AbstractPredictions of the tender price of proposed buildings are needed during early design. Conventionally, these are prepared manually and based on limited data.A computer system was produced which prepares estimates using a library of data containing rate, quantity and date for the constituent elements of previous buildings, inflation indices and statistical models.Because the scatter of the rate vs quantity data for each element in a sample of buildings was so variable between samples, no one model was adequate. Thirty two different models were included, together with a criterion for selecting the most appropriate. This ensured the most precise prediction possible.The co-efficient of variation of the ratio forecast to actual price measured for a set of system-produced estimates grouped by building type ranged from 10 to 19%. Current manual practice ranges from 6 to 21%, but this is only achieved much later when design is complete and more reliable data is available.