Final Year Project · BSc (Hons) Computer Science · University of Westminster · 2026 to 2027 · work in progress
Forecasting the Direction of Everyday Routine from Passive Sensing, with Calibrated Confidence
University of Westminster, London
Abstract
Phones and wearables record how people sleep, move and use their devices, week after week. Most work on this data asks what state someone is in right now. This project asks where their routine is heading. Given a few weeks of a person's passive sensing data from the public GLOBEM dataset, the model matches their partial trajectory against people with similar histories and forecasts the direction of change over the following weeks. A k nearest neighbours baseline comes first, with a sequence model as a stretch goal. Because a forecast about a person is only useful if its confidence is honest, the second contribution is calibration: every forecast is evaluated on cohorts the model has never seen, using expected calibration error, Brier score and reliability diagrams, with Platt and isotonic recalibration.
Keywords: longitudinal behaviour modelling, passive sensing, trajectory forecasting, probability calibration, human digital twins
1. Introduction
GLOBEM made multi year passive sensing data public and showed how hard it is for behaviour models to generalise across people and years [1]. Work built on it mostly detects a current wellbeing label. Surveys of human digital twins note that modelling how a person changes over time is still largely conceptual [2]. This project takes one small, testable piece of that idea: forecasting the direction of routine change.
2. Approach
Weekly features for sleep, physical activity and phone use are built from GLOBEM. For a partial trajectory, the nearest neighbours in the training cohort vote on the direction each feature moves next. This keeps the baseline interpretable: you can see which people the forecast was built from.
3. Evaluation
Models are trained on one cohort and tested on others, including later years. Accuracy alone is not enough. A model that says it is 80% sure should be right about 80% of the time, so calibration is measured directly and corrected where it drifts [4].
4. Scope
The horizon is kept short and honest, because GLOBEM covers roughly ten weeks per person per year. This is a calibrated forecaster, not a full digital twin, and it makes no clinical claims.
5. Status
Proposal in November 2026, interim demonstration in February 2027, final report in April 2027.
References
- Xu, X. et al. (2022) GLOBEM Dataset: Multi-Year Datasets for Longitudinal Human Behavior Modeling Generalization. NeurIPS Datasets and Benchmarks.
- Lin et al. (2024) Human digital twin: a survey. Journal of Cloud Computing.
- Simulating Life Paths with Digital Twins: AI-Generated Future Selves Influence Decision-Making and Expand Human Choice (2025). arXiv:2512.05397.
- Guo, C. et al. (2017) On Calibration of Modern Neural Networks. ICML.