Imagine you are a telecom operator with €27bn to invest. You have to choose between hundreds of thousands of possible options. Different network areas, different technologies and different cell tower locations. All with different costs, revenues and returns. The only catch: the expected returns for all individual options vary widely and are hard to know in advance.
In aggregate, that is the yearly challenge for all European telecom operators combined. It is a multi-variable optimization for which the outcomes have to be predicted. Traditional approaches fall short, especially when done at scale. As it turns out, AI can help solve the puzzle. Machine learning algorithms can sift through customer, usage, network and other data to predict which investment options are most valuable to pursue. The recent surge in data traffic and changes in demand patterns make this a critical capability for operators seeking to further enhance their networks once normal life resumes.
growth in data traffic
“Is the connection bad on your end or mine?”
Ongoing glitches in video conferences can have a detrimental effect on virtual meetings. Along with remote working, the normal traffic congestion of most people’s daily commute has also gone digital. Streaming-quality bandwidth used to be mostly an evening or weekend requirement for Netflix binge watching. But connectivity now belongs to the list basic necessities. People may find themselves in daily competition for WiFi with their kids who are streaming home schooling classrooms in the morning.
In pre-Corona times, the average global data traffic grew at a rate of about 26% per year according to Cisco. Specifically mobile data traffic is booming with 46% per year (global average). Current lock-down measures to curtail the spread of the COVID-19 virus are causing a seismic shift in internet usage. Vodafone reports a 50% growth in mobile traffic in some European countries . British Telecom claims that traffic on its fixed network climbed as much as 60% compared to normal weekdays. These are growth rates operators would normally expect to see in a year. One way to cope with this surge in demand is to ask major streaming services like Netflix to reduce bandwidth utilization at the expense of picture quality. But that is not a sustainable fix.
Like many other industries, telcos have their challenges in responding to COVID-19. Lower (retail) sales, postponed 5G roll outs, capacity scarcity in customer service, to name a few. In the long run, telcos will benefit from the increased need for reliable connectivity as a critical commodity. But delivering on this promise will put more emphasis on required network upgrades.
“Traditional approaches to telecom network investments leave money on the table, because they tend to focus primarily on the cost side.”
Telecom operators in Europe invest around €27bn per year on network upgrades . At the same time, providers are facing flat or declining revenues in most mature countries due to saturated markets and heavy competition. So how can telecom operators upgrade their networks by rolling out 5G and Fiber-to-the-Home, while also earning a decent return on their investments? Upgrades cannot all be done at once, so where should they invest first? Our societal shift in the use of digital tools may open up new revenue opportunities for telcos. But they already have opportunities to increase returns by becoming more savvy in making investment decisions.
Traditional approaches to telecom network investments leave money on the table, because they tend to focus primarily on the cost side. The most cost-efficient areas go first. This makes sense: in densely populated areas the investment per connection goes down drastically. However, costs are only half of the equation. Equally important is the customer lifetime value impact: how do network upgrades influence customer loyalty (and hence market share) and contract value? Where will demand grow the most? These factors have shown to differ substantially between areas and depend on factors such as demographics and current network speeds. For instance, areas with current DSL bandwidths 7-12 mb/s have shown a 4 times higher market share growth during the first 18 months after fiber
roll out compared to areas with 12-50 mb/s as current network speed .
Operators find it hard to predict these customer effects and thus make very rough assumptions when making investment decisions. Often resulting in lengthy discussions between Operations and Commercial departments which cannot agree on prioritization.
 Perimeter of European Telecom Network Operators (ETNO)
 Real example from a European country.
Devising an optimal network upgrade strategy consists of developing thousands of mini-business cases. Upgrade decisions are done on a local level, either being a geographic area for fiber roll out or individual cell towers for mobile. These upgrades are typically done over a multi-year period, so their sequential order is a major value driver. We want the most valuable business cases to go first. Moreover, upgrades can also involve multiple options: investing in new fiber-to-the-home, upgrading existing copper capacity, using community WiFi offloading or hybrid fixed-mobile solutions. The crux lies in upfront determination of the expected value creation of all these single investments and to then order them accordingly. This is where AI comes in.
To help telecom operators optimize their network investment decisions, we developed an AI powered solution – called Smart ROCE – that calculates one single metric for each individual investment option. This metric is called ROCE (Return on Capital Employed) and is determined with help of multiple machine learning algorithms that analyze the entire customer base and usage data, enriched with external data.
“One European operator achieved a 120% increase in expected returns on investments in fixed network upgrades”
Fixed Network Upgrades
For investments in the Fixed network, value creation is modeled using three algorithms. The first two algorithms predict for each location in a potential investment area the churn and acquisition probabilities. The third one predicts the average revenue per location by determining the likelihood of cross and up-selling. All three models depend on available network speed and predict over a 10 year period, taking evolving market needs for bandwidth into account. Moreover, they also account for differences in demographics such as household composition, income levels, age, etc.
The combination of these algorithms provides a prediction of market share and revenue evolution for a particular area, depending on the network infrastructure. The outcomes of these predictions are used to calculate the change in Customer Lifetime Value (CLV) for each individual household after a network upgrade, by subtracting the prediction for the case of not upgrading from the prediction for the case of upgrading. The ROCE can subsequently be defined as this value creation divided by the investment.
Smart ROCE was used by a European operator to prioritize areas in their fixed network for fiber roll-out or DSL upgrades. The Smart ROCE engine even allowed to propose hybrid scenarios: use copper upgrades as a 5 to 7 year bridging technology to avoid market share drops due to insufficient bandwidth, and do fiber roll out after that.
The figure below depicts the outcome of this optimization, yielding an overall estimated 120% growth in investment returns (NPV).
Mobile Network Upgrades
Optimizing mobile network upgrades is more challenging than it is for fixed network upgrades. For fixed networks, the following relations can be analyzed and predicted with relative ease:
- How upgrades improve the performance of the network. Achievable bandwidths are known, depending on the existing technology infrastructure. In some cases, specific performance metrics are known at location level, such as line quality. Network upgrades typically have clear effects on bandwidth improvements.
- How the network performance impacts the customer experience. The bandwidth is the dominant driving factor for the customer’s experience. So if we know from the infrastructure the expected (or measured) bandwidth at a customer location, we can derive a good estimate of experienced quality.
- How the customer experience influences customer behavior. We can analyze customer behavior (churn, acquisition, cross and up-sell) in relation to the experienced quality, and hence correlate it with network upgrades.
For mobile, these relationships are less straightforward. Firstly, mobile network performance can be expressed in many different metrics: dropped call rates, throughput, fall-back ratios (from 4G to 3G for instance), signal strength, and more. Secondly, which of these have the biggest influence on the subscriber’s experience? Thirdly, if a customer moves through a network, how to allocate the network performance of a site to a specific customer? Lastly, at what point does network performance start to have material influence on customer behavior? Should we focus on mobile sites close to subscriber homes, since the performance of those masts will be most important? Or should we analyze experienced performance during daily commutes, for instance causing structural connectivity issues in the train?
These were intriguing challenges to tackle and a full treatment is beyond the scope of this article. But let’s illustrate one of the findings here.
The modeling objective was to determine how mobile site upgrades influence customer behavior such as churn and how much (customer) value this would create. Our approach consisted of the following steps:
- Find network performance metrics that have a (causal) relationship with customer churn and use these as features to predict churn.
- Determine how specific site upgrades would improve network performance and experienced customer quality.
- Calculate the expected investment returns (ROCE) for all possible site upgrades and use that for prioritization in upgrade planning.
The first step involved analyzing and modeling a very skewed data set: the mobile network in this European country was of relatively high quality. To relate performance issues to customer churn was like looking for a needle in a haystack. To solve this puzzle, we selected a subset of the data for which we knew that “home coverage” was below a critical threshold. In other words, we kind of ‘oversampled’ bad network performance and then tried to predict churn.
Figure 2 shows one of the results: the predicted impact of increased 3G fall back on churn. This relationship suggests that when the fall-back percentage increases from around 20% to >40%, customer churn goes up by 1.5%-points.
Different algorithms were used, such as XGBoost and random forest. The first showed a very erratic partial dependence plot, pointing to overfit. A random forest model was chosen and validated on different out-of-sample data sets (without the ‘oversampling’). Stable performance (measured as the area under the Precision-Recall Curve) implied good generalization. That is, we could apply it to the haystack.
The final solution for Smart ROCE Mobile involved multiple “engine parts” that required sophisticated analysis, such as predicting the impact of network upgrades on mast performance and predicting future market demand. The final approach was used to prioritize sites for 5G upgrades.