Abstract

Classical nonparametric smoothing, and in particular Wahba's cubic smoothing spline framework, applies a symmetric curvature penalty that treats every observation identically, regardless of local information content. This is structurally mismatched with financial time series, which are heteroskedastic, non-stationary, and causally directed. To address both shortcomings, we introduce the Information Flow Smoother (IFS), an entropy-regularised estimator constructed within a reproducing kernel Hilbert space.

The IFS is built around the Information Flow Kernel, formed as the pointwise product of the Brownian motion covariance, which captures information accumulation over time, and an Ornstein-Uhlenbeck correlation function imposing a decaying memory horizon. Positive definiteness follows from the Schur product theorem. A key structural property is causal monotonicity: more recent observations exert strictly greater influence on the estimated curve, consistent with efficient market dynamics.

The estimator minimises an objective combining three components: a heteroskedastic goodness-of-fit term with residuals weighted by inverse exponentially weighted moving average variances; a Kullback-Leibler penalty shrinking the trend toward a linear prior via the Kimeldorf-Wahba correspondence; and a roughness regularisation term within the Information Flow kernel space. Both regularisation parameters are selected by generalised cross-validation. A representer theorem reduces the problem to a tractable finite linear system, and spectral analysis confirms eigenvalue decay at a quadratic rate, yielding minimax-optimal mean squared error convergence.

We evaluate the IFS against two Wahba spline baselines on daily S and P 500 prices from January 2017 to December 2024, spanning three volatility regimes: the pre-pandemic bull market, the March 2020 crash, and the 2022 Federal Reserve rate-hiking cycle. On the volatility-weighted mean squared error criterion, the IFS outperforms the cross-validation-optimal Wahba spline by a factor of two and the degrees-of-freedom matched spline by a factor exceeding thirty. Residual calibration is substantially better across all regimes, and regime-stratified analysis confirms IFS residuals remain statistically homogeneous in both calm and turbulent periods. Rolling out-of-sample evaluation further confirms that the in-sample advantage of the near-interpolating Wahba spline does not generalise, while the IFS does.
These results demonstrate that embedding causal structure and heteroskedastic precision weights directly into the kernel and objective function produces material improvements in signal extraction quality for financial time series.