Interpretable time-series models such as neural additive autoregression predict the next step as a sum of small learned functions, one per input variable. Their known weakness is that they cannot represent effects where two variables act jointly, such as the product of velocity and its gradient that drives fluid advection. The obvious fix is to add a few pairwise terms. We argue this fix has a hidden cost, and that the cost is also the opportunity. A pairwise term is trained only on the joint values its two inputs actually took in the data. Once a long forecast starts to drift, each variable can stay inside its own normal range while the pair moves into combinations never seen in training. There the pairwise function extrapolates freely, and its error is fed back as input on the next step. Our hypothesis is that in glass-box forecasters, the error that compounds over a rollout comes mostly from a few interaction terms leaving their joint training support. Because the model is an exact sum of terms, this can be measured exactly rather than estimated after the fact, and it can be fixed by watching only those terms.
The proposal builds a sparse-interaction autoregressive forecaster. It adds an exact "drift ledger" that splits error growth at every rollout step across the model's terms. It also adds a cheap stabilizer that checks each pairwise term's two-dimensional input against its training support and pulls it back when it strays. If the hypothesis holds, the trade between accuracy and interpretability in sequential forecasting becomes a design choice we can manage: interactions buy accuracy, a ledger shows where they cost stability, and a low-dimensional check pays that cost. The check is far cheaper than the full-state diffusion denoisers used to stabilize black-box emulators.
Where the literature stops
Three findings motivate this work. Neural Additive Vector Autoregression fits one network per variable's lag history and sums them to predict the next step. It explicitly leaves cross-variable interactions out and never measures multi-step rollout error. Sparse Interaction Additive Networks (SIAN) show on static tabular data that a few selected interaction sub-networks close the accuracy gap to black-box networks. Its average rank is 3.00 against 3.86 for a plain network and 3.71 for NODE-GAM, and a few 2D interactions explain 84% of variance on Bike Sharing. SIAN never feeds its outputs back as inputs. Sum-of-Parts proves that per-feature additive explanations incur error that grows exponentially with dimension. It needs about 20 backbone passes per input, which is prohibitive inside a long rollout. We reuse SIAN's interaction-selection procedure as our method. We take NAVAR as the model we extend. We treat Sum-of-Parts as a constraint that rules out group-masking designs at rollout time.
On the drift side, ProbTS documents that autoregressive decoders accumulate error over 96-step horizons. Non-autoregressive decoders avoid this but smooth away high-frequency detail. Thermalizer stabilizes chaotic emulators by detecting drift with a scalar noise classifier and denoising the full state with a diffusion model. Both treat drift as an undifferentiated quantity with no account of what caused it. The interaction-attribution tools SHAP-IQ, SVARM-IQ and Integrated Hessians estimate pairwise effects post hoc on single static inputs. They need many model evaluations, and in the Hessian case they need ReLUs swapped for SoftPlus. Applying them along a 1000-step trajectory is impractical. More importantly, they explain the model's output, not how a perturbation grows. The closest bridging work is the time-series concept bottleneck, which aligns transformer layers with a linear autoregressive surrogate and a timestamp concept. Its concepts are independent and it does not study multi-step error. Stochastic Concept Bottleneck Models model dependencies between concepts through a joint covariance, but only on static image and tabular classification.
Core Claim
Consider a forecaster whose one-step update for variable $k$ is
$$\hat x^{k}_{t+1} = \hat x^{k}_{t} + \sum_{j} f_{kj}(h^{j}_t) + \sum_{(i,j)\in S_k} g_{kij}(h^{i}_t, h^{j}_t).$$
Here $h^{j}_t$ is the recent lag window of variable $j$. The $f_{kj}$ are single-variable networks. The $g_{kij}$ are pairwise networks over a sparse set $S_k$ chosen by SIAN-style heredity search on a black-box teacher.
The one-step Jacobian $J_t$ (how a small change in the current state changes the next state) is then exactly the identity plus a sum of per-term Jacobians $J^{(m)}_t$. For a rollout error $e_t$, the linearized one-step growth rate $e_t^\top J_t e_t/\|e_t\|^2$ splits exactly into per-term shares $a^{(m)}_t = e_t^\top J^{(m)}_t e_t/\|e_t\|^2$. These shares sum to the total with no sampling and no approximation beyond linearization. This completeness is the same property Integrated Hessians obtains for output explanations. Here it applies to error growth, and it comes free because the model is a sum.
The claim has three falsifiable parts. First, pairwise terms improve short-horizon accuracy over NAVAR. Second, without intervention, the cumulative shares $\sum_t a^{(m)}_t$ concentrate on a small number of pairwise terms. The two-dimensional inputs of those terms leave their training support measurably earlier than any single-variable input leaves its marginal range. Third, stabilizing only those flagged terms recovers long-horizon stability at little short-horizon cost.
Research Program
The first stage builds the forecaster in two settings.
For multivariate series, terms act on per-variable lag windows. The interaction set is chosen by Archipelago-guided heredity search over a trained black-box teacher, as in SIAN, and the sub-networks are run with block-sparse layers.
For spatiotemporal PDEs, the same template is applied as a shared local stencil. The variables are a point's neighborhood values across physical channels, following the separate-channel idea that PDE-Transformer found improves fine-tuning by 2.7x to 4.4x. Here channel coupling is written as explicit, plottable pairwise shape functions rather than buried in attention. This setting is attractive because the physics is known. In Burgers' equation the nonlinearity is literally a pairwise product, so we can check whether the selected terms match the true coupling.
The second stage is the drift ledger. For every rollout step we record each term's growth share $a^{(m)}_t$. For each pairwise term we also record a support score: the log-density of its current two-dimensional input under a kernel density estimate fit on training pairs. Density estimation in two dimensions is easy and reliable, which full-state density estimation is not. The ledger yields a ranked, human-readable account of which coupling drove each divergence and when it left familiar territory.
The third stage is the stabilizer. When a flagged pairwise term's input falls below a support threshold, it is projected to the nearest high-density training point before the term is evaluated. During training, a Lipschitz penalty is applied only to terms the ledger flags on validation rollouts. Thermalizer must hand-tune its denoising thresholds $s_{init}$ and $s_{stop}$. Our single threshold is set by a quantile of training density, and every intervention is logged against a named term.
Baselines
The naive reference is persistence, which repeats the last observed value, together with the plain NAVAR model with no interactions trained on identical data. NAVAR is the do-nothing version of our idea. If adding interactions does not beat it at short horizons, the first claim fails. If NAVAR outperforms our stabilized model at long horizons, the interactions are not worth their cost.
The strongest black-box competitor on multivariate forecasting is an autoregressive transformer decoder with Reversible Instance Normalization, as configured in ProbTS. That paper reports RevIN "significantly reduced" autoregressive error accumulation but gives no number in the file. No reported result found on Lacuna for ProbTS MSE or CRPS at horizon 96, so we will reproduce it in the ProbTS codebase. We will also run its best non-autoregressive decoder, since that is the reviewer's default answer to drift.
On chaotic PDEs the strongest stabilizer is Thermalizer. It reports that standard emulators of Kolmogorov flow reach non-physical states by step 5000 and that thermalization extends stable horizons by up to two orders of magnitude, at a cost of roughly 60% of one emulator step. For one-step accuracy on PDEs, the reference is PDE-Transformer, which reports a 42% average nRMSE improvement over Fourier Neural Operators. We run it from the official repository.
All methods get the same training trajectories and the same hyperparameter-search budget in GPU-hours. All are scored on identical held-out initial conditions. Black-box baselines additionally receive noise-injection training, adding Gaussian noise to inputs during training, which is the cheapest generic drift remedy.
Three ablations isolate the key ingredient. The first applies the same Lipschitz penalty and support projection to all terms rather than only flagged ones. The second uses a marginal-range clip on single variables instead of the joint support check. The third uses the ledger with random term selection.
We count it as beating the competition if the stabilized model matches or exceeds the RevIN transformer's horizon-96 MSE and CRPS on at least half the ProbTS datasets. On Kolmogorov flow, it must reach a stable horizon within a factor of two of Thermalizer's at lower per-step cost. In both cases it must clearly beat the untargeted ablations.
Why not just…?
Why not just use NAVAR, which is already interpretable and has no fragile interaction terms? Sum-of-Parts proves per-feature additive models lose accuracy exponentially with dimension, and SIAN shows the empirical gap on tabular data. In physics the gap is structural: an additive model cannot represent advection at all. Our first experiment measures this gap directly on Burgers and on the CauseMe nonlinear vector-autoregression data. If NAVAR turns out both stable and within noise of our accuracy there, we will report that interactions are unnecessary in that regime.
Why not just train a black box with noise injection and RevIN? This is a serious threat to the accuracy claim, and we may lose on raw horizon-96 error for high-dimensional datasets such as Traffic. The proposal does not claim the glass-box model beats black boxes everywhere. It claims stability within a stated margin while giving an exact, term-level account of drift that a noise-trained black box cannot provide. The head-to-head comparison above decides how large that margin is.
Why not just run SHAP-IQ or Integrated Hessians on the black-box emulator's rollout? These tools explain outputs at a single input, and each costs many forward passes. Even if we computed Jacobian-based interaction scores along a trajectory, the per-pair split of a dense network's Jacobian is an estimate with no completeness guarantee. We will test this by computing SHAP-IQ pair scores on the teacher at sampled rollout steps. We then check whether stabilizing the pairs it ranks highest works as well as stabilizing the pairs our ledger ranks highest, at matched cost.
Why not just regularize or clip every term uniformly? If uniform treatment works as well, the ledger is merely descriptive and the targeting claim fails. This is exactly the first ablation. We expect uniform Lipschitz penalties to hurt short-horizon accuracy, because they flatten terms that are both accurate and safe.
Proposed Experiments
The study uses four testbeds. The first is CauseMe's nonlinear vector-autoregression and river datasets, where generating structure is partly known. The second is ProbTS multivariate datasets (ETT, Electricity, Weather, Exchange) with autoregressive rollouts to 96 steps. The third is PDEBench 1D Burgers and 2D reaction-diffusion, scored with RMSE plus PDEBench's conservation and Fourier-spectrum errors. The fourth, a stretch test, is Kolmogorov flow, scored on stable horizon and kinetic-energy spectrum against Thermalizer. We are explicit that a stencil glass-box model may not reach U-Net one-step accuracy there.
The decisive checks follow the three claims. The first check is whether interactions beat NAVAR at one to eight steps. The second is whether, on unstabilized rollouts, the top three pairwise terms account for most of the cumulative growth share, and whether their support score drops before any marginal leaves its training range. We will plot the lead time between these two events. On Burgers we also check whether the flagged term is the true advection coupling. The third is whether targeted stabilization closes the long-horizon gap to the black-box and Thermalizer baselines while the uniform and marginal-clip ablations do not.
If the second check fails and drift is spread across many terms, we will still report the ledger as an exact diagnostic, but the support-gating stabilizer will be dropped as a contribution.
References
- Neural Additive Vector Autoregression Models for Causal Discovery in Time Series
- Sparse Interaction Additive Networks via Feature Interaction Detection and Sparse Selection
- Sum-of-Parts: Self-Attributing Neural Networks with End-to-End Learning of Feature Groups
- ProbTS: Benchmarking Point and Distributional Forecasting across Diverse Prediction Horizons
- Thermalizer: Stable autoregressive neural emulation of spatiotemporal chaos
- SHAP-IQ: Unified Approximation of any-order Shapley Interactions
- SVARM-IQ: Efficient Approximation of Any-order Shapley Interactions through Stratification
- Explaining Explanations: Axiomatic Feature Interactions for Deep Networks
- Interpretability for Time Series Transformers using A Concept Bottleneck Framework
- Stochastic Concept Bottleneck Models
- PDE-Transformer: Efficient and Versatile Transformers for Physics Simulations