- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper asks whether traffic forecasting benefits more from a longer history of the target sensor ("temporal depth") or from the recent history of more sensors ("spatial breadth"). On the four LargeST subsets with traffic flow at 15-minute intervals, it measures the forecasting error over a grid of history lengths and neighbour counts using ridge regression and a small MLP-Mixer. It reports that the largest gain comes when the history covers one full day, that neighbours do not help as much, and that congestion usually reaches adjacent sensors within one 15-minute interval. On a freeway corridor with 30-second speed data from FT-AED, nearby sensors do help when traffic is "near capacity". Finally, the paper proposes a model without spatial input, based on an adapted TimesFM with a historical-average prior and a gate for atypical traffic, which reaches the lowest average MAE on all four LargeST subsets.
Critical strengths
1. The paper studies where the accuracy of traffic forecasts comes from, instead of proposing another model. Most traffic forecasting models are graph-based and assume that data from other sensors is the main source of accuracy. The paper checks this assumption directly. Such studies are rare, and their results show where model developers should spend their effort, which is more useful than a small gain on a benchmark. Some results are surprising and useful: Table 2 shows that random sensors recover most of the gain of the nearest ones, and Figure 2 makes the trade-off between history length and the number of neighbours easy to see.
2. The analysis is careful, and the paper states its limitations openly. It uses two models of different complexity (ridge regression and an MLP-Mixer), an ablation study that replaces the nearest sensors with random ones, and an analysis of the "lead–lag information" between adjacent sensors based on traffic-flow theory. It also reports the cases where other sensors do help, namely "sudden-onset events" and traffic "near capacity" when the data are recorded every 30 seconds. The weaknesses below concern how well the experiments support the main conclusion, and I would like to see this study improved and published.
Critical weaknesses
1. The evidence does not show that spatial information adds little once the history is long.
The central claim is that "once the target's history spans one daily cycle, additional spatial observations add little unique information" (lines 471-472). However, the paper's own experiments do not establish it:
- The results of ridge regression contradict it. At , ridge regression still gains 0.5-0.9 MAE from neighbours (lines 248-251), which is comparable to the gain of the temporal-only recipe over Graph WaveNet on San Diego (0.75 MAE).
- The MLP-Mixer gains almost nothing from neighbours at , but it becomes unstable with many neighbours (Appendix B.4), which the paper attributes to the "estimation and optimization cost" of many inputs (lines 891-892). Therefore, the lack of gain may be a limitation of this model and may not hold for the data.
- No graph model is trained with a long history (lines 483-484), and the only baseline that reads one day of all sensors, RPMixer, outperforms the temporal-only recipe on San Diego (17.25 vs 17.43 MAE, Table 4). Moreover, SBATransformer [1] uses data from all sensors with a 96-step input and reports lower MAE than the paper's best model on all four subsets at every horizon (e.g., 15.92/17.69/19.77 vs 15.96/18.55/20.69 at horizons 3/6/12 on GBA). Its evaluation pipeline may differ, but the authors' research log in the supplementary material mentions this work and notes that comparing a 512-step recipe with 12-step graph models favours the recipe (
FINDINGS.md:656-659). However, the paper mentions neither SBATransformer nor this unequal comparison.
As a result, the main comparison changes two factors at once: the recipe uses 512 steps and no spatial input, while the graph baselines use 12 steps and spatial input. At an equal history of 12 steps, the graph model wins (18.18 MAE for Graph WaveNet vs 21.20 for the 12-step MLP-Mixer, Table 16), and at an equal long history, which is what the central claim is about, no such comparison exists.
2. The conclusions rest on a single type of data.
All forecasting results use 15-minute traffic flow from LargeST, and the "nonlinear allocation plane" is measured only on San Diego (line 483). According to the paper's own explanation, whether other sensors help depends on how often the data are recorded: they help little at 15-minute intervals (LargeST), where congestion reaches the next sensor within one time step, but they do help at 30-second intervals (FT-AED). METR-LA and PEMS-BAY, the most common benchmarks, use 5-minute intervals, which lie between these two settings, but the paper leaves them for future work (Section 6). Without them, statements such as "predictive information is far more effectively captured through temporal depth than additional spatial observations" (lines 30-32) claim more than the experiments show.
Moreover, the abstract states that performance "is determined primarily by whether the input covers one complete daily cycle" (lines 18-20). Yet the recipe still improves by 0.80 MAE when its history grows from one day to 5.3 days (18.23 vs 17.43, Table 14). This is larger than the gain from adding the four nearest sensors in Table 2 (0.54 MAE) and ten times larger than the gain from attention over other sensors on top of the adapted TimesFM (at most 0.08 MAE, line 257).
3. Sections 4.3, 4.4 and 5 are hard to follow.
Unlike Sections 3-4.2, these sections put several analyses into long paragraphs, use terms defined only in the appendix, and read like an unedited draft. Section 4.3 mixes three different analyses, and its text does not match Figure 3. Table 3 cannot be understood without Appendix H. Section 5 reports its main results through many numbers in the text, uses the gate before introducing it, and leaves the MLP-Mixer, which the paper calls its main evidence, out of Table 4. I suggest splitting Section 4.3 into parts with one claim and one figure each, defining all terms of Table 3 in the main text, and building Section 5 around one table that includes the MLP-Mixer.
4. Some choices were made on the test set, and key results cannot be verified.
- The forecasting recipe never uses the validation split. According to the authors' research log, all its numbers are test results (
FINDINGS.md:24-25), including the sweep that selected the 512-step history (FINDINGS.md:101-115) and the decision to keep the calendar embeddings (FINDINGS.md:128-140). The paper does not disclose this. - The retrained Graph WaveNet and STID use early stopping on test windows (
scripts/e1_baselines.py:166,226), although Appendix A.3 states that early stopping uses validation data. - The FT-AED study (Section 4.4, Appendix H), the only experiment where neighbours clearly help, has no code in the supplementary material.
- Section 4.2 states that random sensors give the same benefit as the nearest ones "at every history length" (lines 265-266), but this ablation study was run only at and (
scripts/modal_grid.py:64).
[1] Chen et al. "Towards Scalable and Structured Spatiotemporal Forecasting," 2025
Questions for authors
- How does a model that reads other sensors, such as Graph WaveNet or PatchSTG, perform when trained with 96 or 512 steps of history, and how does the recipe compare with SBATransformer under the same evaluation pipeline?
- Can at least the errors of ridge regression over the grid of history lengths and numbers of neighbours (the "allocation plane", Figure 2, left) and the analysis of the "lead–lag information" between adjacent sensors (Section 4.3) be repeated on METR-LA or PEMS-BAY? As an optional suggestion, the datasets from [2] seem to fit this study well: they cover city road networks with up to 100,000 road segments and real road connectivity, with speed and volume at 5-minute resolution. Since adjacent road segments in a city are close to each other, these data could show whether spatial input becomes useful when congestion takes longer than one time step to reach the next segment.
[2] Velikonivtsev et al. "Metropolis-Scale Road Network Datasets for Fine-Grained Urban Traffic Modeling," 2026