{"id":"3b498fbb-b5eb-4ae8-9f04-b22f822f8c84","arxiv_id":"2603.15802","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"ApolloPFN trains a time-aware prior-data fitted network on synthetic time series with exogenous variables and outperforms existing zero-shot forecasters on M5 and electricity price benchmarks.","lead":"ApolloPFN is a new forecasting model that uses external signals like price and weather directly, without any fine-tuning. It beats other zero-shot forecasters on retail and electricity benchmarks by training on synthetic data designed to mimic real-world time series.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Synthetic prior lacks temporal lags and nonstationary/trend structure, and the abstract names datasets never evaluated; the zero-shot generalization claim is therefore not established beyond contemporaneous periodic dependencies.","rationale":"The paper's central claim is zero-shot SOTA with exogenous covariates. For a PFN, zero-shot generalization is entirely determined by the training prior. The Sec. 4.1 prior is far more restrictive than 'realistic temporal patterns': the shared root is a zero-mean sum of two sinusoids and the SCM has no cross-time edges. This means the model has never observed a mechanism where an exogenous input at time t−ℓ affects the target at time t, nor a slowly evolving trend or a sudden level shift. The paper's own limitation paragraph explicitly concedes that unrepresented dependencies cannot be recovered. The reported exogenous benchmarks (electricity, M5) involve contemporaneous effects, so they do not exercise lagged or nonstationary mechanisms. The abstract's mention of UCI Air Quality and Solar Energy without any evaluation is additional evidence that the 'diverse set' claim is overstated. Other issues—no error bars, no code—are reproducibility concerns, but they do not bear directly on whether the mechanism supports the claimed generalization. The existing CONDITIONAL verdict remains appropriate: the paper reports genuine SOTA on the tested benchmarks, but the breadth of the zero-shot claim is not established until the prior's coverage of lagged and nonstationary dependencies is either demonstrated or the evaluation is expanded to the datasets named in the abstract.","tokens_in":14930,"tokens_out":12952,"duration_ms":136424,"concrete_test":"Augment the Sec. 4.1 sampler with (i) an option for time-lagged parent edges, e.g., y_t = f(x_{t-7}) + ε, and (ii) a root process that includes a stochastic trend (e.g., random walk plus sinusoids). Retrain ApolloPFN with this augmented prior and compare against the published model on (a) a held-out synthetic test set containing lagged and trending series and (b) the M5/electricity benchmarks. If the original model performs markedly worse on the lag/trend tasks while the augmented model does not, the missing temporal structure is confirmed; if benchmark performance is unchanged, the periodic-contemporaneous prior is sufficient for the tested scope.","verdict_should_be":"UNCHANGED","load_bearing_attack":"For a PFN, zero-shot generalization is determined entirely by the training prior. In Sec. 4.1, the only temporal structure is the root process v_t,r = α1 sin(φ1 t) + α2 cos(φ2 t), with φ drawn from log-uniform [1,10]. The SRNGN graph defines edges only among variables at the same time step; there is no x_{t-ℓ} → y_t or y_{t-ℓ} → y_t dependency. Thus the prior assigns zero probability to lagged exogenous effects, stochastic trends, level shifts, and regime-switching nonstationarities. The conclusion explicitly claims SOTA in regimes with 'sharp changes, discontinuities, or heterogeneous responses,' and the final limitation paragraph concedes that 'behaviors or dependencies that are not represented in the training data cannot be reliably recovered at test time.' Yet the only exogenous benchmarks reported are electricity prices (contemporaneous load/generation) and M5 (contemporaneous price/promo); the abstract also names UCI Air Quality and Solar Energy, which never appear in Sec. 5. So the evidence for a 'diverse set' of exogenous regimes is missing, and the mechanism that would explain transfer to genuinely novel lagged or nonstationary regimes is absent. This is the load-bearing gap in the central zero-shot claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces ApolloPFN, a prior-data fitted network (PFN) for zero-shot time-series forecasting with exogenous covariates. The authors modify TabPFN's synthetic data generation by (i) proposing a single-root graph-generation algorithm (SRNGN) and (ii) making the root process time-dependent via sinusoidal excitation. Architecturally, they add RoPE and absolute positional encodings and remove the train/test attention mask so that future exogenous values inform all forecast points. The model is evaluated on electricity price forecasting (sCRPS), M5 aggregations (RMSSE), and classical univariate M-series benchmarks, claiming state-of-the-art zero-shot performance against TabPFN-TS, Moirai, Chronos, and Sundial, with only 11M parameters.","tokens_in":15250,"tokens_out":6433,"duration_ms":65965,"significance":"If the claims hold, the paper makes a useful contribution: it extends the PFN paradigm to time series with exogenous covariates while remaining zero-shot, and its ablations (SRNGN vs. RGN, RoPE, full attention) provide concrete evidence about which design choices matter. The model is comparatively small and competitive with much larger foundation models. However, the empirical support is narrower and weaker than the text asserts: two benchmark families named in the abstract and conclusion are not evaluated, all tables are point estimates without uncertainty quantification, and the synthetic prior lacks lagged or nonstationary temporal structure. The core zero-shot generalization claim is therefore not yet established at the level of confidence the paper projects.","major_comments":[{"comment":"The abstract states that ApolloPFN is demonstrated on 'M5, electric price forecasting, UCI Air Quality, and Solar Energy datasets.' Section 5, however, evaluates only electricity price, M5, and univariate M-series benchmarks. UCI Air Quality and Solar Energy never appear in the full text. Since the paper's central claim is SOTA performance 'across a diverse set' of exogenous benchmarks, this missing evidence is load-bearing. The authors should either add the missing experiments or remove the unsubstantiated dataset names.","section":"Abstract and Section 5"},{"comment":"All results are point estimates with no standard errors, confidence intervals, number of seeds, or significance tests. Several reported differences are small (e.g., Table 3: M4(M) ApolloPFN 0.092 vs. Moirai 0.093; M1(Y) ApolloPFN 0.142 vs. Sundial 0.119, where ApolloPFN is not best). Without uncertainty quantification, the reader cannot tell whether the SOTA claims reflect stable improvements or noise. The paper should report variability across repeated evaluations (or at least bootstrap intervals) and, where appropriate, paired significance tests.","section":"Tables 1, 2, 3"},{"comment":"The synthetic prior's temporal structure is limited to a sum of one sine and one cosine per root node, with all SCM edges connecting variables at the same time step. There are no lagged dependencies (x_{t-ℓ} → y_t or y_{t-ℓ} → y_t), no stochastic trends, and no regime-switching processes in the training distribution. The conclusion (Sec. 6) explicitly concedes that dependencies absent from the training data cannot be recovered at test time. Yet the paper claims SOTA 'particularly in regimes where exogenous covariates drive sharp changes, discontinuities, or heterogeneous responses.' The only exogenous benchmarks shown involve contemporaneous covariates (load and price/promotion). This does not establish transfer to lagged or nonstationary regimes. The authors should either extend the prior with lagged edges and nonstationary root processes, or temper the claim and add benchmarks (e.g., t","section":"Section 4.1 and Section 6"},{"comment":"The absolute positional encoding formula is written as Ω_{t,2d+1} = sin(2π t 2^{2d+1} / 2^{12}) and Ω_{t,2d} = cos(2π t 2^{2d} / 2^{12}). As written, the frequency increases with d, so for d ≥ 6 the period is less than 2π and aliasing occurs. This is not the standard Vaswani-style encoding referenced in the text. If this is a typo in the manuscript, it must be corrected; if it is actually implemented, it is unclear how the architecture works. Since positional encodings are a core contribution, this formula needs clarification and the correct definition should be stated.","section":"Section 4.2, Eq. (2)"}],"minor_comments":[{"comment":"Typo: 'strong zero-short performance' should be 'zero-shot performance.'","section":"Section 5 opening"},{"comment":"Typo: 'TrancatedGamma' should be 'TruncatedGamma.'","section":"Appendix B.2"},{"comment":"Typo: 'Riemman approximation' should be 'Riemann approximation.'","section":"Appendix A"},{"comment":"Typo: 'generatied' should be 'generated.'","section":"Section 4.1"},{"comment":"Missing space in 'ApolloPFNbeats.' Also consider stating explicitly the number of independent trials used to produce each table.","section":"Table 1 caption"},{"comment":"The paper does not state whether code, model checkpoints, or evaluation scripts will be released. Given the dependence on a custom synthetic-data generator and architectural choices, a reproducibility statement would strengthen the paper.","section":"Reproducibility"}],"recommendation":"major_revision","confidential_remarks":"The abstract-benchmark mismatch appears to be a holdover from an earlier draft, and the missing error bars are standard for many forecasting papers, but together they weaken the central SOTA claim. The prior-limitation concern raised in the stress test is real but fixable: either broaden the synthetic prior or narrow the claims. I do not see internal inconsistency or circularity; the model is genuinely evaluated zero-shot on held-out benchmarks. The paper is worth a revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThe short version: ApolloPFN is a reasonable engineering extension of TabPFN-TS, and the authors identify real failure modes in the base model. But the central zero-shot claim is not as well supported as the abstract suggests. The synthetic training prior contains no lagged dependencies and no trend or regime-switching structure, so the model's success on contemporaneous exogenous benchmarks (electricity load, M5 promotions) does not demonstrate generalization to the 'sharp changes, discontinuities, heterogeneous responses' the conclusion brags about. The abstract also names UCI Air Quality and Solar Energy, but neither appears in the experiments.\n\nWhat's genuinely new: The SRNGN graph generator with a single root and longer paths speeds up training relative to RGN, and the combination of RoPE plus full sample attention in a PFN is a sensible fix for order invariance. The failure mode analysis of TabPFN-TS (recency bias, frequency dependence, confidence intervals) is well illustrated and useful. The reported wins on M5 and electricity price are credible, though we only have point estimates and no error bars or significance tests.\n\nWhere it gets soft: The root node excitation is just a sum of two sines/cosines with frequencies sampled from log-uniform [1,10]. That is high-frequency periodic noise, not the kind of low-frequency trends, level shifts, or lagged exogenous effects that real forecasting problems throw at you. The graph edges are contemporaneous, so the training distribution assigns zero probability to x_{t-l}->y_t or y_{t-l}->y_t. The authors' own limitation paragraph admits that dependencies not in the training data cannot be recovered. That is exactly the problem: the benchmarks used are all contemporaneous, so they do not stress the missing structure. The M5 and electricity price data are known at forecast time, so the model can condition on them directly; there's no need to learn a lag. Thus the 'diverse set' claim is supported by only two benchmark families, and one of them (M-series) has no exogenous variables at all.\n\nThe abstract/dataset mismatch is a real editor-level red flag. No code or data release is mentioned, which makes the empirical claims hard to check. The ablations are mostly qualitative (Figure 4) or a single training curve (Figure 3), so the contribution of each component is not quantified tightly.\n\nWho should read this: anyone working on PFNs for time series, or on zero-shot forecasting with exogenous variables. It's a useful case study in how synthetic priors determine what a PFN can and cannot do. It deserves a serious referee, but the revision needs to either fix the data generation to include lagged and nonstationary processes, or soften the zero-shot generalization claims and add the missing datasets.\n\nMy recommendation: send it to review, but with a strong request for code, uncertainty quantification, and an honest re-scoping of the claims.","headline":"Plausible PFN adaptation to exogenous time series, but the zero-shot claim exceeds the synthetic prior and the reported experiments.","tokens_in":15768,"tokens_out":4063,"would_cite":true,"duration_ms":39102,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that a prior-data fitted network can do zero-shot forecasting with exogenous covariates if the synthetic training prior and the attention mechanism are made time-aware.","keywords":["zero-shot forecasting","prior-data fitted networks","time series foundation models","exogenous covariates","synthetic data generation","positional encodings","electricity price forecasting","M5 competition"],"falsifier":"Train ApolloPFN with root nodes sampled as random walks (or regime-switching processes) instead of sine/cosine mixtures, keeping everything else fixed, and compare on M5 and electricity price benchmarks; a sharp performance drop would confirm the periodic-root prior is the load-bearing component. Alternatively, run ApolloPFN zero-shot on series with known long-range dependence (e.g., ARFIMA or Markov-switching data) and look for substantial degradation relative to models trained on such data.","tokens_in":14848,"feed_emoji":"📈","tokens_out":4765,"duration_ms":49308,"temperature":0.7,"pith_summary":"ApolloPFN is a forecasting model built on the prior-data fitted network idea: instead of training on labeled real series, it learns to imitate Bayesian inference from synthetic datasets. The paper's claim is that the reason earlier PFN-based forecasters fail is that both their training distribution and architecture treat time points as exchangeable, and that fixing this—by generating graphs with a single root and periodic root-node dynamics, and by adding rotary and absolute position encodings plus full attention—lets a small 11M-parameter model handle exogenous variables such as promotions, prices, load and weather zero-shot. On electricity price forecasting and M5 aggregations, ApolloPFN outperforms the next-best model in most settings, and it also holds its own on univariate M-series benchmarks. Why this matters: most deployed time series foundation models ignore exogenous covariates, so a model that natively uses them without fine-tuning could improve forecasting where external drivers cause sharp changes.","feed_headline":"Zero-shot forecaster with exogenous inputs beats larger time series models","feed_subtitle":"An 11M-parameter model trained on synthetic data tops M5 and electricity price benchmarks in zero-shot settings.","key_machinery":"The load-bearing machinery is the synthetic data prior. The paper introduces SRNGN, a graph generator that yields a single root node and long causal paths, and redefines root-node values as α1 sin(φ1 t) + α2 cos(φ2 t) with randomly sampled amplitudes and log-uniform frequencies. Those periodic roots propagate through the graph so every feature and target is time-correlated. On the architecture side, rotary position embeddings (RoPE) are applied in the sample-attention head to make nearby observations interact more strongly, absolute sinusoidal encodings add global position, and full attention lets future exogenous values inform each forecast horizon. Together these make order and relative di","core_discovery":"The central discovery is that the i.i.d. assumption embedded in TabPFN-TS-style training is the main obstacle, not the PFN paradigm itself. By replacing the random growing network with a Single Root Node Growing Network (SRNGN) and by exciting root nodes with randomly sampled sine and cosine functions over time, the synthetic datasets acquire temporal autocorrelation and long causal paths; combining this with rotary position embeddings (RoPE), absolute positional encodings, and full sample attention produces a model that treats order and nearness as relevant. The paper reports that these changes resolve characteristic failure modes—missed ordered patterns, weak trend extrapolation, lack of r","pith_inferences":["A reader might infer that the sine/cosine root process is the real bottleneck: regimes with strong nonstationarity, random walks, or long-memory dependencies are not represented in the prior, so the model may silently revert to poor extrapolation in those settings.","The SRNGN single-root topology couples all exogenous features to one common cause; datasets whose exogenous variables act independently on the target may be harder for the model than datasets where features and target share a common driver.","The time-aware interventions (RoPE + full attention) could be transferred to other transformer-based forecasters or tabular PFNs as a generic recipe, though the paper does not test that."],"forward_implications":["Zero-shot forecasting with exogenous variables is achievable without task-specific fine-tuning, reducing deployment overhead.","A prior-data fitted network with only 11M parameters can be competitive with—and on several benchmarks better than—models 30 to 70 times larger.","The specific failure modes of TabPFN-TS (order invariance, weak trends, lack of recency, over-wide intervals) are attributable to its i.i.d. training prior and can be corrected by data and architecture changes.","Because ApolloPFN learns from synthetic data, the same PFN training loop can be reused for new domains by changing the prior rather than by collecting real data."],"fun_headline_variants":["Time-aware AI forecaster beats larger models on M5","Exogenous-aware zero-shot forecaster tops M5 and electricity","11M-parameter synthetic-trained model wins zero-shot benchmarks","Causal synthetic priors unlock zero-shot exogenous forecasting"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The weakest premise is that the synthetic training distribution—graphs with a single root whose values are a sum of one sine and one cosine—is representative enough of real time series with exogenous drivers; if real dynamics are nonstationary, regime-switching, or long-range dependent, the learned zero-shot prior will not transfer.","fun_headline_variants_meta":{"raw":{"variants":["Time-aware AI forecaster beats larger models on M5","Exogenous-aware zero-shot forecaster tops M5 and electricity","11M-parameter synthetic-trained model wins zero-shot benchmarks","Causal synthetic priors unlock zero-shot exogenous forecasting"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001226,"raw_usage":{"total_tokens":4886,"prompt_tokens":761,"completion_tokens":4125,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":505,"completion_tokens_details":{"reasoning_tokens":4058}},"tokens_in":505,"tokens_out":4125,"duration_ms":27591,"temperature":1.0,"reasoning_tokens":4058,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T18:03:49.122765+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train ApolloPFN with root nodes sampled as random walks (or regime-switching processes) instead of sine/cosine mixtures, keeping everything else fixed, and compare on M5 and electricity price benchmarks; a sharp performance drop would confirm the periodic-root prior is the load-bearing component. Alternatively, run ApolloPFN zero-shot on series with known long-range dependence (e.g., ARFIMA or Markov-switching data) and look for substantial degradation relative to models trained on such data.","supporting_citations":[],"review_version":1}