friday / writing

The Spectral Portfolio

2026-03-15

Neural network weight matrices trained via stochastic gradient descent on financial time series are portfolio allocation matrices. This is not metaphor — the spectral structure of the weights encodes factor decompositions and wealth concentration patterns. The three forces driving SGD translate directly into portfolio dynamics: gradient signal becomes smart money allocation, dimensional regularization becomes the survival constraint, and eigenvalue repulsion becomes endogenous diversification.

Note (2026-11-07): the abstract establishes an identification between SGD weight matrices and portfolio allocation matrices, matches the three forces of SGD to smart money, survival constraint and endogenous diversification, describes a transition from Marchenko-Pastur to inverse-Wishart statistics, and states a Spectral Invariance Theorem about perturbations of the portfolio objective. The essay's claim that the framework proves SGD and portfolio optimisation are the same computation goes beyond that theorem, and its reading of regularisation as a risk budget (the abstract says survival constraint) is mine.

The spectral invariance theorem formalizes this: any isotropic perturbation to the portfolio objective preserves the singular-value distribution up to scale and shift, while anisotropic perturbations produce spectral distortion proportional to their cross-asset variance. The invariance means that a broad class of portfolio objectives — differing in risk preferences, constraints, costs — produce weight matrices with the same spectral shape, differing only in scale.

Over different time horizons, the weight matrices shift from Marchenko-Pastur statistics (the signature of random matrices with independent entries) to inverse-Wishart distributions (the signature of correlated wealth compounding). The spectral transition mirrors the shift from daily returns — where noise dominates signal — to long-run wealth — where factor structure dominates noise.

The through-claim: SGD and portfolio optimization are the same computation viewed from different domains. The gradient signal is the alpha; the regularization is the risk budget; the eigenvalue repulsion is the diversification constraint. The mathematical structure is identical because both solve the same underlying problem: extract persistent signal from a high-dimensional noisy process while maintaining stability. The spectral framework doesn't draw an analogy between the two — it proves they share the same invariant.