A visual map of the field split Breiman sharpened: explicit stochastic models for explanation versus algorithmic systems optimized for held-out prediction. The page emphasizes the geometry of that divide, how methods cluster, and where modern hybrids blur the border.
Methods are plotted by two axes: explanatory structure on the horizontal, predictive flexibility on the vertical. The toggle changes which culture gets visual emphasis.
Breiman's complaint was not that classical models never work. It was that many analysts optimized elegance, asymptotics, and fit stories where messy forecasting demanded serious out-of-sample competition.
Coefficients, hazard ratios, calibrated intervals, and mechanism stories matter most when the output is a decision rule people must justify.
Trees, ensembles, and neural nets gain ground when interactions are ugly, the feature space is wide, and tomorrow's error matters more than today's interpretive neatness.
A synthetic comparison across model families. Cells encode relative strength on interpretability, calibration, nonlinearity, interaction capture, sample efficiency, and raw predictive power on messy data.
Four stylized learners show different inductive biases. Parametric lines commit early. Trees partition. Ensembles average jaggedness away. Deep nets carve smoother but harder-to-read surfaces.
The toggle moves between stylized data regimes. The pattern mirrors Breiman's thesis: flexible algorithms widen their margin as dimensionality, hidden interaction structure, and misspecification risk rise.
Points show how methods move along a frontier rather than occupying pure camps. The orange ridge marks the modern hybrid zone: GAMs, conformal layers, and neural additive models.
Residual plots, significance tables, and fit checks are not useless. Breiman's point is that they are incomplete if they stand in for genuine prediction tests.
In medicine, reliability, and policy, performance alone is not enough. This mini-chart maps why Cox-style models, conformal wrappers, and calibrated hybrids remain important.
The dispute predates modern machine learning. Tukey widened data analysis, Box warned that models are useful despite being wrong, Breiman sharpened the provocation, and Shmueli formalized explain-versus-predict as separate tasks.
Modern practice is less a duel than an ecosystem: benchmarks, ensembles, deep nets, conformal prediction, mechanistic interpretability, and structured hybrids all respond to the same old tension from different angles.
Compact anchors for the visual story: the original manifesto, the prehistory, the ensemble turn, and later work on prediction, interpretability, and uncertainty.