Predictive Attendance Analytics and Forecasting

A research and development direction rather than a deployed system. The question is whether historical attendance patterns across a large, distributed workforce carry enough structure to anticipate shortfall and coverage risk in advance — so that the response is planned rather than reactive. The modelling techniques under investigation belong to the same class of approach used in the Godtrader market-research platform for signal generation and strategy evaluation.

Attendance in the organisation is currently understood retrospectively. Divisional officers can see who has been absent and what the attendance rate was, but the question that matters operationally is prospective: given the pattern of who is likely to be absent next month, what should be planned now? Without a forward view, coverage is managed reactively, and reactive management of a distributed workforce is expensive.

What made it hard

• Began by establishing whether a signal exists at all — a forecasting exercise without a baseline is not interpretable, so historical variation had to be characterised before any model was fitted.
• Separated the problem into detecting predictable structure and deciding what to do with a prediction, because the second is an operational judgement rather than a modelling question.
• Recognised the domain-transfer problem: a model fitted to one division's attendance may not hold for another, which suggests a hierarchical treatment rather than a single pooled model.
• Framed the work to produce an explainable output — a forecast a divisional officer can act on, with the reasoning visible — rather than an opaque number.

The approach under investigation follows the pattern established in the market-modelling work. Historical attendance is treated as a time series with both trend and regime characteristics: regime detection classifies periods by their underlying behaviour, and a supervised classifier estimates the probability of absence from the features that precede it. The methodological interest is in combining the two — regime alone is too coarse to plan against, and a classifier without regime context produces a probability the reader cannot interpret. The intended output is a forward coverage-risk signal per division, with its basis shown.

Feature layer

Historical attendance series prepared into features capturing trend, level, volatility, and calendar structure.

Regime layer

Regime classification to establish the behavioural context in which any prediction is made.

Prediction layer

Supervised classification producing a probability of absence for each subsequent period.

Interpretation layer

An explainable forward signal per division, with the basis for the prediction visible to the person acting on it.

• Assembled a historical attendance dataset spanning divisions, with attention to consistency of capture method across divisions.
• Characterised baseline variation before modelling, to establish whether any predictable structure exists.
• Investigated regime detection as a precondition for interpreting short-horizon forecasts.
• Built and compared candidate predictive models against a simple baseline, keeping the baseline comparison explicit.
• Framed the deliverable as an actionable forward signal rather than a forecast accuracy figure, because a number a divisional officer cannot act on does not change anything.

• Historical attendance series prepared into modelling features.
• Regime classification establishing context for short-horizon forecasts.
• Supervised prediction of absence probability per period.
• Explicit comparison against a simple baseline to test whether complexity earns its place.
• An explainable output designed for operational use.

• This is an ongoing research direction. The method is developed and tested; it is not yet deployed as a production system, and no operational performance figures are claimed.
• The modelling techniques belong to the same class of approach used in the Godtrader market-research platform, where machine-learning components support signal generation and strategy evaluation.

Screenshots for this project have not been published yet.