Where the Graph Matters: Tail-Stratified Recalibration for Dynamic Spatio-Temporal Forecasters

Traffic and air-quality forecasters increasingly build a new graph of sensor-to-sensor connections at every time step, called a dynamic graph topology, on the theory that relationships between roads or cities change with conditions. Yet a careful audit found that on standard benchmarks most of the predictive signal comes from each sensor's own history, and that the extra accuracy from spatial message passing is modest on average. We think the average is the wrong place to look. Most time steps are ordinary, such as free-flowing traffic or the regular rush hour, and a seasonal model of the sensor's own past already predicts them well. The rare time steps where the forecast is far from that seasonal expectation, such as an accident, a bottleneck, or a pollution episode, are the ones that need information from neighbors. They are also the ones that standard mean-error training ignores, because they are few. If this is right, dynamic graphs face a hidden head-tail trade-off: they are trained and judged on the common cases ("head"), while their value lies in the rare ones ("tail").

This note proposes to measure that trade-off directly and then to break it cheaply. The method separates what the dynamic graph encoder knows from how the final output layer uses it. We first test whether an ordinarily trained encoder already contains the information needed to forecast large deviations. If it does, freezing it and retraining only a small output layer on tail-balanced data should improve forecasts of rare events without hurting ordinary ones. Vision work found the same pattern for rare object classes. If it does not, the failure tells us the dynamic graph itself must change. Either answer is new: no paper in our file reports dynamic-graph forecasting accuracy broken down by how unusual the target is.

Core Claim

We make two linked, falsifiable claims. The first is diagnostic. Start from the residual, meaning the gap between the target and a seasonal per-sensor baseline. On that residual, the error reduction from a dynamic-graph model over a temporal-only model should grow monotonically with the size of the residual. The modest average spatial gain reported by A Critical Audit of Spatiotemporal Forecasting Benchmark Datasets and Baselines would then be a small number of large tail gains diluted by many near-zero head gains.

The second claim is about mechanism. Under ordinary mean-error training, the tail failure sits mostly in the output layer rather than in the dynamic graph encoder. A frozen encoder plus a recalibrated output layer should therefore move the head-tail frontier outward, not just slide along it. This mirrors the finding of The Devil Is in Classification that rare-class failure on LVIS lives in the classifier and not in the shared features. However, graphs give a specific reason to doubt the transfer. Tackling Long-Tailed Distribution Issue in GNNs via Normalization shows that message passing over imbalanced graphs can collapse the features of poorly connected nodes, and a recalibrated head cannot recover information the encoder has discarded. The diagnostic step below decides which of these two pictures holds for dynamic graphs.

Research Program

The base model follows the decomposition that the audit found strongest. A seasonal autoregressive model is fit to each sensor and absorbs recurring temporal structure, avoiding the first-differencing that the audit shows amplifies noise. A dynamic graph network then forecasts the residual. We use two encoders, chosen to separate "dynamic graph" from "any graph." The first is the audit's residual GCRN-GRU, a graph-convolutional recurrent network. The second is D2STGNN, which builds a time-varying adjacency from attention and splits each signal into a part that spreads across the network and a part that is purely local. Its local branch gives us a built-in temporal-only control.

We define head and tail without labels. For each sensor, residual magnitudes on the training split are cut at fixed quantiles into many-shot, medium-shot, and few-shot bins. These are analogous to the bins used in deep imbalanced regression, where CCC-GRPO documents "regression to the mean": rare continuous targets get shrunk toward the average.

The program has three steps. The first is stratified measurement. We report per-bin mean absolute error for SARIMA alone, for SARIMA plus a temporal-only residual network, and for SARIMA plus each dynamic graph encoder, at every horizon. This tests the first claim.

The second step is an oracle probe in the spirit of SimCal's ground-truth-proposal test. We freeze the ordinarily trained encoder and fit a ridge regression from its hidden states to the true residual, evaluating only on the few-shot bin. If the probe sharply lowers few-shot error, the information is present and the output layer is to blame. If the probe barely beats the trained output layer, the representation has collapsed. In that case we measure the per-node feature spread on tail time steps, which is the "node-wise standard deviation" statistic that ResNorm uses, to test whether collapse tracks the moments when a node's dynamic neighborhood thins out.

The third step is recalibration. We keep the ordinary output layer and train a second output layer on the frozen encoder. This layer sees training windows sampled uniformly across residual bins and is trained with a differentiable batch-level concordance-correlation term plus absolute error. The concordance term penalizes mismatch in mean, spread, and correlation over a batch. We borrow it from CCC-GRPO but optimize it directly by gradient instead of reinforcement learning, which avoids their reported sixfold training cost. A small gate, fit on held-out data, predicts from the encoder state the probability that the coming residual is in the tail and blends the two output layers.

Balancing the Scales proves that tail-favoring adjustments are inconsistent with minimizing plain overall error. We take this as a constraint rather than an obstacle. We never claim one operating point wins every metric. Instead we sweep the gate threshold and report the full curve of head error against tail error.

If the probe shows collapse, the program switches to changing the encoder. The test then becomes whether degree-aware rescaling of features, applied per time step to the dynamic adjacency, restores the tail. ResNorm does this only for a fixed graph.

Baselines

The naive reference is SARIMA alone, together with the ordinarily trained residual model, reported per bin. The strongest existing number in the file is the audit's residual GCRN-GRU at an overall MAE of 3.32 on METR-LA (A Critical Audit). We will reproduce it, and if our reproduction differs we will report both. No tail-stratified result for any method was found on Lacuna, so the per-bin numbers are new measurements.

D2STGNN is reported to improve one-hour METR-LA MAE by 2.6% over prior methods. TESTAM is the most direct competitor. It already routes each road among a temporal-only expert, a static-graph expert, and a dynamic-attention expert aimed at non-recurring incidents. Its routing is trained against overall error and was never evaluated for a seesaw. No numerical score for TESTAM is quoted in the file, so we rerun it.

The strongest loss-based competitor is end-to-end inverse-density reweighting of the residual loss. Each window is weighted by the inverse frequency of its residual bin, and the whole network is trained from scratch.

All methods use identical splits, the same SARIMA baseline, the same encoder size, and an equal hyperparameter search budget of 30 trials each, selected on validation overall MAE plus validation few-shot MAE. Every method is run with three seeds.

The ablations isolate the key ingredient. Recalibration without the gate isolates the effect of the gate. The gate and second output layer trained with plain absolute error, without the concordance term, isolate the loss. Recalibration on a temporal-only encoder tests whether the dynamic graph is what carries the tail information.

Beating the baselines means the following. Few-shot-bin MAE must fall by at least 10% relative to the ordinarily trained model. Many-shot MAE must stay within 1% of it, with overall MAE no worse than 3.32 on METR-LA. Both conditions must hold across seeds and for both encoders.

Why not just…?

Why not just use SARIMA or a temporal-only model, since the audit says time dominates? The audit itself shows that a graph model on SARIMA residuals sets the best METR-LA score. Our first step tests the stronger version of this objection. If the dynamic-graph advantage is flat across residual bins, or absent in the few-shot bin, then dynamic topology is not where tail accuracy comes from, and we would report that as the main result.

Why not just reweight the loss by residual rarity, end to end? This is the real threat. In classification, scaling up tail gradients is known to produce the seesaw: tail gains bought with head losses, as described in Feature Fusion from Head to Tail. It is also inconsistent for overall error per Balancing the Scales. Neither result has been checked for spatio-temporal regression, however. The reweighting baseline is therefore run with the full tuning budget. If its head-tail curve matches ours, the decoupling is unnecessary and we will say so.

Why not just use TESTAM, whose dynamic-attention expert already targets non-recurring incidents? TESTAM trains its experts and router against overall error. Frequent small fluctuations therefore dominate the gradients reaching the attention expert, and the paper reports no per-tail-bin errors. We rerun TESTAM under the same stratified evaluation. We also apply our recalibration on top of its frozen experts, which tests whether the two approaches complement each other rather than substitute for each other.

Proposed Experiments

The datasets are METR-LA and PEMS-BAY, where the audit's findings apply directly, and EXPY-TKY, whose 1,843 links and non-recurring events stress the tail. KnowAir is included to test whether the pattern holds beyond traffic, using pollution episodes. Horizons are 3, 6, and 12 steps.

There are three decisive checks. First, the dynamic-minus-temporal error gap must increase across residual bins. If it does not, the first claim fails. Second, the frozen-encoder probe must recover most of the gap between the ordinary output layer and the best end-to-end tail-tuned model on the few-shot bin. If it does not, the second claim fails and the program moves to encoder-side correction. Third, the recalibrated model's head-tail curve must lie strictly outside the curves of reweighting and TESTAM.

As a secondary analysis, we compare a filtration-style encoder that prunes edges, following TimeFilter, against a non-pruning encoder in the few-shot bin. This tests whether pruning "discordant" edges to reduce average error removes the early spatial warning signs of tail events.