James Hannon

Consumer smartwatch readiness scores summarise physiological and sleep data, but their relationships with independently measured physiology and self-reported wellness remain uncertain. This exploratory 14-day repeated-measures observational study examined Google Pixel Watch 3 Daily Readiness in 38 participants (410 participant-days; 220 with recorded readiness scores). Morning heart rate variability was measured independently using Polar H10 and Kubios, and daily questionnaires assessed fatigue, stress, body soreness, and sleep quality. Mixed-effects models separated within-person and episode-mean associations; participant-grouped cross-validation tested incremental wellness prediction. Within participants, each 10-point increase in readiness relative to the episode mean was associated with 6.2% higher root mean square of successive differences (95% CI 2.7–11.7%), a 0.15-unit higher parasympathetic index (0.08–0.23), a 0.12-unit lower sympathetic index (−0.25 to −0.05), and 0.99 beats/min lower heart rate (−1.91 to −0.42). Respiratory rate was not associated. Higher readiness was associated with more favourable fatigue (odds ratio 1.19, 95% CI 1.03–1.38) and stress (1.26, 1.08–1.47) ratings, but not body soreness or sleep quality after multiplicity correction. No episode-mean association survived correction, although estimates were imprecise. Daily Readiness did not improve out-of-sample prediction. These findings show convergence with one physiological input domain but do not validate the composite or assess its use alongside load metrics.

Heart rate variability (HRV), particularly the root mean square of successive differences (RMSSD), is widely used as a non-invasive indicator of autonomic nervous system activity and physiological recovery. This study examined whether daily short-term HRV, measured under standardised morning conditions, was associated with self-reported wellness in a non-clinical adult population. Over a 14-day period, 41 participants completed daily five-minute HRV recordings using a Polar H10 chest sensor and the Kubios mobile app, followed by ratings of sleep quality, fatigue, stress, and physical recovery. Bayesian ordinal mixed-effects models revealed that higher RMSSD values were associated with better self-reported sleep (β = 0.510, 95% HDI: 0.239 to 0.779), lower fatigue (β = 0.281, 95% HDI: 0.020 to 0.562), and reduced stress (β = 0.353, 95% HDI: 0.059 to 0.606), even after adjusting for covariates. No association was found between RMSSD and perceived muscle soreness. These findings support the interpretability of RMSSD as a physiological marker of daily recovery and stress in real-world settings. While the effect sizes were modest and individual variability remained substantial, results suggest that consistent HRV monitoring may offer meaningful insight into subjective wellness—particularly when contextualised and tracked over time.

This paper uses a two-step approach to modelling the probability of a policyholder making an auto insurance claim. We perform clustering via Gaussian mixture models and cluster-specific binary regression models. We use telematics information along with traditional auto insurance information and find that the best model incorporates telematics, without the need for dimension reduction via principal components. We also utilise the probabilistic estimates from the mixture model to account for the uncertainty in the cluster assignments. The clustering process allows for the creation of driving profiles and offers a fairer method for policyholder segmentation than when clustering is not used. By fitting separate regression models to the observations from the respective clusters, we are able to offer differential pricing, which recognises that policyholders have different exposures to risk despite having similar covariate information, such as total miles driven. The approach outlined in this paper offers an explainable and interpretable model that can compete with black box models. Our comparisons are based on a synthesised telematics data set that was emulated from a real insurance data set.