This presentation shows an overview of use-cases in analytic models in telco. Analytic models are characterized by the fact that they produce a (predictive) score/feature for every customer/cell-tower/handset/call as opposed to a BI reports, where aggregated information is displayed.
- 1 Marketing
- 1.1 Acquisition model
- 1.2 Cross-and upsell models
- 1.3 Retention models
- 1.4 360 degree profile
- 1.5 Call center notes
- 1.6 Social data
- 1.7 Fraud models
- 1.8 Next Best Action components
- 1.9 Customer Lifetime Value
- 1.10 Competitor campaigns
- 1.11 Recharge model
- 1.12 Quality
- 1.13 Capacity
- 1.14 Social network
- 1.15 Known numbers classification
- 1.16 Known locations classification
- 1.17 Roaming analytics
- 1.18 Prepaid identification
- 2 Call center
- 3 Monetization
- 4 Financial
An acquisition model selects a group of customers with a higher likelihood of becoming a become customer from a larger pool of prospects. The larger pool can consist of ex-customers, customers from another company (e.g. a bank) or can come from acquired data such as yellow pages (B2B).
Cross-and upsell models
Cross-and upsell models take existing customers and asses the likelihood to migrate a customer to the more premium service (upsell) or to a additional carried product (cross-sell). Depending on the breath of the company, cross-sell can be a mobile data product on top of voice service or can be a broadband internet connection additional to mobile services.
Product recommendation is the product part of Next Best Action (see NBA). For every customer, a prediction is made for every product (or product group) indicating the likelihood of purchase, based on similarity of customers who already purchased that product. Typically, a top 3 product predictions are made available in the (inbound) call center and on a website. Scores are generated in a nightly batch run and there is no need for real-time model scoring.
Product growth is typically seen from the perspective of one product. Using a product migration approach, the combination of products of a customers is central, and it is visualized how customers migrate from the one pattern to the other pattern. Those transition probabilities can be used for marketing purposes to enhance or prevent certain migration patterns or they can be used in What-If simulation scenarios.
A special mention is given to the product migration from prepaid to postpaid, as this is typically perceived as a very profitable move: a customer ties up into a contract for a specified time. The migration model selects a group of customers from a pool of prepaid customers that are likely to respond to the postpaid offer. Note that, on the side of such a model, a credit check model has to be used in order to prevent inflow of customers with a high probability to default.
A retention model predicts which customer is likely to leave. In combination with customer value, it then can be decided what the effort needs to be to keep/retain the customer. Typically, a retention offer is made that gives a customer a discount to renew a contract.
Product churn model
A predictive model that assess the likelihood that a customer is leaving a certain product or product group. This is seen as an early warning for customer churn.
Customer churn model
A predictive model that assess the likelihood that a customer is leaving the company. Postpaid models with a contract end date have as main predictor 'time to contract end'. Prepaid models are more challenging due to the fact that churn is never observed (i.e., does the prepaid customer switches sim or does he just not call) and the fact that little information is known about the prepaid customer.
Rotational churn model
In the prepaid market, many customers rotate sim cards in order to call as cheap as possible. A rotational churn model asses such users and hence, it can be prevented to have those customers included in the regular retention campaigns. Instead, rotational churn campaigns incentivize customers to stay on the network for longer periods of time.
360 degree profile
The 360 degree profile typically stands for all information that can be captured about a customer and this made available for marketing purposes. As customer demographics are (likely) already available, a mention is made about the newer features that can be extracted.
Location features extract common locations of a customer such as home, work, shopping, vacation locations. Those features can be used to make campaigns more relevant. Data is obtained via GPS location of attached cell-towers.
Mobility features extract patterns such as commutes and other common travel patterns. In addition, it assesses mobility as the number of miles a customer travels. Campaigns are made more relevant using this information. Data is obtained via GPS location of attached cell-towers.
Temporal features include the detection of wake and sleep patterns, high activity patterns and detection of social time. Those features can be used to make campaigns more relevant. Data is obtained directly from Call Detail Records (CDR).
From a wide range of sources, customer activity gives rise to unbridled amount of information. The interest classification model brings this back to a number of categories that can be dealt with by the marketing organization and can be used to make messaging more relevant.
Call center notes
Call center notes come as unstructured data. In some cases, speech to text creates written notes out of recorded messages.
Customer sentiment extraction
Customer sentiment extraction assesses mood at the start of a call and follows the evolution of the mood during the call. In real-time situations, a supervisor intervention can be made if a customer doesn't seem satisfied with the provided solution. In offline mode, the mood features can be used in the personality type classification which, in turn, can be used to make messaging more relevant.
Product quality model
Extracted from the call center notes, common problems with products or services can be detected. Pre-set categories typically do not cover the possible range of issues.
Agent satisfaction model
Customer sentiment towards the end of the call can be used to assess if the call center agent satisfactorily solved the customer issue. Aggregated over a period of time, on agent level, this is KPI that is of more use than average time to solve.
Call-center call likelihood
Assessment of the likelihood of a customer calling the helpdesk. In understanding the characteristics of those customers, alternative (cheaper) methods may be found to achieve the same, or specific groups of customers can be given a different service level (e.g. people without internet access)
Social data comes from the social media channels and in wider sense, from the web (blogs, forums, etc.). Matching social media profiles to internal customers is typically challenging. Options are 1) match on aggregate level, 2) create apps that customer be-friend and gain access to their identifying details such as phone number, 3) via social media partnerships
Privacy laws prevent Deep Packet Inspection (DPI) for marketing purposes. The URL endpoints, however, can be used. The URL endpoint classification feeds into the customer interest profile model as it maps URL endpoints to interests. (E.g., https://vegetarian_recipes.com maps to the category 'foodie').
Social profile extraction
The collection of information available from the social media profile and the extraction of information of the posts made on the social media. Extractions can be sentiments, personality assignment and keyword based to detect interests.
Fraud in telco industry always comes out, as at the end of the day, the customer or the company is harmed. As clear as this may look, traditionally, fraud data is typically not stored in a way that analytics can easily handle. As such, fraud models are human created rule sets rather than probability models.
Theft detection models detect sudden unusual trends in customer behavior (E.g. sudden long distance calls, paid numbers, etc.) and alerts a helpdesk to verify that the phone is still with the rightful owner.
Fair use assessment
When packages are sold with 'unlimited use', the fair use assessment can analyze the usage patterns and assess or alertif/when individual users start using a significant part of the total usage.
A large class of models, traditionally mostly rule based, to assess specific cases of non-intent to pay, by-pass fraud, identity theft, sim card copy, etc.
A class of models, typically built into the switches and routers to detect penetration attempts.
Next Best Action components
Next Best Action models are an attempt to integrate many different models into one generic 'advisor'. The model is characterized by doing everything 'right': the right product, the right message, the right time, the right place, the right etc. As many of the models components have been discussed in the other topics, the additional few that are typically part of the NBA approached are highlighted here.
Channel preference model
The channel preference model assesses the channels that the customer is willing to spend money over. The preferred channel can be different for different product classes.
Contact frequency preference
The contact frequency model relates customer satisfaction and churn to the contact frequency in order to find out the preferred level of communication of a customer with the company.
Generic campaign response
The generic campaign response model estimated the likelihood of a customer giving a positive response to any campaign.
An overload model assesses the contact pressure where a customer switches off (I.e., not open for the message anymore), or gets annoyed.
Customer Lifetime Value
Customer lifetime models range from simple revenue-minus-cost models to advanced mathematical approaches to model future lifetime. A practical approach is a 4 step procedure: 1) assess revenues 2) assess costs 3) assess lifetime 4) use historic data where last year’s CLV is related to this year’s CLV in order to predict next year’s CLV.
Where the revenue model is relatively straight forward, the cost model is more complex. Typical approach is to include Sales, Acquisition and Retention (SAC) costs. Costs for specific services can be assessed by a top-down approach.
The churn model can be used as a lifetime estimation, or specific time-to-churn models can be build, however, this is typically more complex due to the nature of survival type models.
The competitor campaign model uses (local) pricing schemes from the competitor to 1) run the offers against the user base to understand potential price advantages for the customers and 2) predict the effect of those campaigns to the churn and usage rate. Data from the competitor can be scraped from web or can be obtained from local resellers.
The recharge model is specifically for the prepaid market. It predicts when customers will recharge and how much. This information can be used prior to the recharge, to propose a higher package and post-predicted recharge, a retention offer with a discount. Often it is of concern that the customer will find out the scheme and will postpone the recharge in order to get the discount. In practice this effect has not been observed, and if required, can be prevented by slightly randomizing offer time. Network =
Quality mostly is captured as aggregated information in dashboard. In some cases, it is interesting and actionable to capture this on customer level.
Dropped calls detection
Dropped calls may be available as an output of network statistics, yet, there's a number of features on customer level that can be used to understand customer perceived quality, such a consecutive re-connects. Those features are mostly used as predictors for churn, although apologetic acknowledgement of call quality is also known to be effective.
Handset quality analysis
An aggregation of the dropped call features on the level of handset may reveal patterns of certain handsets that do not well with the network. Churn prevention activity follows this analysis.
Cell tower quality analysis
An aggregation of the dropped call features on the level of cell tower or network node may reveal fault patterns of cell towers and network nodes that can trigger preventative actions. Although seemingly more appropriate for a BI dashboard, the cell towers and network nodes are so numerous that a real time detection algorithm for concurrent quality would be appropriate as an alert for network services.
Capacity monitoring is typically done on dashboards, some analyses having more analytics in them and hence is discussed here.
Demand forecasting in location and time predicts the expected network load and allows making resources available to guarantee sufficient throughput at minimum costs.
Black hole identification
Geographical model to understand cell coverage and assessment of potential earnings for cell tower placement given a location. The external data are population density and demographics in order to understand potential customer base.
Analysis of occurrences of congestion, being done as ad-hoc analysis is more flexible than dashboards. Analysis will identify overloaded network node in time and place and makes predictions about the next occurrence.
Social network focusses around connectivity, where as social data concerns the profile and posting information.
Assessment of customer call circle in connective density, concentration and spread. The actual top call circle can be used in campaigns and the change of top call circle can be used to assess life changes.
Leaders are customers with a lot of connections to other people. Authoritive leaders have people reach out to them, disseminative leaders reach out to others. Understanding leadership and the types creates opportunities for differential treatment in campaigns.
Peer campaigns model
Peer model campaigns revolve around incentivizing customers selling service to their (connective) peers. Prediction revolves around the selection of the right peer and the right target.
Assessment of connected peers being on the same network or not.
Known numbers classification
Known number classification revolves around collecting calls to banks, notaries, insurances, food ordering, etc. This data can be used to understand customer interest and life events and as such can be used to for effective messaging. Alternatively, this data, appropriately packaged, can be monetized by selling to external parties.
Known locations classification
Known location classification revolves around collection of locations that a customer visits such as airports, stadiums and countries abroad. As with the number classification, this data can be used to understand customer interest and life events and as such can be used to for effective messaging. Alternatively, this data, appropriately packaged, can be monetized by selling to external parties.
Both inbound and outbound travelers are, revenue-wise, very interesting customers. Inbound revolves around getting temporary customers onto the network, either roaming or with a local sim. Free Wi-Fi hotspots at (air)ports can be used to understand share of wallet. Outbound travelers can be predicted using location features mentioned earlier and can be offered interesting roaming packages in order to them keep them from switching to a competitor network or a local sim.
The prepaid identification revolves around a series of models with the purpose of uniquely identifying a customer. Rotational churn is easy to detect if a customer keeps the same sim card. A phone has a unique IMEI number, and as such, one can detect customers over sim cards, provided that they use the same phone. Customers keep having a unique call circle, and as such, can be identified by their phone circle over the change of phones.
Call center hiring model
Call centers typically deal with a large agent turn-over. This is a large costs, as training a new agent requires education, supervision and touches customer experience. A personality/aptitude/traits test as part of the interview can be related to employment six months later. The resulting predictive model allows to score to be hired candidates with the likelihood of employment six month later and hence will inform the decision to deploy or not.
Call center demand forecasting uses analytical models to predict upcoming call volume so that the appropriate number of agents can be planned for.
Case assignment model
The case assignment model connects the right agent to a case. In the simplest sense, an assessment is made of the case complexity and the more experienced agents are assigned the more complex cases. In a more evolved model, the topic of the case is routed to the agent who is most knowledgably in that area.
Email routing model
The email router is a text classification engine that scans incoming textual information and routes it to the relevant instance in order to reduce human effort. A simple model scans for address changes, complaints, questions etc. A more evolved model is capable capturing the details of, say, the new address, automatically.
With models getting more complex and different models getting combined, the ability appears of having helpdesk/webdesk calls being handled by cognitive entities. Those complex models are capable of providing the relevant answers as response on a customer spoken or written question (in natural language). This is an area in development at the moment.
The monetization of information is largely unexplored and is a sensitive topic. If data is being sold outbound (as opposed to inbound, see next topic), the data needs to be aggregated in such a way that individual customers are not recognizable anymore. Typical levels of aggregation are on postal code, or on (local) demographic segments. Information that external parties are interested in, are spending habits, size of the wallet, interests, messaging preferences and special locations.
Inbound monetization refers to giving external parties access to individual customers, however, safeguarded by the telco, ensuring that the customer identifying characteristics are never revealed. As such an external customer can request a message to be sent to, say, all males between 20-30 years, etc.
Linking telco data to
The monetization of data becomes more interesting when companies embark on a strategic partnership. The area is largely unexplored and is a sensitive topic, however, as with all new technology, it can be used for the better or the worse. Three use-cases are discussed, however, this is the very top of the iceberg. Retail Loyalty card data groups customer purchases over time, over braches, and potentially over brands. Subscription to a loyalty program may require government ID to be part of the provided information, in which case, the linkage to telco data (which also often requires some part of government ID) becomes easy. Numerous methods exist that do (ad-hoc) entity resolution to match telco to retail data in case no official ID is available. Retail data reveals a great sense of purpose of customers (E.g., diet - or and specifically, dieting, smoking behavior, alcohol consumption, purchase of specific activity articles as swimwear, diving goggles, etc.) From the telco perspective, this information can be used to better understand the customer and tune the messaging (E.g., knowing that someone is heading to a vacation) and from the retail perspective, the telco information can be used to make messaging relevant in time and place (E.g., a mobile advertisement will better work when it is received during a train ride than when it is received during top activity at work).
Banking & insurance data reveal the size of customer wallet, and the distribution of the wallet over services and products, the need for certainty and details as family composition. As such, combining this with telco data provides a view for the telco where services can be offered in accordance to creditworthiness, and sized to fit for family composition, life stages and outlook. For the banks and insurances, the telco data provides clues to make messaging relevant in time and place.
Media data reveals (experienced) identity of a person. I.e., the news-mindedness of a person, genres of movies, the type of TV shows one watches, etc. In addition, it shows great detail on how people spend time. If the telco isn't capable of finding out when someone wakes up in the morning, the power switch of the TV certainly will. Telco data can be used by media industry to connect advertisements over channels and over locations to more efficiently (less wasteful) reach the customer. The telco benefits from the media data by additional classification of interest, location, and temporal customer behavior in order to personalize messaging.
Telco's prevent loss by predicting if customers will be able to pay the bills. Traditionally an area of man-made rules, predictive models can be used to validate the rules, fine-tune the rules, or even replace the rules by a (preferably real-time)scoring model
Not limited to telco, there's a wide class of models available to enhance the financial forecasting. Currently, often, controllers use linear trends models or simple level corrected historic trend models, to predict future result, yet, more advanced models are available that account for non-linearity and confounding external influences.
Not limited to telco, there's a wide class of model available to enhance the risk assessment. Currently, often, controllers use a limited set of What-If scenarios to assess potential outcomes and their risk. Yet, the more advanced models are simulation based, take the chances of events occurring into account (distributional assumptions) and compute risk in quantitated amounts.
Not limited to telco, there's a wide class of model available to enhance the financial auditing. Currently, often, auditors use a fixed scheme sampling approach to check financial details. The use of anomaly models can enhance the auditing by automatically detecting situations that are deviant from the norm, controlled for a large number of other factors.