Explainable Volterra Networks Bring Transparent Nonlinear Dynamics to Time Series Forecasting

Deep learning has transformed time series forecasting, but a persistent complaint shadows every benchmark table: the best models are black boxes. A recurrent network or Transformer may nail next week’s electricity load while offering almost no insight into which past observations mattered, how long their effects persisted, or which variables interacted to shape the outcome. A new architecture called Explainable Volterra Networks, or XVN, now tackles that complaint head-on by rebuilding the forecaster around a classical mathematical object that keeps memory and nonlinear interactions visible in its very indices.

The Volterra series, developed for nonlinear system analysis, extends linear convolution to nonlinear dynamical systems by expressing an output through delayed inputs and their higher-order interactions. In this representation, interaction order, variable identity, and time delay are explicit in the indexing of the coefficients, so memory effects and nonlinear interactions remain directly traceable. The research team behind XVN observed that a forecasting model is simply a map from a finite history of observations to a future value. If the underlying sequence evolution contains memory effects and nonlinear interactions, that map is naturally a nonlinear functional of delayed inputs, and the Volterra series provides a structured polynomial representation of exactly such functionals.

Turning this classical structure into a modern neural forecaster required solving two central challenges. The first was structural embedding: the classical Volterra series represents nonlinear systems with memory through indexed kernels, and a neural implementation had to discretize and train this structure while preserving the semantic correspondence between coefficients and variable-delay interactions. The second was redundancy control: explicit high-order interactions generate symmetry-induced redundancy in feature construction and parameter redundancy in high-order coefficient tensors. XVN addresses both through two components, Discrete Volterra Reparameterization and Redundancy-Aware Sparsification, wrapped in a multi-channel architecture whose channel weights and linear output projection remain explicit learnable parameters.

Discrete Volterra Reparameterization converts the classical integral series into learnable Discrete Volterra coefficient tensors. The team recognized that the discretized nth-order summation is precisely the Frobenius inner product between a kernel tensor and the nth Kronecker power of the flattened input window. XVN therefore constructs explicit interaction features from the input window and learns coefficient tensors whose entries remain indexed by specific feature interactions at specific time delays. Redundancy-Aware Sparsification then removes redundancy in both the feature space and the parameter space: a fixed mask retains only monomials with non-decreasing index order, eliminating permutation-equivalent terms, while a low-rank coefficient parameterization with magnitude-based core sparsification keeps the surviving high-order coefficients compact, trainable, and inspectable.

Interpretation flows from what the authors call effective coefficients. Because the channel weighting and output projection are purely linear, the final prediction remains an exact polynomial function of the input, and the effective coefficient of any interaction can be computed exactly by direct parameter inspection, aggregating realized coefficients, channel weights, and the output projection through the same linear aggregation used in the forward computation. From these effective coefficients, the framework extracts three dimensions of analysis: feature importance scores for each output component, temporal memory patterns indexed by feature and time delay, and pairwise or higher-order cross-variable interaction effects. Unlike post-hoc attribution methods such as SHAP, LIME, and Integrated Gradients, which explain trained predictors through secondary approximations and have been shown in benchmark studies to produce unstable temporal attributions, these quantities are derived from the model’s actual parameters.

The controlled validation is striking. When trained on the Lorenz 63 chaotic system, whose governing equations are known, XVN recovered the dynamical parameters sigma, rho, and beta within a maximum error below ten to the minus three. On a controlled sparse delayed Volterra system with twelve planted interactions, XVN achieved a combined Top-k hit rate of 0.89, an F1 of 0.82, a Spearman rank correlation of 0.87, and sign accuracy of 0.90, with first-order recovery strongest at values above 0.9. Compared on the same system, TSMixer with SHAP reached a combined F1 of only 0.49 and GRU with Integrated Gradients 0.47, while SINDy with a matched delayed polynomial library achieved the strongest recovery at 0.86, expected for a system-identification method whose candidate library aligns with the data-generating process. Local perturbation tests further confirmed coefficient semantics: sign agreement for first-order coefficients reached 0.93, second-order interaction agreement 0.86, and lag-response consistency 0.89.

Real-domain analysis then showed that the same coefficient machinery produces physically meaningful patterns. For oil temperature prediction on the electrical transformer temperature dataset, the learned feature importance revealed a clear active-reactive pairing, with active loads consistently outweighing corresponding reactive loads at each level, consistent with copper loss dominance in transformer heating. Temporal memory analysis extracted fast electrical time constants of roughly 1.2 to 1.8 hours and an oil temperature constant of 3.2 hours, consistent with the 2 to 4 hour ONAN transformer specifications documented in IEEE C57.91-2011. Second-order coefficients suggested physically meaningful couplings, including an active-reactive heating term consistent with IEC 60076-7. On the Weather dataset, vapor pressure deficit showed shorter temporal memory than potential temperature, matching the meteorological principle that vapor pressure responds more rapidly than thermal inertia allows temperature to respond. On Exchange, learned currency importance broadly followed the BIS 2013 market-share hierarchy with an average relative error of 8.2 percent, and cross-seed stability tests showed the dominant variables, temporal profiles, and top pairwise interactions were reproducible across independent training runs.

Crucially, XVN remains competitive as a forecaster. Across twelve chaotic-system configurations, including Double Pendulum, Lorenz 63, and N-body, XVN achieved the lowest mean squared error in every setting. On Lorenz 63 at horizon 12, its error of 3.05 times ten to the minus four was roughly four times better than the second-best model GRU and vastly better than Informer. On the five real-world benchmarks, XVN ranked lowest or tied-lowest in 18 of 20 configurations, with Weather improvements of 6.0 to 13.4 percent over the next-best model at all four horizons. The two exceptions were ETT at short horizons 96 and 192, where SparseTSF’s cross-period sparse structure captured simple periodicity more efficiently; XVN surpassed all baselines at the longer ETT horizons of 336 and 720, where accumulated cross-variable coupling and nonlinear response become more influential. Because XVN trains a single-step predictor applied autoregressively, its parameter count stays fixed at 37.4K across all prediction horizons, whereas representative Transformer baselines require roughly 300 to 500 times more parameters. A direct multi-step variant, MIMO-XVN, reduced error at the longest horizons but expanded its output head with the horizon, reaching 1.10M parameters on ETTh1 and 9.64M on Weather at horizon 720.

Ablation studies confirmed that both redundancy-reduction mechanisms matter. Removing all masks caused the largest degradation on ETTh1, with mean squared error increases of 46.6 to 68.1 percent, while removing the low-rank module and individual mask components produced intermediate degradations. Rank sensitivity analysis placed the capacity plateau at R=16, where coefficient reconstruction error fell to 0.052 and support F1 and sign accuracy rose to 0.82 and 0.90, with larger ranks adding resource use but no prediction gain. Polynomial order selection followed a validation-based rule: Lorenz 63 and the sparse delayed system favored second order, a third-order synthetic system favored cubic terms, and Weather showed similar error across orders. The authors also mapped the method’s boundaries through stress tests. Under a regime shift, support F1 fell from 0.82 to 0.56 and sign accuracy from 0.90 to 0.68, because a single static coefficient set summarizes multiple regimes. Non-polynomial mismatch reduced support F1 to 0.62, and halving the training data widened training-validation gaps most sharply for third-order models, supporting early stopping and the smallest stable polynomial order when training windows are limited.

The broader message is that understanding and accuracy need not be traded off. XVN demonstrates that a trainable, scalable neural forecaster can carry its dynamic explanation inside its own parameters, with feature importance, temporal memory, and cross-variable interactions traceable through effective coefficients rather than reconstructed after the fact. The authors are careful about the strength of the evidence: the experiments support coefficient-level interpretation and selected forms of mechanistic consistency, while causal claims in real-world systems would require intervention or causal-design studies. Future work points toward state-conditioned channel gates, piecewise Volterra or mixture-of-experts models, non-polynomial basis functions, and stronger validation protocols for real-domain interpretation. For scientific, engineering, and financial applications where users need to know which past observations matter, how long their effects persist, and which variables interact, XVN offers a template for forecasters that show their reasoning in their coefficients.

Subject of Research: Machine Learning and Artificial Intelligence

Article Title: Explainable Volterra networks: Structured nonlinear dynamics for time series forecasting

Article References: Huang, Y., Jiang, Y., Li, T., Fan, J., Wang, M., Li, H., & Jiang, H. (2026). Explainable Volterra networks: Structured nonlinear dynamics for time series forecasting. Results in Engineering, 32, Article 113194. https://doi.org/10.1016/j.rineng.2026.113194

Image Credits: AI Generated

DOI: 10.1016/j.rineng.2026.113194

Keywords: time series forecasting, explainable AI, Volterra series, nonlinear dynamics, neural networks, interpretability, chaotic systems, machine learning, Explainable, Volterra, networks, Structured