A Better Screen Is Not a Crystal Ball

This series has examined one specific subject: early-stage new technology-based firms being screened for scarce support before their futures are known. The final question is whether customer transaction maturity can beat gut feel—and what that result is worth once survival, growth, causation, and generalization are kept separate. The defensible answer is yes in a narrow predictive comparison, but no as a forecast of winners. Customer relations can improve triage; they cannot remove uncertainty from venture building.

That answer may sound less dramatic than either side of the usual argument. It is more useful. A decision-maker does not need a signal to reveal the entire future in order for the signal to improve an early screen. They need to know what information it adds, how much confidence it deserves, what decision it can support, and what evidence must replace it as the venture develops.

What the customer signal actually buys

The focal study compares 272 NTBF business plans from a high-tech competition in Baden-Württemberg, submitted between 2006 and 2014, with observed survival five years later. Of the 272 ventures, 165 survived, so a rule that predicted survival for everyone achieved 60.66% correct classifications. That sample-mean rule is intentionally agnostic: it uses the majority outcome and ignores the individual plan.

The raw non-expert benchmark reached 61.76%, correctly classifying 168 cases. The improvement over the baseline was only 1.10 percentage points, and the paper did not find a significant improvement. That result does not prove that people cannot read useful signals in venture plans. It shows that this specified group of non-experts, using a majority-vote rule on this sample, did not reliably outperform the easiest majority prediction.

The customer-only objective model reached 64.34% correct classifications. In the reported comparison, that is 3.68 percentage points above the sample mean and 2.58 points above the raw non-expert majority. Customer transaction maturity was the dimension retained by the cross-validated LASSO procedure, while financier, partner, and human-resource relations did not provide enough stable incremental predictive value under that specification. The customer coefficient was positive and highly significant in the follow-up logistic model, supporting an improvement over the agnostic baseline.

This is the part of the evidence that earns the title’s first clause. A defined, text-derived customer signal did better than the raw intuition benchmark used by the study and did better than simply predicting the common outcome. But the margin is modest, the sample is narrow, and the comparison depends on the prediction rule. The result is not “customer transactions predict the future” in the unlimited sense. It is “customer transaction maturity added information to a five-year survival classification in this study.”

The enhanced subjective model shows why even that sentence needs care. When the share of non-experts predicting survival was used as an input to a logistic model, the reported in-sample accuracy reached 65.80%, slightly above the customer model’s 64.34%. The authors did not treat that observed-sample number as the end of the comparison. They used K-fold cross-validation and found higher predictive error for the subjective measure across the tested settings. The objective advantage therefore belongs to the more cautious, out-of-sample-minded comparison, not to a universal claim that every model built from human judgments will score lower.

The distinction between these results is a portable lesson. “Beat gut feel” is never a complete claim until the reader knows whose intuition, what aggregation rule, what denominator, what outcome, and what validation procedure are involved. A majority vote, a fitted probability, and a cross-validated error estimate are related but not interchangeable. A model can look better in the observed sample and still be less reliable when the cases used to tune it are not the cases used to test it.