It is the 1970s, and forecasting has a seductive disease: add parameters and the fit to past data always improves, right up until the model has memorized the noise and predicts the future terribly. You hold five nested time-series models, each fitting the history a little better than the last, and you must choose one to forecast next quarter — on principle, not on which curve hugs the old data most tightly. Penalize complexity in a way you can justify from information theory, not taste. Get it wrong and you ship the overfit model: it dazzled on the backtest and fails the moment real money rides on the forecast — the whole art is trading fit against parsimony before reality, not after, exposes the bloat.
Chose Sparse recovery — right call.
Nowak's work in statistical signal processing and machine learning — active learning theory, sparse recovery, and notably earlier multiscale methods with explicit complexity-regularized estimation (penalized likelihood tree pruning for Poisson inverse problems, around 2000) — means he has actually used and proved risk bounds for penalized model selection, the same mathematical species as AIC. His minimax analyses of adaptive sampling show command of the bias-variance calculus. The gaps: his penalties are calibrated by oracle inequalities for signal reconstruction, not derived from predictive Kullback-Leibler information theory; his settings are images and sensing systems, not short economic time series; and he postdates the era by decades. Among modern signal-processing carriers he is unusually close to the problem's actual mathematics, but the specific 1970s craft — nested ARIMA orders on forty points — is inherited, not his.
LeCun has a specific, citable claim on this problem's theme: Optimal Brain Damage (1990, with Denker and Solla) used second-derivative information to delete parameters that contribute least, explicitly arguing that removing excess capacity improves generalization — a principled, quantitative complexity reduction, cousin to the counting in AIC. His convolutional networks are themselves an argument that built-in structure (weight sharing) beats profligate parameterization. But his anti-overfitting practice is validation-driven engineering inside one flexible model family, not selection among nested parametric models by an information-theoretic criterion; he has often argued, post-2015, that overparameterized networks generalize despite classical warnings, which inverts the problem's premise. On a short 1970s quarterly series his tools cannot even be trained. Genuine conceptual overlap through OBD; little grip on the stated task.
Battle #72 · 8/10/2026, 11:36:11 AM · this result is deterministic: the same two personas on this problem always resolve the same way.