What the Numbers Actually Show

This chapter reads the focal study’s results as a comparison among specific prediction rules, not as a contest between “data” and “human wisdom.” The subject remains the early-stage NTBF and its five-year survival, while the key judgment is how much information customer transaction maturity adds beyond a sample-wide majority prediction and a defined non-expert benchmark. The answer is positive but modest, and its strength depends on keeping in-sample fit, cross-validation, and outcome scope separate.

The headline figures are easy to repeat and easy to misread. The study reports 60.66% correct predictions for the sample-mean baseline, 61.76% for the raw non-expert majority vote, and 64.34% for the customer-based objective model. Those percentages do not all describe the same kind of prediction. Their differences become meaningful only when each is attached to the rule that produced it and to the validation question it was meant to answer.

Three classifiers, three different claims

The sample-mean baseline is the simplest. Of 272 NTBFs, 165 survived, so the survival rate was 60.66%. Because survival was the majority outcome, the reference model predicted that every venture would survive. It was correct for all 165 survivors and wrong for the 107 non-survivors. Its 60.66% accuracy is therefore not evidence that the sample mean understands individual ventures. It is the level a proposed screen must exceed if it is to justify using information from the plans.

The non-expert majority vote improves on that baseline by only 1.10 percentage points, reaching 61.76%, or 168 correctly classified cases. The paper concludes that this difference does not represent a statistically meaningful improvement over the reference model. That conclusion is important because a small numerical edge can look more impressive than it is when the baseline already predicts the common outcome for everyone. The raw intuition measure did not show that reading the plans as non-experts produced a reliably better survival classification than ignoring the plans.

The error pattern adds context. Non-experts correctly predicted 66% of ventures that survived, but only 55% of ventures that did not. Their judgments were therefore not equally useful across the two outcome classes. A screener who wants to identify ventures at risk of failure may care particularly about the second number. Accuracy alone does not say whether a method is good at finding fragile ventures, good at avoiding false alarms, or simply good at following the majority class.

The customer-based objective model reaches 64.34% correct predictions. That is 3.68 percentage points above the sample-mean baseline and 2.58 points above the raw non-expert majority vote in the reported comparison. Its customer coefficient is positive and highly significant in the ordinary logistic model, and the paper treats this as evidence that the objective customer measure improves on the agnostic reference. The result is not a dramatic separation. It is a measurable, bounded improvement in a sample where most ventures survived.

The case count keeps the magnitude in perspective. The raw majority vote gets 168 of 272 cases right, only three more than the 165 cases captured by predicting survival for everyone.