A tissue microarray (TMA) containing diagnostic biopsies was used to develop predictors of outcome in a group of 105 patients having advanced-stage follicular lymphoma (FL). The patients were staged and uniformly treated, and the usable cases had been randomly divided into a subgroup of 50 patients with outcomes identified, and a reserved subgroup of 43 patients whose outcomes were masked for blind testing of the predictors. Using training-input data from some patients with known outcomes, parallel cascade identification developed two predictors of overall survival based on a number of biomarkers. Both predictors had statistically significant performance over the remaining patients with known outcomes. The first predictor had been identified with model architectural settings and encoding scheme chosen, for the particular training input used, to enhance classification accuracy over remaining patients in the known subgroup. The second predictor was obtained without changing the settings and encoding scheme, but from an entirely different training input corresponding to novel cases from the TMA. Not surprisingly, the first predictor showed much higher accuracy over the known subgroup, but when tested over the reserved subgroup of 43 patients, averaged about 58% correct and did not reach statistical significance. The other predictor performed very similarly over the known and the reserved subgroups, with prediction on the reserved subgroup highly significant at p = 0.0056 in Kaplan-Meier survival analysis. We conclude that a predictor based on a number of biomarkers obtainable at diagnosis has the potential to improve prediction of overall survival in FL.