Forecast Every Retail Unit – Even One With No History
A unit-level retail forecasting pipeline built on a time-series foundation model, producing 36-month probabilistic forecasts for every unit — including vacant ones.
- ThinkCol built a unit-level retail sales forecasting pipeline on the Chronos-2 time-series foundation model, replacing manual, portfolio-average forecasts.
- An ETL pipeline turns messy tenant records into unit-level series, fused with Hong Kong macro signals such as inflation, visitor arrivals, and retail trends.
- The model produces 36-month probabilistic forecasts with confidence bands for every unit, including new or empty units with no sales history.
A premium mall operator relied on manual, inconsistent forecasts based on portfolio averages. The team had no way to forecast sales for new or vacant shops, or to test how a change in tenant mix would affect performance before signing a lease.
ThinkCol built a unit-level retail sales forecasting pipeline on the Chronos-2 time-series foundation model. An ETL pipeline turns messy tenant records into clean unit-level series, and the model produces 36-month probabilistic forecasts with confidence bands for every unit, fused with Hong Kong macro signals such as inflation, visitor arrivals, and retail trends.
The pipeline handles four leasing scenarios out of the box — lease renewals, what-if trade-mix changes, units with thin sales history, and brand-new or currently empty units — with no per-unit models to build or maintain.
Why customise? Pre-built forecasting tools need long per-unit history, can't blend messy external data, and can't answer what-if tenant-mix questions.
