate the Expected Impairment for each customer, for each potential path that could be assigned to them. Equally it is possible to develop customer level models to estimate the number of actions of each type that will occur for a customer, if they pass down a particular path. So rather than being bogged down in the complexity of the movement of a customer through this path, evaluating which action is causing which reaction etc, this methodology simplifies the problem to its core components Which path to apply? What is the expected impact on impairment How much resource is required? The improvements (of up to 5% without creating new paths) generated by optimising in this way are coming from two key sources: The pay off between risk and balance E.g. high risk & low balance = low risk strategy, hence less cost The size of difference in outcome for individual customers for different paths E.g. small difference between high risk and low risk strategy = apply low risk strategy and hence low cost E.g. large difference between high risk and low risk strategy = apply high risk strategy to minimise impairment. Constraints can also be applied by the different channels, as there are estimates of volume of customers receiving each communication, based on the strategy applied. This can then allow for cost and resources to be effectively managed. It should be noted that deployment of optimised strategies should be no different from normal challenger strategies - implementing on a small test group is recommended to allow for data gathering, clear test results, and risk mitigation. An Experian white paper Page 9

12 Debt Collection Agency (DCA) optimisation Once the decision has been made to use an external agency to collect the debt, and when there are more than one to choose from, optimising between them becomes a very attractive, quick win option. The reason for this is that different DCAs perform differently on different types of customer, this can be driven by Geography Customer type Specialism in gone aways/trace etc, Actions they take out (phone versus door knocking) The other advantage of this decision point is that collections departments will often have a wealth of data available, as DCAs will have been tested against one another, with a random set of accounts being passed to each. This is perfect for developing propensity to collect models that are specific to each DCA. Again the core elements required to develop the optimised strategy are outlined below: Clearly an internal agency can be included in the optimisation problem to minimise cost, as can second placements of debt, and even debt sale. The results of such a problem can be very rewarding, with up to 10 percent improvements observed. Equally satisfying is that optimisation delivers a win win: Each DCA gets debt that is more aligned to their collections ability, so they collect more, and earn more commission The collections department sees an increase in total debt collected, improving the core KPI In this example, consideration can also be given to batch processing DCA files, as the visibility that a strategy tree delivers is not necessarily needed by either the collections team, or the DCAs. As such a further improvement in performance can be gained by using individual (customer level) optimisation tools. Page 10 Collections optimisation

13 Recoveries Recoveries is similar to DCA optimisation, in that there is typically just one action, and a simple 0/1 outcome. It differs in that there is limited challenger testing between strategies, so data is limited, and because of smaller volumes, complexity is lower, and constraints fewer. However, this is not to say that it cannot be optimised. However, most uplift in performance will be driven by taking the most profitable action, if there is one. The elements of a recoveries optimisation opportunity are outlined below: Dialler optimisation The examples given above all lend themselves to being implemented in a decision agent, or a workflow management tool that can score and segment customers. However, optimisation can also be applied to dialler activity. Again, this is not optimisation of the process of the dialler, but optimisation of which customers get called and when. Below is a schematic of the dialler pyramid outlining how the ultimate goal of obtaining payment from a customer is driven by a number of outcomes that need to occur: An Experian white paper Page 11

14 Fortunately there is typically a lot of data available to model each of these events (although not every individual outcome needs to be modelled, as there will be correlation between outcomes, and thus one or two key drivers can be modelled (e.g. Propensity to get RPC, and Propensity to Get Successful Payment) The approaches to using diallers vary significantly between collections departments, emphasising once more the need for flexibility in a software tool. Outlined again below are the key requirements for taking an optimisation approach into the dialler environment. Significant benefit has been observed in the optimisation of dialler activity, with the improvement coming from understanding the differential between different actions, and selecting customers appropriately, rather than just choosing the best action per customer. Consider for example two customers, with the following. Customer A Propensity to collect (a.m.) = 70 percent Propensity to collect (p.m.) = 50 percent Customer B Propensity to collect (a.m.) = 40 percent Propensity to collect (p.m.) = 10 percent A hierarchical approach (with constraints on call volumes) will pick a morning call for customer A, and an afternoon one for customer B. However, if one swaps these around, the result is a higher total response. When one considers the impact of balance, RPC rates...it is clear that significant uplift is achievable by measuring each accurately, and optimising them. Similar to the DCA optimisation, individual optimisation can be more readily considered for this, and where modelled previously, uplifts in performance can be up to 35 percent. Page 12 Collections optimisation

15 Summary Optimisation is a mathematical methodology that has been demonstrated to add value to many decisioning problems, across many industries. The collections process is no different in that there are many different actions, can be taken on many different customers, and clearly there are opportunities to improve performance by taking an optimisation approach. The key challenges to achieving the benefit are to: Distilling the complexity of the problems into manageable discreet opportunities Creating customer level drivers that predict the truly desired outcome Having a sufficiently flexible tool to manage the multiple drivers, KPIs, and constraints. If this is achieved then optimisation can add benefit across collections: Pre collections Self cure strategy Early/Mid/Late Paths DCA allocation Recoveries By developing off line strategy trees, these can be deployed in various decision agents and workflow management tools, resulting in limited IT impact, and no IT integration. Equally a similar approach can be used to optimise dialler activity. The benefits are significant at each stage with five to 30 percent improvements possible, and correspondingly significant impacts on KPI performance, and ultimately improved profit for an organisation. An Experian white paper Page 13

16 Appendix: Optimisation across the Collections Process The above diagram broadly describes the various stages in the collections process; these are points at which individual optimised decisions can be made, namely: Pre collections strategies These are actions that occur, prior to a customer becoming delinquent. Examples of pre collections strategies can be considered to include: Limit management Authorisations strategies Preventative calls Triage Reaction to inbound calls Pre bill reminder messages Self cure strategies The idea of a self cure strategy is to provide a grace period during which a customer will return to order, or cure, without any actions occurring. Often a grace period will be as long as a full month or cycle, though will also included setting a number of days for a customer during which no collections action will occur. Page 14 Collections optimisation Path setting Collections Path describes a sequential group of actions that are carried out. Other terminology includes collections module, or contact schedule for example. Typically a customer will remain on the path for the duration of the path (e.g. 30/60/90 days), unless they return to order, or are assigned to another path or hold queue (e.g. payment plan) So for example, a collections path could be: SMS at 5 days past due (DPD) Letter at 15 DPD Outbound Phone Call at 25 DPD

17 Path setting is probably the most complex of points within the collections process to consider, and it is worth segmenting ones thinking about it in three stages: First action optimisation i.e. choose the preferred (optimal) contact methodology for each customer Path segmentation when there are x paths to choose from, one can consider choosing the right path for each customer (and one can either consider using only existing paths for greater simplicity) Optimal path development this is most complex, and aims to create a path that modifies both the actions and the timings of these actions (significant data and modelling required for this) Debt Collection Agency (DCA) optimisation A debt collection agency is a specialist collections organisation that is (usually) an external function from the organisation that the debt is owed to. Typically a creditor will firstly decide that a customer needs to be passed to a DCA for resource issues, of because the DCA will have greater chance of success of collection. A panel of four or five DCAs is typically used to ensure good performance from the DCAs through competitive tendering, and contingency in terms of volume. Recoveries The final step in the collections process is typical the Recoveries step. This usually involves more intensive or escalated activity, such as legal activity, going to court, and ultimately debt sale. Customer collections versus account level collections Customer is frequently referred to in this document, and it is a recurring theme of many debates in the collections industry. In this document it is used in its loosest sense of a person upon which a debt is to be collected. More analytically, it means a row in a decisioning problem, as such each of the problems described here can be equally applied at either the customer or account level. Tools There are a number of tools that assist credit risk and collections departments in their ability to collect well. These include: Decision agents. Typically a strategic tool that is used to score and segment customers, and apply different treatments to them (e.g. Probe Strategy Manager can be used to assign customers to collections paths). Workflow Management Tools. These tools can also score and segment customers, but typically are designed to work in a real time environment so they can update actions based on activity that is occurring on the account (e.g. if a customer makes a payment in the morning, then the planned dialler call is not made in the afternoon). Predictive Diallers. A tool that is used to efficiently make use of operational calling resource, when a list of customers is required to be called. An Experian white paper Page 15

18 About Experian s Decision Analytics Global Consulting Practice Experian s business consultants deliver powerful insight that is used by clients to enhance credit-management strategies across their consumer and small-business portfolios and the Customer Life Cycle. Experian s Global Consulting Practice is a credentialed consultancy dedicated to creating measurable and sustainable value for organisations around the globe in financial services, banking, mortgage, automotive finance, telecommunications and utilities, microfinance, retail credit and debt collections. We specialise in analytics-based decision strategies, data-driven products and services, regulatory compliance and fraud risk management across acquisitions, customer management, collections and overall portfolio management. Experian s business consultants provide clients with exceptional credit and fraud riskmanagement strategic insight, detailed enhancement opportunities, and deployment strategies through deep business subject-matter expertise and client intimacy, as well as a client engagement methodology to ensure consistency. We have deep knowledge of data, analytics and software and have demonstrated the ability to synthesise this intelligence with the deep understanding of credit-management principles and practices to solve our clients complex business needs. Average of 20 years experience per consultant Forty-three consultants based throughout Asia Pacific, EMEAI, NA and the UK More than 700 engagements across the globe. Teams with local knowledge of best practices About the author: Adrian Carr, Senior Business Consultant, Global Consulting Practice Adrian Carr has over fifteen years experience working across the credit and marketing functions for a number of blue chip organisations. Carr has held management positions in risk modelling, risk management, marketing analysis and product marketing roles for organisations including LloydsTSB, Orange, HSBC and Marks and Spencer Money. With an honours degree in pure mathematics, and a Pg Cert in Operations Research, Carrs academic background helps drive his analytical and creative approach to problem solving. Previous projects include contact planning for a credit card issuer; credit authorisations strategy for a retailer; and a contact strategy for a multichannel, multiproduct PFS provider. Page 16 Collections optimisation

19

20 Landmark House Experian Way NG2 Business Park Nottingham NG80 1ZZ Experian 12/11. The word EXPERIAN and the graphical device are trade marks of Experian and/or its associated companies and may be registered in the EU, USA and other countries. The graphical device is a registered Community design in the EU. All rights reserved.