⚡ electrical engineering

Temporal portfolio stress-testing for candidate 5G upgrades under future operational burden

This paper proposes a temporal portfolio stress-testing framework that utilizes December traffic and spatial features to predict January operational burdens for 5G upgrades, demonstrating that while advanced constrained portfolios significantly improve spatial equity and engineering efficiency, they offer minimal gains in future burden capture compared to simple forecast-based rankings.

Original authors: runzhe liu, hui chen, yun he, dabin wang

Published 2026-06-29

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: runzhe liu, hui chen, yun he, dabin wang

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a mobile network operator as a massive city with thousands of streetlights (cell towers). Some lights are flickering, some are dim, and some are burning out because too many people are trying to use them at once. The company has a limited budget to upgrade these lights to a newer, brighter "5G" technology. They can't fix them all at once, so they need a list of the top 100 lights to fix first.

The problem is: How do you guess which lights will be the most crowded next month, not just this month?

This paper is like a "stress test" for that list-making process. Here is how the authors approached it, using simple analogies:

1. The "Crystal Ball" Problem

Usually, when companies make these lists, they look at the data from this month to decide what to fix this month. It's like looking at a traffic jam right now and saying, "I'll build a new road here today." But what if the traffic pattern changes tomorrow?

The authors decided to be stricter. They used data from December to make their "to-do" list, but they waited until January to see if their list was actually good. They wanted to know: Did the lights we picked in December actually turn out to be the busiest ones in January?

2. The "Sticky" vs. "Fickle" Signals

They looked at different clues to decide which towers were in trouble:

The "Sticky" Clue (Backflow): This is like a heavy truck that gets stuck in traffic. The data showed that if a tower was busy in December, it was almost guaranteed to be busy in January. This signal was very stable.
The "Fickle" Clue (High-4G/No-5G days): This is like a sudden, short-lived rainstorm. Just because it rained for a few hours in December didn't mean it would rain in January. The authors realized this clue was too unreliable and threw it out of their main list.

3. The "Perfect Score" vs. The "Fair Score"

The team tested different ways to build the list:

The "Score-Only" Approach: This method just picks the 100 towers with the highest predicted traffic. It's efficient, but it's like a greedy shopper who only buys the most expensive items in one specific neighborhood, ignoring the rest of the city.
The "Fair & Balanced" Approach: This method tries to pick towers that are busy, but also ensures that every single city in the province gets at least one tower on the list, and that the towers aren't all clustered right next to each other.

4. The Big Surprise: "Score Saturation"

Here is the most important finding, which the authors call "Score Saturation."

Imagine you are trying to catch water in buckets. The "Score-Only" approach catches a certain amount of water (future traffic burden). The "Fair & Balanced" approach catches almost exactly the same amount of water (within a tiny fraction of a percent).

The Catch: Even though they caught the same amount of "water," the buckets they used were completely different!

The "Score-Only" list picked 50% different towers than the "Fair & Balanced" list.
The "Fair & Balanced" list managed to represent all 16 cities in the province, whereas the "Score-Only" list only covered 9 cities.
The "Fair & Balanced" list spread the towers out more evenly, so they weren't all crowded in one spot.

The Lesson: Once you have a good enough prediction of the future, you don't need to sacrifice efficiency to be fair. You can change the list to be more geographically fair and less clustered without losing much performance.

5. The "Engineering" Twist

Finally, they added a "complexity" check. Some towers are simple (one light), while others are complex (a whole cluster of lights). Upgrading a complex tower is harder and might involve more paperwork or engineering work.

They tweaked their "Fair & Balanced" list to avoid the overly complex towers where possible.

Result: They reduced the number of complex "cell records" they had to deal with by about 10%, while still catching 98.6% of the future traffic burden. It was like finding a way to carry the same load with a lighter backpack.

Summary

The paper doesn't claim to have invented a new way to predict the future perfectly. Instead, it shows that you can build a "fairer" and more "spread-out" upgrade list without losing much efficiency.

Old Way: Pick the top 100 busiest spots. (Good for traffic, bad for fairness).
New Way: Pick the top 100 busiest spots, but force the list to include every city and spread them out. (Almost the same traffic benefit, but much better for the whole region).

The authors conclude that this "stress test" framework helps network operators make decisions that are not just mathematically efficient, but also geographically fair and easier for engineers to handle.

Technical Summary: Temporal Portfolio Stress-Testing for Candidate 5G Upgrades

Problem Statement
Mobile operators face the incremental challenge of selecting which existing macro sites to upgrade to 5G under fixed budget constraints. The decision requires balancing multiple, often conflicting objectives: capturing future operational burden (traffic, backflow, user reach), ensuring regional equity across jurisdictions, managing spatial redundancy, and accounting for engineering complexity. Existing prioritization methods often suffer from two critical flaws:

Temporal Confounding: Many studies construct rankings and evaluate them using the same time period (cross-sectional), measuring internal consistency rather than the ability to predict future operational states.
Preference Uncertainty: Single fixed-weight vectors in Multi-Criteria Decision Making (MCDM) conceal the uncertainty regarding which objectives (e.g., commercial traffic relief vs. public-service coverage) should be prioritized.

Furthermore, a high future burden at a site does not causally imply that an upgrade will yield proportional throughput or revenue gains; the goal of this study is strictly to identify candidates for engineering review based on future load signals, not to estimate post-upgrade causal outcomes.

Methodology
The authors propose a temporal portfolio stress-test framework using data from a provincial mobile network comprising 14,129 physical macro sites across 16 cities.

Data & Temporal Separation: The study utilizes two monthly snapshots. December data serves as the decision-time input (features), while January data serves as the future evaluation target (ground truth). This strict separation prevents data leakage and tests the transferability of decision rules.
Burden Definition: A composite operational burden metric is constructed using empirical percentiles of three December features: user reach, 5G backflow traffic, and average site traffic. High-4G/no-5G days were analyzed but excluded from the primary target due to low temporal persistence.
Forecasting: Future burden is modeled using machine learning (Elastic Net, Random Forest, Gradient Boosting) with city-grouped cross-validation (GroupKFold). This ensures models predict January outcomes for a city without having observed that city's January data during training.
Portfolio Construction:
- Baselines: Includes raw backflow ranking, equal-weight scoring, entropy-TOPSIS, and a forecast-only ranking.
- Robust Scoring: A preference-robust score is generated by sampling 1,000 weight vectors from a symmetric Dirichlet distribution, penalizing sites whose utility varies significantly across plausible preference scenarios.
- Sequential Selection: A greedy algorithm constructs the portfolio by maximizing a marginal value function that combines:
  1. Robust demand utility.
  2. Local redundancy penalty (exponential decay based on 3km distance).
  3. Regional shortfall bonus (ensuring representation across all 16 cities).
  4. Engineering complexity penalty (based on the count of distinct cell records per site).
Evaluation Metrics: Efficiency is measured by the share of total January composite burden captured. Equity is measured via the Jain index of city service fractions. Spatial redundancy is assessed via mean nearest-neighbor distance. Engineering complexity is audited via the total count of associated cell records.

Key Results

Temporal Persistence: 5G backflow showed high persistence between months (Spearman $\rho$ = 0.968), whereas high-4G/no-5G days were unstable ( $\rho$ = 0.252).
Forecasting Accuracy: The Gradient Boosting model achieved strong city-held-out performance ( $R^2$ = 0.847, Spearman $\rho$ = 0.911), confirming that December features effectively predict January burden.
Score Saturation: At a 100-site budget, the "forecast-only" ranking captured 1.349% of the total January burden. Constrained portfolios (spatial-equity and engineering-aware) captured 1.334% and 1.333% respectively. The difference in future-burden capture was negligible (<0.02 percentage points), and statistical tests (paired sign-flip) found no significant efficiency loss.
Equity and Diversity Gains: Despite similar efficiency, the constrained portfolios produced materially different engineering shortlists:
- City Representation: Forecast-only ranked sites in only 9 cities. The spatial-equity robust portfolio represented all 16 cities.
- Equity Index: The Jain index improved from 0.397 (forecast-only) to 0.967 (robust portfolio).
- Spatial Reach: Mean nearest-neighbor distance increased from 18.1 km to 28.7 km, reducing geographic clustering.
- Turnover: The robust portfolio replaced 50% of the sites selected by the forecast-only baseline.
Engineering Complexity: An engineering-aware extension (penalizing high cell counts) reduced the selected cell records from 217 to 193 while retaining 98.6% of the forecast-only capture and maintaining high equity (Jain = 0.946).

Significance and Claims
The paper claims a modest but operationally significant finding: score saturation. Once temporal prediction is strong, the marginal gain in future-burden capture from optimizing purely on forecast scores is minimal. Therefore, operators can impose constraints for regional equity, spatial diversity, and engineering complexity without measurably sacrificing the primary goal of capturing future operational burden.

The framework shifts the focus from finding a "perfect" optimizer to creating a transparent screening layer. It demonstrates that:

Temporal holdout is essential to distinguish descriptive indices from deployable decision rules.
Multi-objective constraints (equity, redundancy, complexity) can redirect the engineering shortlist to be more balanced and manageable, even if the raw "capture" metric remains flat.
The approach does not estimate causal upgrade benefits (throughput/revenue) but provides a robust method for selecting candidates for engineering review under future operational pressure.

The study concludes that this framework is suitable as a candidate-screening layer until richer data on costs, coverage, and post-upgrade outcomes become available for full network optimization.