We selected seven samples. Table 3 shows the years for which we collected data, the number of
companies per sample and how samples were used for estimation or prediction tasks or both. For
each sample collected at time t, company status (healthy vs. failed) was assessed at time t+1, with
an average lag of 12 to 18 months. Balance sheets and income statements were selected from six
consecutive years (t to t − 5) and firms were chosen at random from among those in the database
when they complied with the criteria described above.