Smart home devices struggle to personalize sleep features due to the absence of automatic adjustments based on individual sleep patterns. Examples of automatic adjustments include automatic blue light and temperature adjustments, falling short of user expectations for advanced health solutions. Our business seeks to provide smart home device manufacturers with personalized sleep prediction solutions to enhance user experience. The goal is next-day-ahead forecasts of individuals' bedtimes to promote sleep readiness. The client, a manufacturer of smart home devices, can utilize our forecasting model to enhance their sleep readiness product, using the forecasted bedtime to turn on/off features of their device that enhance sleep readiness. By incorporating our predictive model into their product, they aim to offer users a more intelligent and personalized sleep experience.
The data comprises 84 days of bedtimes from 4 students. The data was collected through a combination of each user's electronic devices as well as self-reported information. Additional predictors of bedtime were explored, including phone use, computer use, walking distance, holidays, and workloads, etc.
In our forecasting solution, we employ various machine learning methods while utilizing simple forecasting methods as benchmarks. The dataset includes 84 days of data, covering the period from September 25, 2023, to December 17, 2023. Roll-forward validation was used to incorporate new daily data into the analysis. We tested multiple statistical and machine learning methods for predicting bedtimes. Models were assessed by comparing forecasting performance. The selection process favored simple models with fewer variables which were shown to have better performance in the final forecasts. Final models were visualized to be evaluated through plots and checked for overall error metrics and bias towards under- or over-forecasting. Also we found forecasting bedtimes within 15 intervals did not improve forecasting accuracy, unfortunately.
Overall, it was discovered that accurate prediction of certain bedtime patterns can be challenging without external information, highlighting the importance of incorporating external factors into forecasting models. It is crucial to recognize that the optimal forecasting models can vary significantly from person to person, underscoring the need for personalized approaches. Utilizing electronic devices to automate data collection processes can improve efficiency and reduce human errors. To enhance existing forecasting models, it is recommended to gather a year-long dataset to analyze seasonal trends and patterns comprehensively. Additionally, exploring the impact of screen usage timing on sleep beyond daily totals, including late-night phone use, can provide valuable insights.