How Efficient Are Handovers in Mobile Networks? A Data-Driven Approach

Abstract

This work analyzes handover (HO) efficiency in mobile networks using real-world data, addressing key challenges that affect connection stability, latency, and network load. Unsuccessful HOs—often caused by suboptimal parameter settings, network congestion, or rapid user movement—are identified as a major problem. To address these issues, the study evaluates key radio frequency indicators such as Received Signal Strength Indicator (RSSI), Reference Signal Received Quality (RSRQ), Signal to Interference plus Noise Ratio (SINR), and the HO ratio, along with mobility-related factors including user direction and the ping-pong effect. A set of key performance indicators (KPIs) is used to quantify performance: Handover Rate (HOR), Handover Ping-Pong (HOPP), and Unnecessary Handover (UHO).The evaluation, based on real-world network data, shows an average HOR of 79.9%, an HOPP rate of 72.45%, and 14.83% of HOs classified as unnecessary. These findings reveal the limitations of traditional static threshold-based strategies and emphasize the need for adaptive, data-driven optimization approaches.The results demonstrate that a comprehensive HO strategy integrating multiple real-time parameters is essential for efficient mobility management and improving overall network reliability and user experience. This study reinforces the importance of leveraging real operational data to refine and validate mobility algorithms in modern cellular systems.

1. Introduction

The HO process is a key mechanism in mobile networks, essential to ensuring communication continuity as a user moves between coverage areas [1,2]. This mechanism allows an active connection, such as a call or data session, to be transferred from one base station (BS) to another without noticeable interruptions [3,4]. Its primary objective is to prevent connection drops or quality losses while the user is in motion, automatically connecting the device to a new BS when leaving the coverage area of the previous one [5]. This process is particularly relevant in high-mobility contexts, such as in vehicles or dense urban areas, to ensure seamless communication [6].

In heterogeneous networks (HetNets), where various types of cells with different characteristics coexist, efficient HO is crucial to maintain connection stability and optimize load distribution across available cells [2,5]. With the advent of technologies such as 5G, which involve a higher density of cells and higher frequencies, the HO process becomes even more relevant [3,7]. In this context, it is critical to ensure low latency and high quality of service (QoS), maintaining a smooth user experience even in high traffic demand scenarios [8].

However, in real-world scenarios, this process is not always executed efficiently, which can lead to disconnections and service interruptions. A critical issue is HO blocking, which occurs when the destination cell does not have sufficient resources to accept the transferred connection, causing communication failures [5]. Another challenge is traffic overload, where inefficient HO management increases latency and decreases QoS by congesting certain cells, while others remain underutilized [5]. Furthermore, inefficient HO decision-making can lead to the ping-pong effect, where a device repeatedly switches between two cells unnecessarily, deteriorating the stability of the connection and increasing the load on the network [3,5]. Latency in the HO process is another determining factor, especially for time-sensitive applications like video conferencing or online gaming [5]. HO failures, where the transfer is not completed correctly, also affect real-time services and reduce overall network reliability [7,8].

Several theoretical approaches have been proposed to optimize the HO process, including methods based on queueing theory, mathematical modeling, static threshold policies and more recently, machine learning (ML) and deep reinforcement learning (DRL) [3,9]. These methods show promising results in simulated environments; however, their application in real commercial networks is limited due to lack of adaptability to changing conditions and high dependency on training data. Furthermore, most evaluations rely on synthetic datasets that do not fully capture network behavior under real traffic and mobility conditions [5].

Given this gap, real-world data analysis is gaining traction as a more reliable foundation for network optimization. In this context, data-driven evaluations provide insights that are better aligned with commercial network dynamics, especially in heterogeneous and high-mobility environments [10,11]. Using field-collected data enables better understanding of HO behavior, revealing inefficiencies and areas for improvement without relying on assumptions from simulated environments.

This work proposes a descriptive, data-driven evaluation of HO efficiency in cellular networks, using real-world measurements collected over a commercial mobile network. The analysis focuses on identifying inefficiencies such as unnecessary HO, ping-pong events and delays, by correlating radio quality indicators (RSSI, RSRQ, SINR) with user mobility patterns. The study uses KPIs such as HOR, HOPP and UHO to evaluate the HO process. Unlike ML-based models, which often require extensive datasets and computational resources, our approach remains descriptive and interpretable, facilitating deployment in operational settings. While no ML models are implemented in this paper, we discuss how the proposed data structure and metrics could be used as input features for ML models in future work. The objective of this study is not only to characterize current HO performance, but also to establish a solid basis for future research that integrates ML-based decision-making or artificial intelligence (AI)-driven policy control in 5G and beyond networks.

2. Related Work

Research on HO in cellular networks has focused primarily on analytical models and simulations with the aim of optimizing the seamless transfer of connections. Traditional approaches have been based on simple rules that utilize individual parameters, such as signal strength or quality, without integrating multiple metrics simultaneously. These rules are applied using predefined thresholds, evaluating the optimal moment to perform an HO based solely on one parameter at a time.

One of the most common approaches involves the use of queueing models, which allow traffic overflow to be modeled and HO failures to be predicted based on simulations of different network configurations [5]. However, these methods are based on simulated data and often do not take into account real mobility conditions, limiting their applicability.

Other studies have addressed HO from a simpler perspective, proposing the use of isolated parameters such as the distance to the BS or the strength of the signal [1]. In these cases, other factors such as quality variation over time or interference from neighboring networks are ignored, which could lead to less precise HO decisions [3]. This type of simple rule has proven to be effective for low-mobility scenarios or homogeneous networks, where the condition variability is minimal.

In HetNets, recent research has applied simulation-based static thresholds to control HO between different cells. The exclusive use of simulated data imposes limitations, as dynamic factors such as network congestion or the real mobility of users are not considered. In many cases, only fixed signal quality thresholds (RSRQ) have been used to initiate HOs, which could result in a high number of UHOs or HO failures under changing conditions [2].

Although simulations have been useful for initial model development, evaluating HO efficiency in real-world environments is key to improving its applicability in commercial networks. The conditions of mobility, interference and network congestion vary significantly in real-world scenarios, making solutions based solely on simulations insufficient [10,11].

Several studies have addressed this problem using real-world datasets from operational networks. For example, in [10], real-world measurements of an operational Long Term Evolution (LTE) network in Asia were used to demonstrate how the use of ML can reduce latency and improve target cell selection. Similarly, in another work, an HO prediction model was validated using a real-world cellular network dataset, covering rural, suburban and highway environments, achieving an accuracy greater than 90% [11].

Other authors have explored AI-based approaches to mobility optimization. For example, in [12], the authors applied reinforcement learning to support dynamic HO decisions in 5G HetNets with unbalanced traffic distributions, improving network load balancing. In [3], a Deep Learning (DL) model was trained to predict signal behavior and anticipate HO failures, demonstrating gains in QoS. However, many of these studies still rely on simulation rather than operational deployment.

Complementing these approaches, [13] conducted a statistical analysis of HO performance in a cellular network in Quito, Ecuador, using data collected through drive tests and the Net Monitor tool (Huawei, Shenzhen, China; accessed: 1 July 2025). The study focused on a national mobile operator and high-traffic urban routes. Using R (version 4.3.1, R Foundation for Statistical Computing, Vienna, Austria) for data processing, the authors analyzed metrics such as RSSI, RSRQ, power margins and measurement repetitions. They observed an HOPP rate close to 10% and found that almost 50% of HOs did not target the geographically closest BS, which may affect overall efficiency. In addition, they proposed new features derived from signal measurements, such as power margins and repetition patterns, which could be valuable for predictive or data-driven optimization models [14].

Recent studies have advanced the use of deep and reinforcement learning techniques for handover optimization in 5G networks. Verma et al. [15] proposed a Double Deep Reinforcement Learning (DDRL) framework that significantly reduces UHOs, link failures and ping-pong effects. Unlike prior works relying on simulated data, their model is trained in a 3rd Generation Partnership Project (3GPP)-compliant real-time environment, enhancing its practical relevance. Similarly, the work by Prado et al. [16] introduced a Multi-Agent Deep Q Network (MADQN) approach that jointly optimizes user-to-BS assignments and handover timing, achieving better fairness, fewer handovers and reduced radio link failures while staying close to optimal performance. In addition, Amaira et al. [17] employed the Proximal Policy Optimization (PPO) algorithm to manage handovers in sliced 5G networks. Their model considers both BS and slice selection, and outperforms traditional Deep Q-Network (DQN) and Maximum Signal-to-Noise Ratio (Max-SNR) methods in simulation by reducing handovers and improving cumulative system performance. These works complement our proposal by demonstrating that rule-based feature extraction, as presented in our methodology, can serve as a practical foundation for training such DRL models under real-world conditions.

These studies demonstrate that analyzing real-world data improves the accuracy of HO prediction and optimization models and provides valuable insights into their performance in commercial networks. Validation in real-world environments is critical for developing more robust solutions that can be applied to future networks.

3. Proposed Methodology for HO Efficiency Analysis

To evaluate HO efficiency in cellular networks using real-world data, a structured methodology is proposed. This methodology includes the definition of successful and failed HOs, along with the analysis of key parameters such as signal strength and quality, user movement direction and the frequency of UHOs. Furthermore, the evaluation is performed using HO-specific metrics, such as HOR, HOPP and HO Detection, which allow for identifying failures in the HO process and optimizing their management in real-world scenarios.

3.1. Definition of Successful and Failed HO

A successful HO occurs when the user is transferred to a new cell with improved signal quality, ensuring stable connectivity and reducing the possibility of connection drops. This means that, after HO, RSSI, RSRQ and RSSNR (Received Signal Strength to Noise Ratio) should improve or at least remain within acceptable ranges [5,9].

However, a failed HO occurs when the HO does not improve connection quality, causes a disconnection or introduces unnecessary jitter between cells (ping-pong effect). These events can be caused by errors in selecting the destination cell, network congestion or incorrect setting of threshold values [3].

3.2. Data Collection

This study is based on a detailed analysis of a dataset collected using the NetMonitor mobile application (Vitaly V, Android app, available on Google Play Store, accessed on 15 July 2025) in the city of Quito, Ecuador, with the aim of investigating the relationship between network conditions and the generation of HO events in cellular networks [18]. Data collection was conducted over a 30-day period, covering various urban contexts and mobility conditions. A total of 29,670 valid records were obtained, distributed across 24 variables with no missing values, including both quantitative and qualitative variables.

The measurements were taken in four representative areas of the city, considering routes under high mobility conditions with varying speeds and directions, data collected from static locations, routes exhibiting population and urban variability, and an area characterized by low cellular coverage, ideal for assessing HO performance under adverse conditions.

The measurements were captured at a frequency of one sample per second. This provided high temporal resolution, suitable for analyzing sensitive events such as HOs. The data collection conditions included both high-speed routes and urban traffic conditions, as well as different times of the day, also covering periods of high network demand.

The measurements collected are grouped into several categories:

Power measurements: connection power (rssi) and strongest power reported (rssi_strongest).

Time measurements: day (date) and hour (time) of connection.

Position measurements: latitude (lat) and longitude (long) of the measurement location.

Signal quality measurements: quality of the received signal (rsrq) and signal to noise ratio (rssnr).

Connection measurements: operator code (net_op_code), network type (net_type), data transmission state (data_state), bytes received (data_rx), bytes transmitted (data_tx), number of UMTS network neighbors (umts_neighbors), number of LTE network neighbors (lte_neighbors) and the ID of the connected BS (psc_pci).

3.3. Experimental Setup

The analysis was based on data collected over 30 days in various urban environments of Quito, Ecuador, using Android smartphones (Samsung Galaxy A71, Samsung Electronics, Suwon, South Korea; Xiaomi Mi 10T, Xiaomi Corporation, Beijing, China) with the NetMonitor application. Devices were operated under diverse mobility conditions (walking, driving and public transportation), and at different times of the day, typically between 06:00 and 22:00, capturing both peak and off-peak periods.

Data was collected across a wide range of urban zones, including residential, commercial and mixed-use areas, enabling the evaluation of HO behavior under different network load and radio propagation conditions. The estimated user speeds varied depending on the mode of mobility: walking (3–5 km/h), public transport (10–25 km/h) and private vehicles (30–60 km/h).

The sampling rate was one measurement per second, chosen to balance the temporal resolution required to capture HO events and the battery consumption of the devices. Although battery impact was not specifically analyzed, this rate proved feasible for extended collection routes.

Environmental conditions such as interference or weather were not explicitly controlled, as the aim was to capture realistic operational conditions. This approach introduced natural variability into the dataset, enriching the analysis of HO efficiency in commercial mobile networks.

3.4. Data Transformation and Cleaning

The dataset was imported into Python (v3.10.13, https://www.python.org/, accessed on 15 July 2025) for analysis and creation of new features. Based on the initial data, additional variables were calculated to enable a more detailed analysis of HO behavior, such as:

Distance between locations: Calculated using latitude and longitude coordinates to determine geographic proximity between measurement points.

Time between locations: The data were derived from the differences in the date and time of the measurements, allowing temporal analysis of movement.

User equipment speed: Estimated by combining distance and time between positions to gauge the speed of movement of the device.

Direction of movement: Calculated as an angle relative to the north, with 0 degrees pointing east, increasing counterclockwise, providing a clear indication of the direction of movement.

Signal rating: Developed to evaluate the quality of the received signal, this composite rating was based on variables with the highest statistical variability.

Position clustering: Since many locations were recorded per second, generating multiple entries for the same location, the k-means clustering technique was applied to group positions by latitude and longitude. Each cluster was represented by its mean, simplifying location analysis.

Previous signal strength and quality: Variables were added to capture signal strength (rssi_ant) and quality (rsrq_ant) beforehand, allowing for analysis of how these preconditions impact the process.

Difference in signal strength and quality: Differences in signal strength (dif_potencia) and quality (dif_calidad) between consecutive measurements were calculated to provide an indicator of connectivity changes during HOs.

Furthermore, data were resampled to ensure a uniform temporal distribution, improving the representativeness of HO events. Resampling, essential in time series analysis, handles irregular data and facilitates the comparison of measurements taken at different times.

The resampling process shows that after application, the separation time between measurements increased, particularly in cases where a significant day change occurred. This confirms that the interval between measurements was correctly adjusted to 1 s.

3.5. Data-Driven Analysis

This study adopts a fully data-driven approach based on real-world measurements rather than simulated data. Using 29,670 records collected across diverse urban contexts in Quito, Ecuador, the analysis incorporates variables such as RSSI, RSRQ, SINR, user speed, movement direction and PSC_PCI changes to detect and evaluate HO events.

HO events were classified using deterministic rules: an HO was labeled as successful if post-HO signal strength and quality improved or remained stable, and as failed or unnecessary if the signal deteriorated, the user returned to the previous cell within one minute or no significant improvement in signal quality was observed. These criteria enabled the calculation of KPIs including HO Rate (HOR), Ping-Pong Rate (HOPP) and Unnecessary HO Rate (UHO).

Although ML techniques are not applied in this study, they are proposed as a future direction. The current work lays the foundation for ML-based HO optimization by defining and extracting features that can be used in supervised training and classification models in future implementations.

3.6. Handover Detection

The HO, which is the process by which a user device switches its connection from one BS to another, was detected primarily through changes in the PSC_PCI value (Physical Cell ID). The exact moments when a change in PSC_PCI occurred were identified, allowing the recording of the precise time when each HO occurred.

A new column called “Handover” was added, which was assigned the value “1” when the current PSC_PCI differed from the previous one, thus allowing measurements in which an HO occurred to be identified. Using this procedure, a total of 646 HO events were detected and recorded during the study period.

A detailed analysis of the number of HOs was performed, including the creation of graphs that depict the frequency and distribution of HOs over time. This analysis enabled the observation of patterns in user mobility and network connectivity management. Figure 1 illustrates a 400 s segment of the dataset where HO processes are marked with a value of 1, while periods without HOs are shown with a value of 0, represented by blue lines and points. Additionally, BS changes are visualized using the normalized PSC_PCI value, depicted by green lines. In the measurement area, the PSC_PCI values ranged from 3 to 500; to enable their visualization alongside the binary HO indicators, these values were normalized to a 0–1 scale. This normalization facilitates the observation of PSC_PCI changes—i.e., BS changes—within the same graph.

In addition to changes in PSC_PCI, other potential HO indicators were examined, such as variations in signal quality (RSRQ, RSSNR), signal strength (RSSI, RSSI_strongest) and user device speed. This multifactorial approach not only validated the detected HO events, but also allowed for a deeper analysis of the conditions that triggered them.

Figure 2 further illustrates how the radiofrequency parameters RSSI, RSRQ, RSSNR and RSSI STRONGEST are related to the execution of the transfer. It can be seen that during an HO process (represented by the blue line and point), these parameters normalized and visualized through shaded areas (red for Reference Signal Received Power-RSRP, green for RSRQ, orange for RSSNR and yellow for RSSI STRONGEST) undergo variation. Ideally, as shown in the figure, these values decrease before the HO and improve afterward. However, one of the objectives of this study is to identify situations in which this improvement does not occur.

3.7. Handover Events

3.7.1. Power Differences

The strength of the received signal is a key parameter in HO processes, as it directly influences the decision to switch from one BS to another. In LTE/LTE-A networks, RSSI is used to determine the appropriate moment to initiate an HO, particularly during events such as A2 and A3, where the signal strength between the current cell and neighboring cells is compared [19]. Furthermore, signal strength plays a crucial role in preventing UHOs and reducing failures during the process, highlighting the importance of hysteresis and time-to-trigger (TTT) [20].

For an HO to be considered successful, the signal strength after the HO must exceed that of the pre-HO state. In other words, the power difference should be positive, ensuring an improvement in connection quality.

Figure 3 displays the variation in RSSI around HO events. Blue bars represent the difference in signal power between consecutive measurements, where negative values indicate a loss in signal strength. Red dots highlight HO events, and those associated with a drop in power are considered failed HOs.

The analysis shows that most HOs occur without significant power degradation, suggesting a generally smooth transition process with adequate signal conditions before and after the HO.

3.7.2. Signal Quality Difference

In cellular networks, signal quality during the HO process is essential to ensure a smooth and uninterrupted user experience. Signal quality is primarily measured using the RSRQ parameter, which evaluates both the stability and strength of the received signal. An effective HO must ensure that this quality is maintained or even improved during the transition between cells. Minimizing signal loss during HOs is critical to avoid noticeable disruptions in the user’s connection. Signal degradation can lead to service failures, affecting call quality, internet browsing and other critical services [5,20].

To assess the effectiveness of an HO, the difference in RSRQ values before and after the event is analyzed. A positive difference indicates that the signal quality was maintained or improved, confirming a successful HO. This evaluation is key to ensuring a continuous and satisfactory user experience during transitions between BS.

However, it is important to clarify that a decrease in RSRQ during the HO does not necessarily imply a drop in connection quality or a failed event, as the RSRQ may still remain within an acceptable range. Therefore, it is more useful to identify cases where, after the HO, the RSRQ value falls below the minimum acceptable thresholds. This condition enables the detection of events in which the quality of the connection is genuinely compromised. This analysis is addressed in more detail in Section 5 (3GPP Thresholds).

Figure 4 illustrates the variation in signal quality (measured by RSRQ) during HO events.

The vertical bars represent the difference in RSRQ before and after each HO, where positive values indicate improved quality and negative values represent degradation. Red dots mark the occurrence of HOs. Events showing a post-HO drop in RSRQ are considered failed, as they may compromise connection stability and user experience. This visualization highlights the importance of signal quality as a criterion for evaluating HO efficiency, particularly in real-world network conditions where poor RSRQ values can lead to service interruptions.

3.7.3. Handover Ratio

The number of HOs performed during network operation is a key indicator to evaluate the performance of the HO process. The total number of HOs experienced by a user or a specific cell provides valuable insight into network efficiency and user experience. Counting HOs is essential for identifying mobility patterns and assessing the network’s capability to manage multiple HOs between cells without degrading service quality. A high number of HOs may indicate coverage or planning issues, while a low number suggests a smooth transition between cells.

To compare the number of HOs with the number of measurements evaluated in each interval, a ratio is calculated. This ratio is obtained by dividing the number of identified HOs by the total number of measurements in the corresponding interval. This calculation allows for the evaluation of the HO frequency concerning the volume of data recorded in each time interval.

Figure 5 shows the number of HO events detected in sliding time intervals of 5, 10, 15 and 20 min, based on a subset of the first 2000 records. Each colored line represents the HO count calculated within a sliding time window corresponding to the specified interval size. The X-axis displays these windows in units equivalent to multiples of 5 min, enabling the comparison of HO evolution at different temporal observation scales. It can be observed that as the interval duration increases, more HOs are identified, as expected; however, there is also greater stability in the curves, suggesting less variability in the events.

Figure 6 presents the HO Ratio (HOR), calculated as the number of HOs divided by the total number of samples, evaluated over four time intervals: 5, 10, 15 and 30 min. This ratio allows the assessment of HO activity independently of the sample size, highlighting how frequently HOs occur within each time window. As expected, the 5 min interval shows the highest ratios due to its sensitivity to short-term fluctuations. This behavior suggests that shorter intervals may overrepresent transient or insignificant HO events. Based on the analysis, a threshold of 0.06 is proposed to filter out overly frequent HOs that are likely unnecessary or noise, thereby improving the accuracy of KPI calculations such as UHO and HOPP.

3.7.4. Trajectory Change

In the HO process within cellular networks, the user’s direction of movement must be considered to optimize both service quality and network efficiency. The direction in which the user moves can significantly influence the selection of the target cell and the overall execution of the HO. If not adequately considered, a suboptimal target cell can be selected, increasing the likelihood of additional HOs and negatively affecting service quality. By integrating the direction of movement into HO algorithms, user displacement can be better anticipated, allowing smoother transitions between cells, which improve the user experience [21].

Considering the direction of movement in HO management allows for smoother and more efficient transitions. Accounting for this variable helps reduce service interruptions and provides a more stable and continuous connection, thereby contributing to a more satisfying user experience.

Figure 7 shows the temporal relationship between HO events and sharp changes in the user’s movement direction. Red markers indicate detected HOs, while green markers represent directional shifts greater than 170°, which are interpreted as significant turns. These abrupt changes in movement may lead to erroneous or unnecessary HOs—especially in scenarios where the user doubles back or reverses direction shortly after an HO. Identifying these patterns is essential for understanding mobility-induced HO inefficiencies and refining decision algorithms to reduce ping-pong effects.

3.7.5. 3GPP Thresholds

The 3GPP standards define the events that trigger HO between BS in cellular networks. These events help avoid premature decisions when selecting a BS during the HO process, ensuring better QoS by maintaining a stable and efficient connection. In addition to providing a standardized framework for managing HO, the 3GPP standards also enable the configuration of key thresholds that help decide when an HO is necessary [22].

These thresholds include the SNR, RSSI and RSRQ. These parameters are critical for assessing signal quality and determining the need for HO [23]. Adjusting these thresholds according to the type of service and network conditions is essential to ensure that the user maintains a high-quality connection for as long as possible. Proper threshold configuration enhances HO efficiency, reducing the likelihood of interruptions and ensuring a seamless user experience.

The typical thresholds defined are as follows.

RSSI: A value above −80 dBm is considered excellent, while a value below −110 dBm indicates weak signal strength [6].

RSRQ: Values above −10 dB indicate good quality, while values below −20 dB are deemed inadequate [6].

SINR: An SINR greater than 20 dB is excellent for data transmission, while values less than 0 dB indicate a noisy signal that affects quality [6].

Analyzing whether identified HOs meet these thresholds is crucial to determining whether they were successful. This evaluation is key to ensure users experience a smooth transition between cells without a significant degradation in service quality.

Figure 8 presents a subset of real measurements showing the behavior of key radio parameters—RSSI, RSRQ and SINR—during HO events, in relation to 3GPP-defined threshold values. Green dotted lines represent acceptable operating ranges, while red lines indicate minimum acceptable thresholds for each metric. HO events are marked with red dots across the timeline. This visualization enables assessment of whether an HO was triggered under adequate signal conditions. Notably, HOs occurring below the critical threshold lines (e.g., RSRQ < −20 dB or SINR < 0 dB) are likely to degrade the connection and can be classified as suboptimal or failed. This analysis is essential for identifying threshold misconfigurations and improving HO decision policies.

3.7.6. Repeated Base Stations

The analysis of repeated BSs is related to the “ping-pong” effect in cellular networks, which occurs when a mobile device repeatedly switches between two adjacent cells. This behavior can lead to frequent disconnections and poor QoS [24]. By examining the number of times a device connects to the same BSs within a given time interval, repetitive HO events can be identified. Reducing these events improves HO management and improves network stability.

Figure 9 shows histograms of repeated BSs within different time intervals (0.1, 0.3, 0.5 and 0.7 min). Very short intervals (e.g., 0.1 min) detect too many repetitions, risking false positives, while longer intervals (e.g., 0.7 min) may miss actual ping-pong events. The 0.5 min interval provides a balanced detection, capturing meaningful repeated HOs without overgeneralization.

In Figure 10, yellow boxes highlight repeated connections to the same BS within short time intervals (approximately 0.5 min), a phenomenon known as the ping-pong effect. This indicates UHOs that reduce network efficiency. The vertical axis shows the BS identifier (PSC/PCI), allowing visual tracking of cell transitions. By filtering out sequences where the same BS is revisited in rapid succession, the analysis can better capture meaningful HOs and reduce noise in mobility patterns.

3.8. Handover KPIs

KPIs related to HOs are essential for evaluating the efficiency of algorithms in mobile networks. These metrics allow for measuring and optimizing the performance of communication systems. Among the most relevant KPIs are the following:

Handover Rate (HOR): This KPI reflects the proportion of successful HOs relative to the total number of attempts. A high HOR generally indicates that mobility management procedures are working effectively, especially in dense urban areas where frequent HOs are expected due to small cell coverage and continuous user movement. It is calculated using the following formula:

$$HOR={\displaystyle \frac{\mathrm{Number}\mathrm{of}\mathrm{successful}\mathrm{HOs}}{\mathrm{Total}\mathrm{number}\mathrm{of}\mathrm{HOs}}}$$

(1)

For its calculation, HO events were extracted from the dataset, and conditions were applied to validate their relevance. Although HOR provides an overall view of HO performance, it does not reflect whether each event was truly beneficial. Some HOs, although technically successful, may not improve service quality—for example, when the user equipment (UE) quickly returns to the original cell. These cases, known as UHOs, are included in the total but indicate inefficiencies in the mobility algorithm.

Handover Ping Pong (HOPP): This indicator measures repetitive HOs between two or more cells within short time intervals, a phenomenon known as ping pong. A high HOPP value indicates connection instability and suggests possible misconfigurations or inefficiencies in the HO decision process [5].

The HOPP calculation starts by chronologically sorting the dataset and identifying HOs that involve cell changes (PSC_PCI). A ping-pong event is detected when the user equipment returns to a previously connected cell within a short time period.

The HOPP rate is calculated as follows:

$$HOPP\_Rate={\displaystyle \frac{\mathrm{Number}\mathrm{of}\mathrm{HOPPs}}{\mathrm{Total}\mathrm{number}\mathrm{of}\mathrm{HOs}}}$$

(2)

This KPI enables the identification of instability patterns in mobility management and supports the adjustment of HO parameters to reduce unnecessary cell switching.

Unnecessary Handovers (UHO): These are HO events that do not result in signal quality improvement and may negatively affect the user experience by increasing signaling load, causing service interruptions or unnecessarily draining the device’s battery. Unlike HOR, which measures the frequency or success of HOs, this metric focuses on the efficiency and necessity of each mobility decision.

UHOs are identified by analyzing sequences of events occurring within short time intervals. An event is classified as unnecessary when the user equipment quickly returns to the original cell, indicating that the initial cell was still more appropriate or that the HO decision did not properly account for the user’s movement direction. These events rarely improve signal quality metrics such as RSSI, RSRQ or SINR and often reflect premature or inefficient HO decisions.

For detection, the dataset was first chronologically ordered, and only records with confirmed HOs were retained. Then, three conditions were evaluated: the time interval between consecutive HOs had to be short, the cell identifier (PSC_PCI) had to remain unchanged and no significant improvement in signal quality indicators (RSRP, RSRQ or SINR) should be observed. When all three conditions were simultaneously met, the event was classified as an UHO. This approach allows for the identification of redundant network transitions that do not benefit user connectivity and may indicate inefficiencies in the mobility algorithm.

The proportion of UHOs is expressed as:

$$UHO\_Rate={\displaystyle \frac{\mathrm{Number}\mathrm{of}\mathrm{UHOs}}{\mathrm{Total}\mathrm{number}\mathrm{of}\mathrm{HOs}}}$$

(3)

This metric helps quantify the effectiveness of mobility decisions and provides valuable insights for network parameter optimization.

To support the choice of the one-minute threshold used in the HOPP and UHO KPIs, an empirical analysis was conducted considering user speed and direction changes. The resulting heatmap (Figure 11) shows that most HOs occur at speeds greater than 5 m/s with low angular variation—conditions typical of stable forward movement. In contrast, HOs occurring at low speed and with minimal direction changes often happen within intervals shorter than 60 s, suggesting ping-pong behavior. Therefore, a 60 s time window was deemed an appropriate threshold to identify suspicious events without excluding legitimate transitions.

4. Results

To optimize the HO process in cellular networks, it is essential that events that do not contribute to improving signal quality be eliminated. This implies that only those HOs that have been successful and meet the parameters established in this research will be retained.

Table 1. Handover analysis by scenario.

Event	Condition	Total	Without HO	HOs	Invalid HOs	% of HOs	% Invalid HOs
Initial Data	-	13,617	12,971	646	-	-	-
Resampling	-	15,388	14,675	713	-	-	-
Power Difference (dp)	HO = 1, dp < 0	15,201	-	526	187	3.46	35.55
Quality Difference (dc)	HO = 1, dc < 0	14,812	-	576	137	3.89	23.78
Handover Ratio (rh)	$t=5$ min, rh > 0	14,910	-	478	235	3.21	49.16
Trajectory Change (td)	HO = 1, cd > 170	14,695	-	693	20	4.72	2.89
3GPP Thresholds	HO = 1, RSSI < −110 dB	14,836	-	552	161	3.72	29.17
	HO = 1, RSRQ < −20 dB	14,675	-	713	0	4.86	0.00
	HO = 1, SINR < 0 dB	14,816	-	572	141	3.86	24.65
Repeated BSs (rbs)	$t=0.5$ min, rbs > 1	14,938	-	450	263	3.01	58.44

details the number of HOs eliminated under each of the conditions analyzed previously. This breakdown allows for effective data purification, ensuring that only successful HOs remain in the dataset. In this way, facilitating future comparison with results obtained through optimization algorithms, such as those based on ML, which are considered for future work.

It is important to note that the data related to eliminated HOs are not completely discarded from the dataset. Instead, an HO value of zero is assigned to these events, while the psc_pci value is maintained equal to that of the previous event. This approach allows for the preservation of data integrity and continuity, which is essential for subsequent analysis, while discarding HOs that failed.

Looking at the table, it can be seen that the HO ratio condition and the number of repeated BSs result in a higher number of eliminated HO events compared to the condition based on the RSRQ threshold, which eliminates no events. Additionally, the direction change condition results in the elimination of 20 HOs. These results demonstrate the effectiveness of the proposed conditions in reducing UHOs, highlighting the importance of considering both the HO ratio and the number of repeated BSs to obtain a dataset that reflects successful HOs. In contrast, the RSRQ threshold is shown to be less effective in filtering HOs that have not been successful, suggesting that caution should be exercised when evaluating the quality of HO events.

The analysis of the main KPIs associated with HO and their impact on the number of UHOs is presented in

Table 2. Handover KPI analysis by scenario.

Event	HOR (%)	HOPP (%)	UHO (%)
Power Difference (dp)	74.00	62.55	18.06
Quality Difference (dc)	81.00	66.84	15.80
Handovers Ratio (rh)	67.00	83.26	1.46
Trajectory Change (td)	97.00	78.79	9.96
RSSI	77.00	71.92	13.04
RSRQ	100.00	79.81	9.53
SINR	80.00	72.03	13.11
Repeated BSs (rbs)	63.00	64.44	15.11

. All indicators are expressed as percentages, allowing a normalized comparison across different conditions. The table compares the HOR, the proportion of HOPP and the percentage of UHO. This structure enables a clear identification of the factors that enhance HO stability and those that contribute most significantly to redundant or inefficient events.

Figure 12 shows the significant variations in HO KPIs in the different stages of the analysis, now represented as percentages for improved comparability. Regarding the HOR, the highest value was obtained under the RSRQ threshold (100%), indicating that all HOs triggered by this condition were successful. However, the repeated BS condition presented the lowest HOR (63%), reflecting a higher proportion of HOs that did not meet the established success criteria.

In terms of the HOPP rate, considerable differences were recorded between the evaluated conditions. The RSRQ threshold again displayed the highest value (79.81%), suggesting that while this configuration improves the success rate of HOs, it can also result in frequent back and forth transitions between cells. In contrast, the HO ratio (rh) condition showed the highest overall HOPP rate (83. 26%, reinforcing that overly frequent HOs can compromise connection stability.

With respect to UHO, the configurations that exhibited the highest percentages were those based on power difference (18.06%) and quality difference (15.80%). These results indicate that although these criteria aim to optimize signal strength or quality, they often generate HOs that do not translate into meaningful improvements in connection conditions.

In contrast, the RSRQ threshold, despite leading to a high number of ping-pong events, demonstrated high efficiency in terms of successful HOs and lower UHOs (9.53%). This highlights the importance of balancing different evaluation metrics when tuning HO decision parameters.

The HO evaluation using real-world data yields an overall average of 80% HOR, approximately 70% HOPP and under 15% UHO across the studied conditions. These values reflect the typical behavior of the HO process under realistic operating scenarios and underscore the importance of multi-criteria analysis for effective mobility management.

While this study focused on rule-based filtering and statistical thresholds for HO evaluation, it sets the groundwork for more advanced strategies. Future iterations could leverage ML models to dynamically learn optimal thresholds or recognize patterns across multiple parameters. However, in this exploratory stage, the emphasis was on interpretable, deterministic criteria derived directly from real-world measurements, ensuring clarity in the impact of each condition on KPI behavior.

5. Discussion

Correct HO evaluation in cellular networks requires considering both signal quality and user dynamics. Indicators such as RSSI, RSRQ and RSSNR are critical to determining whether cell HO improves connection quality. A successful HO occurs when these values exceed established thresholds, ensuring stability of communication. However, selecting these optimal thresholds is challenging, as poorly calibrated values can lead to unnecessary or failed HOs, affecting service continuity.

Another key factor in HO evaluation is the HOR rate, which measures the frequency of BS changes. Defining the analysis window and the appropriate threshold for this rate is essential to distinguish between necessary and excessive HOs. In heterogeneous networks, this parameter must be dynamically adjusted to balance connection quality and stability.

Furthermore, user dynamics significantly influences HO efficiency. Abrupt changes in direction of movement, such as U-turns, should not trigger UHOs, as the device could reconnect to the previous cell. This factor becomes even more relevant when considering user speed; at high speeds, rapid cell changes may be necessary to maintain QoS, while at low speeds, an early HO could generate instability.

HOPP is another key metric in the HO evaluation. This occurs when a device repeatedly switches between two BSs in a short period of time, indicating poor threshold settings or lack of coverage stability. Detecting these events should not be done in isolation, but in conjunction with other factors such as user speed and direction, to avoid erroneous decisions that introduce instability into the network.

As noted above, building a model based on real data to accurately assess successful HOs is a complex task due to the interaction of multiple parameters. The way these metrics are applied and combined directly influences the results. In this work, the conditions were analyzed independently; However, a more advanced model—potentially incorporating ML—could be developed in future work to integrate these parameters in an optimal sequence. This would allow for the construction of adaptive decision models capable of real-time threshold adjustment based on spatiotemporal context. Additionally, such learning-based systems can be designed to align with emerging paradigms such as 5G and 6G network slicing, where mobility decisions must be optimized per service slice. This approach would allow for a more precise and adaptable HO management system, improving the efficiency of the process in commercial cellular networks.

The proposed method, based on real-world KPIs and mobility-aware filtering, can be directly adapted into operator-side RRM (Radio Resource Management) policies. For instance, dynamic adjustment of thresholds such as HO rate or directionality conditions can be implemented in Self-Organizing Networks (SON). While not formalized into 3GPP protocol specifications, this data-driven logic aligns with principles of intelligent HO management and could be deployed as part of vendor- or operator-specific optimization layers.

6. Conclusions

Evaluating HO in cellular networks using real-world data allows a better understanding of the impact of this process on the QoS and operational efficiency of cellular networks. This study demonstrated that optimized HO management contributes to reducing connection instability. To mitigate this problem, it is essential to adjust HO parameters and thresholds, avoiding unnecessary switches between BSs and improving service continuity. For an accurate HO assessment, key indicators such as RSSI, RSRQ and RSSNR were identified, along with metrics such as HOR, HOPP and UHO. Furthermore, factors such as user speed, distance from BS and direction of movement play a key role in detecting unnecessary and failed HOs.

Continuous analysis of these parameters will allow HO management to adapt to dynamic network conditions, ensuring greater efficiency. It has been shown that the order in which the evaluation criteria are applied significantly influences the results. Identifying successful HOs does not depend solely on a single indicator but on an optimal combination of parameters. Adjusting decision algorithms based on these indicators improves system accuracy and reduces the failed HOR, optimizing the user experience.

Another critical aspect is the time interval used in the HO evaluation, as adjustments to this parameter can affect the detection of successful HOs. Incorporating additional factors, such as the speed of the user and the distance to the BS, is recommended to achieve a more adaptive and accurate decision-making model.

The use of real data in HO evaluation represents a significant advantage over simulation-based models, as it more accurately reflects network operating conditions. This facilitates the calibration of parameters and algorithms, allowing for more effective optimization of the HO process. In future research, the future integration of real-world data with ML algorithms presents a promising strategy to further improve HO management.

Designing a comprehensive HO management system that combines multiple conditions and parameters is essential not only to ensure efficient HO but also to optimize overall network performance. This approach allows for the development of more robust and dynamic solutions, ensuring that the HO process is increasingly accurate, stable and efficient.

Although the proposed evaluation framework showed strong results, it was applied to a single operator and city, which limits its generalizability. Future work will extend this methodology to other geographic and operational contexts, and explore integration with emerging technologies such as network slicing in 5G/6G. This study deliberately avoided applying ML models at this stage, focusing instead on interpretable, rule-based metrics derived from real measurements. However, the availability of such datasets opens the door to future integration of supervised or reinforcement learning models that can dynamically adapt to user behavior, mobility patterns and network conditions across heterogeneous environments. This capability is especially relevant in the context of managing mobility across multi-slice 5G and upcoming 6G architectures, where quality-of-service differentiation requires more intelligent and context-aware HO mechanisms.