Abstract

This paper develops a self-contained mathematical framework for smoothing
heteroskedastic financial time series within a reproducing kernel Hilbert space
(RKHS) setting. The central object is the Information Flow Kernel, a positive
definite kernel constructed as the pointwise (Hadamard) product of the Brownian
motion covariance and an Ornstein–Uhlenbeck correlation function, whose positive
definiteness is established rigorously via the Schur product theorem. Building on the
Kimeldorf–Wahba correspondence between Bayesian estimation and regularisation,
we derive an entropy-regularised objective that jointly optimises data fidelity, temporal
smoothness, and divergence from a linear trend prior. We provide complete,
self-contained proofs of existence and uniqueness via the direct method in the
calculus of variations, a representer theorem for a direct-sum RKHS, a closed-form
matrix solution, and minimax-optimal convergence rates derived from the spectral
analysis of the associated integral operator.
A methodological point governs how the empirical results should be read: neither
the Information Flow Smoother (IFS) nor Wahba’s cubic smoothing spline is a
forecasting method. Both are retrospective smoothers whose hat matrix is defined
over the full observed window simultaneously; there is no causal fitting structure.
Empirical comparisons are therefore conducted on three distinct criteria: in-sample
volatility-weighted residual calibration, which is the native operating criterion of the
IFS; regime-stratified fit quality, which assesses performance under heteroskedastic
stress; and an anchoring stability test, which asks whether better heteroskedastic
calibration produces a more stable terminal fitted value when the smoother is
extrapolated linearly into a hold-out window.
We evaluate all three criteria across three train/test splits spanning qualitatively
different market regimes: a full bull-bear-bull cycle (2015–2022 train), the post-
COVID recovery period (2015–2020 train), and the Federal Reserve rate-hiking stress
episode (2015–mid-2022 train), and against two Wahba benchmarks: the classical
uniformly-weighted cubic smoothing spline, and a heteroskedastic variant fitted
with the same precision weights as the IFS, which isolates the marginal contribution
of the Information Flow kernel from that of reweighting alone. The IFS achieves
lower volatility-weighted error than both benchmarks across all three regimes, at its
own GCV-optimal flexibility, at matched effective degrees of freedom, and with the
advantage confirmed as statistically robust to the serial dependence of daily returns
via block-bootstrap confidence intervals. The sensitivity heatmap confirms that the
IFS advantage on vol-weighted criteria is a structural feature of the heteroskedastic
weight design rather than an artefact of any particular train/test partition; this
evidence is drawn from a single equity index and its generalisation to other assets
remains to be established.