Make the signals usable
Detect outliers, fill missing observations, examine rolling windows, and explore transformations and stationarity.
Time-series analysis · 2022
Detecting changes in everyday app usage.
Understanding how smartphone habits evolve by preparing, analysing and forecasting daily behavioural signals across a large user sample.
A changing habit is difficult to forecast.
Daily screen time, phone-check frequency and app usage can fluctuate substantially. Missing observations, seasonality and shifts in behaviour complicate the interpretation of a single time series.
The goal was to explore how statistical analysis and forecasting can help distinguish recurring patterns from meaningful changes in usage.
From noisy activity to interpretable trends.
The project used exploratory time-series analysis, data preparation and forecasting benchmarks to understand behaviour over time.
Compare forecasting approaches
Evaluate a persistence baseline, seasonal ARIMA and Prophet on daily numbers of phone checks.
Investigate shifts
Use change-point analysis, including PELT, to explore potential changes in smartphone behaviour.
What the experiments indicated.
Prophet performed best in the reported comparison
The original project reports that Prophet outperformed persistence and seasonal ARIMA for forecasting 30 days of phone-check counts across an experiment involving 300 users.
Context matters for time-series analysis
Exploring seasonality, autocorrelation and missing data helped build a better understanding of the behavioural signals before modelling.
Change detection extends the analysis
Change-point methods offered a complementary perspective to forecasting by looking for shifts in usage patterns.
The simplest baseline deserves a seat at the table.
Forecasting is a comparison problem, not just a modelling problem. Testing a method against plausible alternatives matters more than choosing the most fashionable algorithm.
The project also reinforced the value of careful preprocessing and communicating what can, and cannot, be inferred from behavioural data.
Explore the original analysis.
Read the source project for the full methodology, change-point work and forecasting experiments.