People affected by cancer want information about their prognosis but clinicians have trouble estimating and talking about it. We sought to determine the nature and accuracy of medical oncologists' estimates of life expectancy in newly referred patients with incurable cancer. With reference to each patient, medical oncologists estimated how long they thought 90, 50, and 10% of similar patients would live. These proportions were chosen to reflect worst case, predicted, and best case scenarios suitable for discussions. After a median follow-up of 35 months, 86 of the 102 patients had died with an observed median survival of 12 months. Oncologists' estimates of each patient's worst case, predicted and best case scenarios were well-calibrated: 10% of patients lived for fewer months than estimated for the worst 10% of similar patients; 50% lived for at least as long as estimated for 50% of similar patients (predicted survival), and 17% lived for more months than estimated for the best 10% of similar patients. Oncologists' estimates of each patient's predicted survival were imprecise: 29% were within 0.67-1.33 times the patient's actual survival, 35% were too optimistic (>1.33 times the actual survival), and 39% were too pessimistic (<0.67 times the actual survival). The proportions of patients with actual survival times bounded by simple multiples of their predicted survival were as follows: 61% between half to double their predicted, 6% at least three to four times their predicted, and 4% no more than 1/6 of their predicted; similar to the proportions in an exponential distribution (about 50%, 10% and 10% respectively). Ranges based on simple multiples of the predicted survival time appropriately convey prognosis and its uncertainty in newly referred people with incurable cancer.