Tommaso Proietti | Tor Vergata
Trend Filtering with Fractional Splines
Wednesday, November 4, 2026 h. 13:00-14:00
Sala del Consiglio
Abstract
The talk introduces a novel signal plus noise model, where the signal is a fractional spline (FS) process. The latter is generated by a linear combination of shocks, where the impulse response function follows a generalized harmonic sequence, governed by the memory parameter. The fractional spline process encompasses standard trend models, such as the random walk and the integrated random walk, while for fractional memory parameter it provides an alternative to standard fractional noise. We discuss two specifications, the fixed effects model, which generalizes trend filtering to the fractional case, and
the random effects specification, and develop estimation and signal extraction by penalized least squares with l_1 and l_2 penalties in the former case, and by maximum likelihood estimation and conditional mean inference for the latter. We then characterize the time series properties of the fractional noise process in terms of memory, and, in the stationarity case, we derive the autocovariance function and spectral density, and compare them to those of the
fractional noise process. The practical relevance of the FS framework is demonstrated through applications to inflation dynamics and realized volatility, illustrating its forecast performance and its ability to capture complex long-memory patterns in economic data, as well as persistent trends in tree-ring based temperature reconstructions.

