Predicting Where They Go Next: Next-Location Models

Human movement is deeply habitual. We return to the same coffee shop on Tuesday mornings, follow the same commute corridor, and cluster our social lives around a predictable orbit of venues. This predictability, which we rarely consciously register, has become one of the most studied problems in applied machine learning. The field of next-location prediction (NLP) asks a deceptively simple question: given where someone has been, where will they go next? The implications for open source intelligence (OSINT) are profound and, depending on one's vantage point, either thrilling or deeply unsettling.

This article examines how NLP technology intersects with modern OSINT practice: where the academic and operational worlds have converged, where they remain stubbornly apart, and what the near future likely holds.

What Next-Location Prediction Actually Is

In academic machine learning, next-location prediction is a well-scoped technical task. A model ingests a sequence of historical trajectory data, including GPS traces, cell tower pings, check-in records, and social media geotags, and returns the most probable next destination. Research in this area has accelerated considerably over the past decade, with early Markov chain models giving way to recurrent neural networks, attention mechanisms, and now transformer architectures.

Hong et al. (2022) proposed a transformer decoder architecture that improved on prior state-of-the-art methods by a margin of more than five percent in F1-score across two large GPS tracking datasets, by incorporating not only location and time but travel mode as an auxiliary learning signal.^[1] The core insight is that movement is multimodal: a person traveling by train has a different set of plausible destinations than one who just stepped off a bicycle.

Academic NLP research has broadened well beyond trajectory sequences. Luca et al. (2025) demonstrated that overlapping training and test trajectories inflate benchmark accuracy scores, warning that reported model performance in the literature may significantly overstate real-world generalization.^[2] Meanwhile, other researchers have shown that integrating inferred activity semantics, such as whether a person is likely heading to work, running errands, or dining out, substantially improves prediction accuracy over purely spatiotemporal models (Chen et al., 2025).^[3]

The underlying assumption, supported by decades of mobility research, is that human movement exhibits strong regularity. The practical consequence is high: a 2022 survey found that next-location prediction has become relevant to applications ranging from traffic management and urban planning to epidemic prevention and what the authors more delicately described as "suspicious target tracking" (Dawit et al., 2022).^[4]

The OSINT Parallel: Pattern of Life Analysis

Within the intelligence and law enforcement communities, NLP as a concept travels under a different name. Pattern of life (POL) analysis is the operational practice of mapping a subject's routine behaviors and movements over time to establish a behavioral baseline, detect deviations from it, and anticipate future actions. It is, functionally, next-location prediction performed with a mixture of human judgment and automated tooling.

Investigators studying routine behaviors and movements establish behavioral baselines against which anomalies become visible, with unusual deviations pointing to important investigative leads (Axligence, 2024).^[5] The language differs from the academic literature, but the underlying inference is identical: historical trajectory implies future trajectory.

Social media has become a primary data source for this type of analysis in the OSINT context. Law enforcement investigators can anticipate suspect movements by recognizing that subjects regularly tag posts at recurring locations such as gyms or cafes, allowing them to map habitual presence and predict future attendance (Burley, 2025).^[6] This is manual pattern-of-life analysis executed against open-source geotag data, rather than a trained NLP model, but the inferential logic is functionally equivalent.

Commercial Platforms Bringing Prediction to Practice

The gap between academic NLP and operational intelligence is narrowing, driven by a generation of commercial platforms that package predictive location analytics for government and law enforcement customers.

Movement Intelligence Vendors

Venntel, a subsidiary of Gravy Analytics, markets what it describes as track-level movement data enabling exact behavioral insights and movement history over time, with analytic tools that turn raw geolocation signals into context-rich operational intelligence (Venntel, 2025).^[7] The platform is marketed explicitly to federal law enforcement, national security organizations, and defense agencies. Its core capability goes beyond historical mapping: the explicit pitch is connecting historical movement patterns to behavioral intent and anticipating future movement.

LexisNexis Risk Solutions offers a similar capability within its Accurint Virtual Crime Center. The platform's TraX module enables investigators to conduct pattern of life analysis using automated, data-enhanced call detail record analysis (LexisNexis Risk Solutions, 2025).^[8] This is predictive location work at scale, surfaced through a law enforcement-grade interface with built-in evidential compliance features.

Cybercheck, operating under what it terms a "CyberDNA" framework, uses machine learning and automation to surface location information criminals leave in their digital wakes, enabling investigators to reconstruct travels and identify suspects without requiring warrants or subpoenas to access initial intelligence (Lindsay, as cited in Butkus, 2026).^[9] The forward-looking dimension of this, forecasting where a subject is likely to be rather than merely where they have been, is built into the platform's analytical logic.

Predictive Policing and Hotspot Mapping

At a more aggregate level, predictive policing platforms use location intelligence to forecast where criminal activity is likely to occur. These systems analyze historical crime data, real-time data feeds, and social network analysis to identify areas of elevated risk (Davies, 2024).^[10] While this is place-level rather than individual-level prediction, it shares the core architecture of NLP: past spatial patterns predict future spatial events.

The geospatial intelligence layer is also expanding. Pertsol (2025) notes that integrating machine learning with location data enables predictive analytics where future scenarios can be predicted based on historical data, including link analysis that examines connections between individuals based on shared locations to map out networks.^[11] This is the connective tissue between traditional OSINT network analysis and NLP: not just who was where, but who will be where, and with whom.

The Data Problem and the OSINT-Specific Challenge

Academic NLP models typically require dense, longitudinal, individual-level GPS or cell tower data to train and infer from. This creates a fundamental tension with the OSINT context, where data is gathered from public or semi-public sources rather than from direct access to a subject's device or carrier records.

OSINT geolocation instead relies on a patchwork of signals. EXIF metadata embedded in photographs can contain GPS coordinates, timestamps, and device identifiers (Neotas, n.d.).^[12] Social media platforms, particularly Instagram, TikTok, and Snapchat, provide geotagged content that can be aggregated into informal trajectory records. IP address geolocation, while imprecise, adds another layer. Satellites provide imagery that can confirm or refute presence at a location. When combined, these signals can approximate the kind of trajectory data that academic NLP models require, but with lower precision and greater noise.

Law enforcement investigators can use geolocation data to track potential suspects by identifying the location of their electronic devices, since individuals tend to keep their devices close to them (Blinkman, 2024).^[13] The passive signal leakage from connected devices, including Wi-Fi access point associations, Bluetooth pings, and app-generated location broadcasts, creates a rich ambient data environment that commercial OSINT platforms have learned to exploit without requiring direct device access.

Anticipatory Intelligence: The Broader Context

Next-location prediction in OSINT does not operate in isolation. It sits within a broader movement toward what the 2019 National Intelligence Strategy called "anticipatory intelligence," described as intelligence that addresses new and emerging trends, changing conditions, and underappreciated developments (Office of the Director of National Intelligence, as cited in Dahl & Strachan-Morris, 2024).^[14]

The direction of OSINT as a discipline reinforces this orientation. OSINT is moving from a static library model toward live-stream monitoring of protest locations, crisis responses, war zones, financial movements, and extremist activity in encrypted spaces, becoming a live radar rather than an archival resource (Azutech, 2025).^[15] NLP fits naturally into this real-time, anticipatory posture: rather than reconstructing what happened, analysts want to know what is about to happen.

Law enforcement and military operations are increasingly integrating next-generation OSINT technologies, including geospatial intelligence, AI-driven facial recognition, and real-time social media monitoring for improved situational awareness (ShadowDragon, 2026).^[16] Location prediction is the logical next capability in that stack.

Where the Gap Remains

Despite the convergence described above, a meaningful gap persists between academic NLP research and routine OSINT practice. Several factors explain it.

First, the terminology does not travel. Intelligence practitioners use "pattern of life," "movement analysis," and "behavioral prediction" rather than "next-location predictor" as terms of professional art. This means that practitioners may be doing NLP-equivalent work without labeling it as such, and that academic researchers working on NLP models may not be aware of how their work is being adapted at the operational level.

Second, the human-in-the-loop norm remains strong within professional OSINT. AI should not yet be making decisions alone in investigative contexts; human and machine collaboration is required to ensure accurate, ethical results (Blackdot Solutions, 2025).^[17] Fully automated individual next-location forecasting, while technically feasible, does not align with current professional norms or legal frameworks in most jurisdictions.

Third, legal constraints create headwinds. Geofence warrants, which historically allowed law enforcement to identify devices present at a location during a given time window, became a flashpoint for Fourth Amendment litigation. Google's decision to eliminate internal location history storage in late 2024 removed one of the most widely used sources of passive movement data for investigators (Butkus, 2026).^[9] The legal terrain around predictive location work, as distinct from retrospective location evidence, remains unsettled.

Finally, bias and feedback loops are significant concerns. Predictive policing systems that identify a neighborhood as high-risk can lead to increased surveillance presence, which produces more arrests, which reinforces the algorithm's initial prediction (Davies, 2024).^[10] Individual-level NLP applied to OSINT contexts carries analogous risks: a subject flagged as likely to appear at a location may be surveil under increased scrutiny regardless of the model's actual reliability.

The Road Ahead

The near-term trajectory, if the parallel can be forgiven, is toward tighter integration. As OSINT data sources multiply and AI-assisted analytics become standard in law enforcement and intelligence platforms, next-location prediction will increasingly be embedded in broader workflow tools rather than deployed as a standalone capability. Practitioners may not know they are using NLP; they will simply see a map highlight that suggests a subject is likely to appear at a particular location within a particular time window, with a confidence interval drawn from historical trajectory analysis.

The academic literature, meanwhile, is moving toward more honest evaluation of real-world generalization. Luca et al.'s (2025) work on train-test overlap suggests that many published NLP models perform substantially worse in novel mobility conditions than benchmark numbers imply.^[2] Closing that gap will be necessary before NLP can be trusted for high-stakes anticipatory intelligence applications.

What is already clear is that the question is no longer whether location prediction will be part of OSINT. It already is, in multiple forms, at multiple scales. The more pressing questions are ones of governance: who decides when a prediction is accurate enough to act on, what disclosure obligations attach to AI-assisted location forecasting in legal proceedings, and what safeguards prevent the feedback loops that turn predictive surveillance into self-fulfilling persecution.

Those questions will not be answered by better algorithms. They will require policy frameworks that, as of March 2026, remain largely unwritten.