Many financial and physical series are well described by Lévy processes: random walks that combine a smooth drift, Gaussian noise, and jumps of every size. They are usually estimated in the Fourier domain, because the transition density has no closed form but the characteristic function (the Fourier transform of the increment distribution) does. Every such estimator assumes the process never changes. Real series do change, and the history arrives unevenly. Old eras may be sparse and recent ones dense, or a recent regime may be short while the long archive behind it is stale.
The usual fix forgets old data at one rate, through a sliding window or an exponential discount, for all features at once. This proposal argues that one rate is the wrong unit of forgetting. Each Fourier frequency of the characteristic function describes a different part of the process. Low frequencies are dominated by drift, variance, and large jumps, which regimes rewrite quickly. High frequencies are dominated by the swarm of tiny jumps, which may be structurally stable across decades. We propose to estimate the current process by denoising, separately at each frequency, the time series of local Fourier estimates across eras. That lets each frequency choose how much history it trusts. The resulting "now" estimate is then inverted with the existing regularized Fourier machinery to recover today's jump distribution.
Core Claim
Split the record into blocks $k=1,\dots,K$, where block $k$ holds $n_k$ increments; under temporal imbalance the $n_k$ differ wildly. For each block, form the empirical characteristic function $\hat\phi_k(u)=n_k^{-1}\sum_{j\in k} e^{iu\Delta X_j}$ at frequency $u$. Its log, divided by the sampling interval, $\hat\psi_k(u)=\Delta^{-1}\log\hat\phi_k(u)$, estimates that block's characteristic exponent (the function that, through the Lévy–Khintchine formula, encodes drift, volatility, and jump measure).
For fixed $u$, the sequence $\hat\psi_1(u),\dots,\hat\psi_K(u)$ is a noisy, possibly non-stationary curve over time. Its noise level is known up to plug-in, because the variance of $\hat\phi_k(u)$ is $(1-|\phi_k(u)|^2)/n_k$. We estimate the current exponent as
$$\tilde\psi_K(u)=\mathcal{W}^{-1}\,\mathrm{ST}_{\tau(u)}\big(\mathcal{W}[\hat\psi_{1:K}(u)]\big)_K,$$
where $\mathcal{W}$ is a discrete wavelet transform over the block index, $\mathrm{ST}_{\tau(u)}$ soft-thresholds coefficients with a frequency-specific threshold $\tau(u)$ scaled to the heteroscedastic block variances, and the subscript $K$ takes the final ("now") value. The Lévy density at time $K$ is then recovered by plugging $\tilde\psi_K$ into the regularized inverse Fourier step of Belomestny, Orlova and Panov. Its spectral cut-off $U$ is chosen from the effective sample size implied by the thresholding at each $u$, not from the raw $n$.
The new claim has two parts, and both are testable. First, a single exponential discount or window is strictly suboptimal for estimating today's jump distribution whenever different frequency bands drift at different rates. Frequency-wise denoising recovers the oracle per-band window without knowing the drift rates. Second, the transform domain makes the guarantees available for heavy-tailed jumps. Each $\hat\phi_k(u)$ is an average of bounded unit-modulus variables, so it is sub-Gaussian even when the jumps themselves have infinite variance. That is exactly the noise condition the wavelet "now" estimator requires.
Relation to Prior Work
The method is a composite of two existing tools, and the novelty lies in where they meet.
The temporal mechanism comes from Adaptive Estimation and Learning under Temporal Distribution Shift. That paper shows non-stationarity is sparse in a wavelet basis and gets minimax "now" estimates of a drifting mean by soft-thresholding, outperforming adaptive sliding windows on Dubai housing prices. It works with one scalar mean, uniform time steps, and sub-Gaussian noise. We apply it to a continuum of Fourier frequencies, allow unequal block sizes, and inherit sub-Gaussianity from boundedness instead of assuming it.
The inversion machinery comes from three sources. Belomestny's composite characteristic function estimator recovers Lévy densities under a random business clock. The moving-average Lévy estimator achieves $(\log n)^{-(s+k)/2}$ rates via spectral low-pass inversion. Burnaev's inversion formula handles the singularity of the jump measure at zero. All three hold the triplet fixed in time; we keep their inversion step and replace its input.
The testing tool comes from Chen, Peng and Yu. They fit Markov jump models by empirical likelihood on conditional characteristic functions and find that 1965–1999 Treasury bill rates reject diffusions in favour of infinite-activity inverse-Gaussian Ornstein–Uhlenbeck dynamics, weighting all 34 years equally. Their result motivates us, because it says small-jump structure matters for real rates. It also provides our test for whether the jump measure itself has shifted.
Two applied systems handle imbalance and staleness together, but neither has any spectral content, nor do they forget at a component level. The Historical Data Reuse system for promotion-time conversion prediction retrieves similar past promotions and importance-weights them. C3DFusion blurs older camera frames and densifies the current one. Both decide relevance per sample or per frame. We decide it per frequency, which is the distinction this proposal tests.
Research Program
The theory step bounds the error of $\tilde\psi_K(u)$ uniformly over a frequency band. It extends the wavelet result to heteroscedastic noise with known variance ratios $1/n_k$. It must also handle the logarithm: where $|\phi_k(u)|$ is small, the log is unstable, so a lower bound on $|\phi|$ over the band is required, and that bound restricts the usable $U$. We then propagate this bound through the regularized inversion to get a rate for today's Lévy density. The rate depends on the per-band variation of the exponent's path rather than on $n$. The headline result we aim for is a regime where small-jump features attain near-full-sample rates while drift attains short-window bias.
The construction step implements the estimator with Haar and DB8 bases and a fixed grid of frequencies. It adds a cut-off rule that picks $U$ from estimated per-frequency effective sample sizes. It also adds a weighted variant of the conditional characteristic function empirical likelihood, using our per-frequency temporal weights, so the method can test whether the jump measure has changed between eras.
The analysis step asks, on real data, whether frequency bands actually drift at different rates. If they do not, the method collapses to a single discount, and that is a reportable negative finding.
Baselines
The naive reference is the pooled estimator: the moving-average Lévy Fourier procedure applied to all history with equal weights, and the conditional characteristic function empirical likelihood estimator likewise unweighted. It is what current theory prescribes, and it shows the cost of ignoring drift.
The strongest practical competitor is the same inversion fed by an exponentially discounted characteristic function, with the half-life tuned by rolling-origin validation on the same data, together with a tuned sliding window. It is strongest because, with a good half-life, it already captures most of a single-rate drift. It receives the same validation budget as our threshold constant.
The strongest existing method is the scalar wavelet "now" estimator of the temporal shift paper. We apply it to the block time series of a few fitted parameters (drift, volatility, jump intensity, activity index) and then reconstruct the exponent parametrically. It shares our temporal machinery but forgets in parameter space rather than frequency space. No reported result on Lévy-density or exponent recovery was found on Lacuna for any of these baselines, and the temporal shift paper reports no numerical errors for Dubai housing. Every number will therefore come from our own runs.
The key ablation is our estimator with one threshold shared across all frequencies. It isolates whether the per-frequency choice, not wavelet denoising itself, drives any gain. A second ablation drops the heteroscedastic $1/n_k$ scaling, to measure the contribution of the imbalance correction.
Beating the competitors means lower integrated squared error of today's exponent and Lévy density on synthetic data, and better one-step-ahead scores on real data, at matched tuning. The improvement over the tuned exponential discount must be statistically significant across seeds or rolling origins.
Why not just…?
Why not just tune an exponential discount on the characteristic function? This is the real threat. If all parts of the process drift at one rate, a well-tuned half-life is essentially optimal and our method adds variance through extra thresholds. We do not argue around this. The synthetic suite includes a homogeneous-drift condition in which we predict no gain. The real-data analysis reports the estimated per-band variation directly, so readers can see whether heterogeneity exists. If it does not, the conclusion is that one discount suffices for Lévy inference, which is still a useful negative result.
Why not detect change points and estimate on the last segment? Hard segmentation throws away the long, possibly stable small-jump history at every break. By the $(\log n)^{-(s+k)/2}$ rate of the moving-average estimator, shrinking $n$ is costly precisely for the fine jump structure. We include segmentation followed by pooled inversion as a baseline. It should lose on the high-frequency part of the error and roughly tie on drift.
Why not fit a neural jump-diffusion in the time domain? The neural jump-diffusion point process models finitely many discrete jumps with fixed weights. The rate data in Chen, Peng and Yu favour infinite-activity jumps, which that model cannot represent, and its fixed weights do not address drift at all. We consider it out of scope for the main comparison and say so, rather than run an unequal contest.
Proposed Experiments
The synthetic study simulates tempered-stable and compound-Poisson-plus-diffusion processes whose drift, volatility, jump intensity, and small-jump activity index drift on separately controlled schedules. Schedules include smooth trends, one abrupt break, and identical rates. Block sizes follow either a growing profile (sparse past, dense present) or a shrinking one (long archive, short new regime). The metrics are integrated squared error of $\psi_K$ over the frequency band and $L_2$ error of today's Lévy density against ground truth, over 200 seeds.
The decisive check is the frequency-sliced error. Our method should match the full-sample error on high frequencies and the short-window error on low frequencies only when the schedules differ. Its advantage over the tuned discount should vanish when they coincide.
The real-data study uses the 1965–1999 three-month Treasury bill series from Chen, Peng and Yu. It has a known structural episode, the 1979–1982 monetary experiment, against which we can check where the estimated exponent moves. We create imbalance by thinning early decades to coarser sampling. Evaluation is rolling-origin: score the predicted next-block characteristic function by its integrated squared distance to that block's empirical characteristic function, and score next-block increments by log predictive density via inversion. We also run the weighted empirical likelihood test of whether the jump measure, as opposed to only drift and volatility, differs between pre- and post-1979 regimes.
As a second real setting, we apply the same protocol to daily CSI 300 constituent returns from CSMD (2021–2024). There the recent window is short relative to useful history, which directly tests the "stale archive, short new regime" profile.
The proposal fails if, at matched tuning, the shared-threshold ablation or the tuned exponential discount equals our method on the real series. It also fails if the per-band variation estimated from data shows no heterogeneity. In that case we would report that frequency-wise forgetting is unnecessary for these markets.
References
- Statistical inference for moving‐average Lévy‐driven processes: Fourier‐based approach
- Adaptive Estimation and Learning under Temporal Distribution Shift
- Statistical inference for time-changed Lévy processes via composite characteristic function estimation
- Inversion formula for infinitely divisible distributions
- Parameter estimation and model testing for Markov processes via conditional characteristic functions
- Capturing Conversion Rate Fluctuation during Sales Promotions: A Novel Historical Data Reuse Approach
- Towards Temporal Fusion Beyond the Field of View for Camera-based Semantic Scene Completion
- Neural Jump-Diffusion Temporal Point Processes
- CSMD: Curated Multimodal Dataset for Chinese Stock Analysis