Your Phone’s Location Codes Could Reveal Road Speeds Without Any Sensors

Every time a mobile phone crosses an invisible boundary drawn by a telecom operator, the network quietly logs the event. Those boundary crossings, recorded as sequences of Location Area Codes, are usually treated as administrative noise. A new study argues they are actually a free, city-wide traffic sensor. Researchers from the China Academy of Transportation Sciences and Beijing University of Technology have shown that by matching these code sequences against the roads they cover, travel speeds can be estimated from ordinary Call Detail Records, with no roadside detectors, no cameras, and no dependence on GPS-equipped probe vehicles.

The work, published in Frontiers in Future Transportation, tackles a stubborn problem in traffic monitoring. Conventional technologies such as inductive loop detectors, roadside cameras, and dedicated short-range communication systems deliver high accuracy but cover only a small fraction of the road network, and deploying them across entire cities is prohibitively expensive. Floating Car Data from GPS-equipped taxis extends coverage substantially, yet probe penetration in many cities remains low; in Beijing, the study notes, the taxi fleet represents roughly 2.5 percent of all vehicles. Call Detail Records, by contrast, are generated as a byproduct of routine network operations for nearly every subscriber, at marginal incremental cost, making them an attractive complement for monitoring the vast majority of uninstrumented road segments.

CDRs are not, however, a ready-made traffic dataset. Each record carries a timestamp, a base station Cell-ID, and a Location Area Code identifying the administrative grouping of the serving cell, but the positioning accuracy is limited by base station spacing, typically 200 to 2000 meters in urban areas and larger in suburbs. Worse, records are produced whenever a user triggers a network event, such as a call, a text message, a cell handover during an active call, or a location update during standby, so sampling follows human behavior rather than a fixed schedule. The researchers analyzed China Mobile data in Beijing and found that each user generates about 53.8 valid records per day, roughly 2.1 events per hour on average, but with a daytime inter-record interval of about 1800 seconds that stretches to around 4000 seconds at night. Conventional GPS-oriented map-matching algorithms simply cannot cope with trajectories this coarse and irregular.

Earlier mobile-phone-based methods have split into two camps, each with a fundamental weakness. Handover-based approaches exploit the transfer of an active call between base stations as the user moves, comparing real-time handover sequences against pre-surveyed reference patterns along target roads. Under controlled conditions this can approach GPS accuracy, but handovers occur only during active calls, and in typical urban environments fewer than 5 percent of users are on a call at any given moment, leaving samples extremely sparse during off-peak hours. Cell-sequence-based methods, meanwhile, exploit the sequence of base stations a device associates with during standby, which works for all subscribers, but cell coverage radii of several hundred meters to several kilometers mean a single road segment may intersect only one or two cells, producing sequences too short to uniquely identify the road, especially in low-density suburban networks.

The new framework inserts an intermediate matching layer between these two extremes: the Location Area Code. LACs are coarser than individual cells but more stable, aggregating multiple cells into administrative zones. Crucially, normal location updates are triggered whenever a standby user crosses an LAC boundary, so LAC transitions are recorded for essentially every mobile device, not just those on calls. The method works in two stages. First, for each target road, the researchers construct forward and reverse reference sequences of the LACs covering the road, ordered along the direction of travel. Second, raw communication records are organized into pre-matching sequences, with consecutive identical LAC records collapsed into a single representative record chosen near the middle of the time span, and these sequences are compared against the road references using a similarity index. Records whose time gaps exceed 60 minutes are excluded, and successful matches yield individual speed estimates based on event timestamps and the projected positions of base stations, which are then averaged across all matched users.

Field experiments revealed how strongly the cellular environment shapes the method. On the urban South Third Ring Road, about 95 percent of a moving phone’s connection time is spent with base stations within 300 meters of the roadway, whereas on the suburban Sixth Ring Road and Jingcheng Expressway the equivalent 95 percent thresholds extend to roughly 1600 and 1200 meters. LAC coverage follows the same gradient: a single LAC covers an average of only 1.86 kilometers of the Third Ring Road, but 10.83 kilometers of the Jingcheng Expressway. This means the same number of LAC transitions spans vastly different road distances depending on the environment, a property that turns out to be decisive for parameter selection.

Validation against GPS floating car data from more than 60,000 Beijing taxis showed the method capturing real congestion dynamics. On three sections of the South Third Ring Road, the LAC-based speed curves closely tracked the GPS reference, faithfully reproducing speed dips at midday and afternoon peaks. The length of the LAC sequence, denoted n, proved critical on urban roads. When n increased from 2 to 4, the mean absolute error fell from 0.204 to 0.185 on Section 1, from 0.235 to 0.115 on Section 2, and from 0.307 to 0.246 on Section 3, with the largest improvement on the straight, highest-density section where short sequences are most ambiguous. But pushing n to 5 raised the error again, for example from 0.115 to 0.179 on Section 2, revealing a U-shaped error curve produced by the superposition of two competing effects: matching bias, which shrinks as sequences lengthen, and sampling variance, which grows as the eligible user pool contracts.

That precision-coverage trade-off is steep. On the South Third Ring Road, effective sample sizes peaked at n = 2, reaching roughly 1000 samples per hour on Section 1, about 400 on Section 2, and fewer than 100 on Section 3, but plummeted to fewer than 100 per hour everywhere by n = 5. Most mobile users cross only one or two LAC boundaries during a typical trip, so longer sequences exclude a heavy-tailed majority. Counterintuitively, the urban optimum at n = 4 operates with only about 50 to 100 samples per hour, demonstrating that in dense networks, eliminating incorrectly matched users more than compensates for losing ambiguous ones. On the Jingcheng Expressway the picture was entirely different: mean absolute errors of 0.177 to 0.210 barely varied across all n values, because each LAC covers over 10 kilometers of expressway, so even the shortest sequence spans enough road to uniquely identify the corridor. The researchers recommend n = 4 for urban arterials but n = 2 to 3 for suburban expressways.

One systematic discrepancy emerged on urban roads: CDR-based speeds ran consistently below GPS speeds during congestion dips, while on the expressway the two sources aligned almost perfectly. The researchers attribute this to the difference between person-based and vehicle-based sampling. Floating car data comes almost entirely from taxis, a vehicle-averaged measure, whereas CDR sampling weights each vehicle by its number of active mobile devices. On urban arterials, buses carry high passenger densities but operate at lower speeds amid frequent stop-and-go, pulling the person-averaged estimate below the vehicle-averaged one. On a homogeneous expressway with no bus stops, the two measures converge. The authors caution that this vehicle-composition mechanism could not be directly verified in the present study, since the data lack information on passenger occupancy and travel modes, and they suggest CDR-derived urban speeds should be interpreted as a lower-bound estimator of general traffic speed.

The study has clear limits. Validation covered only two corridors in a single metropolitan area, the roughly 1800-second daytime sampling interval constrains real-time applicability, and the method assumes static base station coverage footprints that may shift under dynamic load balancing. Performance may also vary across operators, whose LAC configurations and location-update strategies differ. Still, the findings demonstrate that LAC-sequence matching is practically feasible and error magnitudes are comparable to those reported for cell-sequence methods, which typically achieve 10 to 25 percent errors. The researchers propose three directions forward: an adaptive framework that tunes sequence length dynamically to local network density and sample availability, machine-learning approaches to disaggregate trajectories by travel mode before speed estimation, and validation on additional road types and evolving 5G architectures, where ultra-dense small cells may both challenge and enhance the matching paradigm. If those steps succeed, the invisible boundaries drawn by telecom operators could become one of the cheapest traffic sensors ever deployed.

Subject of Research: CDR-based road travel speed estimation using a Location Area Code sequence matching method

Article Title: Call detail records (CDR)-Based road travel speed estimation: a location area code (LAC) sequence matching method

Article References: Wu, P., Pang, Q., Lai, J., Yao, H., & Zhou, K. (2026). Call detail records (CDR)-Based road travel speed estimation: a location area code (LAC) sequence matching method. Frontiers in Future Transportation, 7, Article 1884665. https://doi.org/10.3389/ffutr.2026.1884665

Image Credits: AI Generated

DOI: 10.3389/ffutr.2026.1884665

Keywords: Call Detail Records, Location Area Code, travel speed estimation, traffic monitoring, Floating Car Data, map matching, mobile phone data, intelligent transportation systems, Beijing, precision-coverage trade-off, base station density, transportation systems modeling