acceptodds
Under review as a conference paper at ICLR 2027

MAFF: Mask-Aware Forecasting Framework for Time Series with Missing Values

Abstract

Deployed time series are rarely complete: readings drop out, devices go offline, subsystems fail, and recent observations are lost before they are used. The standard remedy is to impute the gaps and forecast on the completed series, treating the repair as neutral. It is not. Imputed entries enter the instance-normalization statistics that modern forecasters depend on: interpolation deflates the window variance and shifts its mean, the model cannot distinguish fillers from measurements, and the resulting bias is applied back to the forecast at denormalization. We identify this contamination as a mechanistic, architecture-independent source of error, and we eliminate it. MAFF, our mask-aware forecasting framework, keeps observed entries as the only source of statistics, injects a gated summary of the missingness structure at an architecture-adapted position in the backbone, and substitutes a learnable token for fully-missing channels. The repair is parameter-free: it removes the contamination without adding capacity, and parameter-matched ablations isolate its contribution from the wrapper that carries it. Across four missingness protocols (random, block, correlated-group, and tail) at 30% missingness with masks shared across models, MAFF improves MSE on all five benchmarks when coupled with CMoS. We evaluate four backbone families independently: every one of them improves, and the gains are strongest and most consistent on CMoS, the convolution-based backbone whose statistics are most exposed to fillers.

Primary area:
learning on time series and dynamical systems
Keywords:
time series forecasting, missing values, mask-aware modeling, instance normalization, imputation artifacts, multivariate time series

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_onZQOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

MAFF addresses forecasting from incomplete time series within an impute-and-forecast pipeline. Its central intervention computes instance-normalization statistics from observed entries only and reuses them for denormalization, excluding fillers from the statistics applied to both inputs and forecasts. The framework also injects gated per-channel missingness summaries into backbone-specific feature pathways and provides a learnable MASK token for fully missing channels. Evaluation covers five datasets, four backbones, four synthetic missingness protocols, and four forecasting horizons, using shared masks across methods and three mask seeds with a fixed training seed. Reported aggregate MSE reductions are 2.7% for CMoS, 0.7% for iTransformer, 0.4% for CycleNet, and 0.2% for DLinear. Component ablations provide evidence for observed-only normalization, while clean-data controls and reported mask statistics leave questions about attribution and protocol comparability. The MASK token does not activate in the reported experiments.

Critical strengths
  • S1. Clear mechanistic problem formulation and lightweight pipeline integration. The paper identifies a concrete pathway through which fillers affect forecasting: they enter instance-normalization statistics, which are subsequently reused during denormalization. Equation 1 directly excludes fillers from those statistics without introducing a learned imputer or additional parameters for this component. The normalization-only ablation retains the baseline's parameter count and reports a 2.8% aggregate MSE reduction on the ETTh1/CMoS subset with horizons 96 and 336, with positive gains under P2, P3, and P4 (Section 6 and Table 8). This is meaningful evidence that at least part of the improvement does not require additional capacity. The intervention is simple and potentially practical, although excluding fillers does not guarantee recovery of the complete window's true statistics.
  • S2. Transparent empirical reporting of negative results, boundary conditions, and per-window behavior. The evaluation spans four architecture families and four missingness protocols with shared masks across methods. The authors report absolute MSE/MAE, component ablations, clean-data controls, and mask-seed variability, while explicitly disclosing small or negative gains, the inactive MASK token, and the dropout-stream difference between paired runs (Sections 5, 6, Appendices A, F, and G). The per-window analysis further shows that 36.2% of test windows regress under P2 in the examined ETTh1/CMoS setting despite a positive aggregate gain. This reporting makes the method's benefits and limitations assessable and provides a useful basis for investigating the unresolved attribution questions.
Critical weaknesses
  • W1. The missingness-specific contribution is not sufficiently separated from wrapper capacity and training-path effects. CMoS improves even on complete inputs, by 4.8% on ETTh1 and 4.0% on ETTm1 after pooling horizons. Using the displayed values in Table 1, its gain under missingness is smaller than the corresponding clean-data gain at the same dataset and horizon in 60 of 80 cells, including all 20 P1 cells. This does not establish that the missing-data gains are artifacts, but it makes a direct attribution control important. The paper acknowledges additional wrapper parameters, a normalization-epsilon difference between MAFF and native CMoS (10−5vs10−1010^{-5} vs 10^{-10}), and different dropout draws caused by wrapper construction. The clean-data result is a useful diagnostic, but it is not an established bound on these effects under missing inputs and cannot simply be subtracted as a causal correction.

The normalization-only ablation is valuable parameter-matched evidence, but it is limited to ETTh1 and two horizons, and the manuscript does not establish whether epsilon and other training-path differences are also matched in that comparison. Moreover, Table 6 measures variability over mask seeds with a fixed training seed, leaving training-seed robustness unresolved. This is particularly consequential for the 0.2 to 0.7% aggregate gains outside CMoS. The interpretation of the feature pathway also needs clarification. Sections 5 and 6 describe the embedding as conveying gap position, whereas the three specified summaries can be identical for gaps of the same length at different temporal locations. Protocol-specific priors or the completed input may still provide positional information, but this differs from explicitly encoding location in the mask summaries. Likewise, the threshold sweep establishes insensitivity to the tested values of 𝜏, not the necessity of gating. These are secondary limitations on the mechanistic explanation rather than independent grounds for rejection.

W2. The reported mask statistics are inconsistent with interpreting the protocols as having a common 30% realized missing rate. Sections 3 and 4 state that each protocol targets 30% missingness. However, Table 16 reports mean longest missing runs under P2 of 53.86 for ETTh1/ETTh2, 53.83 for ETTm1/ETTm2, and 92.33 for Weather, with look-back length L=96. For each window and channel, the longest missing run cannot exceed the total number of missing entries. Under the table's stated averaging convention,

bw,c≤Nw,cmiss⟹rˉ≥bˉ96.b_{w,c} \leq N_{w,c}^{miss} \Longrightarrow \bar{r} \geq \frac{\bar{b}}{96}.

These values therefore imply mean realized missing rates of at least approximately 56.1% on the ETT datasets and 96.2% on Weather. A nominal generator parameter of 0.3 could differ from the realized rate, but the manuscript does not explain this discrepancy. Shared masks still support within-protocol comparisons between the baseline and MAFF. The concern is that missingness amount and structure are confounded when comparing gains across protocols, weakening the claim that the ordering reflects exploitable structure rather than severity. The same clarification is needed to interpret the missingness-strength sweep.

Questions for authors
  • Q1, addressing W1: attribution and statistical reliability. Were normalization epsilon, backbone initialization, minibatch order, and stochastic draws matched in the normalization-only ablation? Please clarify and, where necessary, repeat a representative part of the existing ETTh1/CMoS ablation with these factors controlled and paired results across several training seeds. A capacity-matched wrapper control with constant mask features would help isolate the injection contribution; to test that contribution specifically, it should retain the same observed-only normalization as full MAFF. Reporting normalization-only results under none, with matched epsilon, and absolute MSEs for the clean-data controls would further clarify the source of the offsets. These are targeted controls for the proposed explanation, rather than a request to repeat the entire benchmark grid.

  • Q2, addressing W2: mask-generation clarification and validation. How does the nominal ratio determine block counts, lengths, overlaps, and channel coverage? Please reconcile Table 16 with the 30% setting and report realized missing fractions for each dataset/protocol, including the strength sweep. Why is the P3 gain reported as +8.9% at every tested ratio, and does the nominal ratio change its generated masks? If realized rates differ substantially, does the protocol-ordering conclusion survive a comparison at matched realized rates on a representative existing setting?

  • Q3, addressing the feature-pathway interpretation in W1: positional information and mathematical specification. Please provide the exact equations and dimensions mapping the three summaries through gating, the GELU projection, and residual injection. Are any raw-mask or explicit positional features used beyond the three stated summaries? If not, how should the statements that the embedding conveys gap position be understood? In particular, please distinguish information explicitly represented by the summaries from information inferred through the completed input or a protocol-specific prior.

  • Q4, addressing the component-level evidence in W1: gate necessity. Figure 7 already evaluates threshold sensitivity. On that existing setting, how does MAFF perform with gc≡1g_c ≡1, keeping the summaries, projection, and training setup unchanged? This would clarify whether the gate adds value beyond the learned projection or whether the evidence supports the simpler ungated feature pathway.

3 comments

We thank the reviewer. W1, W2 and the feature-pathway questions all bear on one quantity: how much of the wrapper's gain belongs to the mechanism we claim. We ran 564 additional runs.

Q1 Attribution. You ask for a wrapper that retains observed-only normalization, holds the mask features constant and matches capacity. We implemented it: the full wrapper, the same parameter count and observed-only normalization, with the three summaries zeroed so the injection path carries no missingness information.

ProtocolFullZeroedDifference
P10.51750.5177−0.03%
P20.83830.8386−0.03%
P30.54340.5436−0.03%
P40.58060.5808−0.03%

The paired difference is 0.03% (24 pairs, all with the same sign, paired standard deviation 0.010%), so the mask information does contribute; the residual branch of the same path spans CMoS −5.56% to DLinear +0.38% on clean data, two orders of magnitude larger.

The component decomposition on clean data (protocol none) shows the other side, ETTh1, three seeds, two horizons.

Backboneoffnormalization onlyfull
CMoS0.54080.54080.5107
iTransformer0.43910.43910.4386
CycleNet0.47790.47790.4735
DLinear0.44180.44180.4435

These are the none-row absolute values requested in Q1. The normalization-only arm is bit-identical to the baseline on all four backbones, maximum absolute difference 5.96e-8; epsilon moving from 1e-10 to 1e-5 gives 1.03e-7 relative, below resolution. The offset comes entirely from the injection branch and changes direction, CMoS −5.56%, CycleNet −0.90%, iTransformer −0.10%, DLinear +0.38%, which accounts for the 60 of 80 cells.

DLinear is an additional negative control: no instance normalization, nothing for the repair arm to act on, difference zero under all four protocols. This also answers the dropout-stream point: the wrapper does change the random stream, but here both arms run the same definition and forward path and the difference is numerically zero.

On minibatch order and initialization, both arms use identical definitions, the same seed and batch order, differing only in where the statistics come from, at 937 parameters each.

Training seeds: three (2021, 7, 13), mask seed fixed, ETTh1, three backbones, four protocols, two horizons, two arms, 144 runs. For a fixed configuration the training seed moves the error by 0.02% to 2.29%, median 0.12%. Paired difference of the repair arm from the baseline follows (%, negative means lower MSE; each cell gives horizon 96 then horizon 336).

BackboneP1 point-wiseP2F short blocksP3F correlated groupsP4 tail
CMoS+0.54 / +0.46−0.26 / +0.02−2.32 / −1.68−4.79 / −5.04
iTransformer+0.35 / +0.24−0.90 / −1.83−0.78 / −0.23−2.24 / −1.43
CycleNet+0.87 / +0.62+0.20 / +0.26−1.07 / −0.94−1.72 / −1.09

21 of 24 cells share a sign across the three seeds and all seven cells above 1.5% do, with a maximum seed standard deviation of 0.92%. The three with mixed signs are CMoS short blocks at 336 (+0.02%, seeds −0.07, +0.04, +0.07), iTransformer correlated groups at 336 (−0.23%, −0.61, −0.41, +0.33) and iTransformer tail at 336 (−1.43%, −3.72, −2.26, +1.70); there the seed spread exceeds the effect and we do not rely on that cell. Point-wise missingness is consistently positive across the three backbones and the three seeds, from +0.24% to +0.87%.

Q2 to Q4, with the strength sweep, follow in the next comment.

Q2 Mask generation, realized rates and protocol ordering.

ProtocolETTWeatherLongest run (ETT / Weather)
P130.1%30.0%3.5 / 3.5
P285.4%96.9%53.5 / 92.5
P317.0%18.8%16.3 / 18.1
P430.2%30.2%29.0 / 29.0

On block counts and coverage: P2 computes its block count as L⋅C⋅r/ℓL \cdot C \cdot r / \ell with ℓ=0.1L\ell = 0.1L, but every block covers all channels, so coverage scales with the channel count, about 85% at 7 and 97% at 21, with overlapping starts keeping it slightly below the bound. P3 fixes the intra-group block length at 20% of the window, so the ratio only reaches sampling inside an already zeroed region and the realized fraction stays at 17% to 19%. Hence the same number at every ratio: the ratio never reaches the mask, which is not the model being insensitive to it. The corrected implementation is a separate protocol and P1 to P4 stay bit-identical; afterwards P2 realizes 28.0% at a nominal 30% and no longer scales with channels, and P3 realizes 30.4% and scales with the ratio (9.8%, 30.4%, 50.0%).

On whether the ordering survives at matched rates we tested three backbones by three strengths (ETTh1, three seeds, paired differences, negative means lower MSE):

BackboneProtocolr=0.1r=0.3r=0.5
CMoSshort blocks+0.02%−0.19%−0.60%
CMoScorrelated long blocks−0.11%−2.45%−6.09%
CMoStail−0.72%−5.64%−12.32%
iTransformershort / corr. / tail−0.38% / −0.36% / −0.17%−1.10% / −0.59% / −0.99%−0.41% / −0.37% / −1.46%
CycleNetshort / corr. / tail+0.34% / +0.23% / −0.26%+0.33% / −1.41% / −1.53%−0.06% / −3.38% / −4.07%

The ordering has two parts. Tail strongest and correlated groups second holds at matched rates, and the backbones share the shape. Block missingness ranked second because its realized fraction was 85%, not because of its structure; at the matched 28% it falls into the noise with mixed signs, so the revision restricts the gain to long contiguous gaps and tail gaps. The same protocols on ECL (321 channels): +1.10%, −0.36%, −2.95%, −6.68%.

Q3 Equations and position. The three summaries (missing rate, longest contiguous missing run and fully-missing indicator) give the same input for equal-length gaps at different positions, so they carry no position; position is carried by the per-position normalized values and by the completed input. Sections 5 and 6 attribute the gain to the model knowing where the gap is without separating information the summaries explicitly represent from information inferred from the input, and the revision corrects that and gives the full equation: x←x+α(Proj([x,m])−x)x \leftarrow x + \alpha(\mathrm{Proj}([x, m]) - x), where m=g⊙[miss rate,longest run/L,fully-missing]m = g \odot [\text{miss rate}, \text{longest run}/L, \text{fully-missing}], Proj:RL+d→RL\mathrm{Proj}: \mathbb{R}^{L+d} \to \mathbb{R}^L, and α\alpha is a per-channel scalar initialized small; backbones with aligned tokens project the embeddings at the same position. Proj\mathrm{Proj} is initialized to the identity on the zz block and to zero on the mm block and the bias, so training starts from no injection, and no raw mask or explicit positional feature enters this path beyond the three summaries.

Q4 Necessity of the gate. Across τ∈{1,4,16}\tau \in \{1, 4, 16\} the error changes by less than 2e-3%, so the gate is not a sensitive hyperparameter. Together with the paired control in Q1, the body of the gain of this path comes from the learnable residual, and the gate and the summaries are corrections of the same small order; the revision describes the gate as an initialization detail.

We hope these clarifications address your concerns. We would be glad to provide further details if needed.

Thank you for auditing the mask generator and running the matched controls. The results strengthen the case for observed-only normalization, particularly for tail gaps on CMoS. Given the small incremental gain from the mask summaries and the variation in protocol ordering across backbones, I am keeping my rating at weak rejection. I have no further questions.

Reviewer_n3gtOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

MAFF augments impute-and-forecast models with observed-only instance normalization and gated per-channel missingness summaries injected at architecture-specific locations. It also includes a learnable token for fully missing channels, although this component never activates in the reported evaluations. Experiments cover four backbones, five benchmarks, four forecast horizons, and four synthetic missingness protocols, using shared masks and three mask seeds with a fixed training seed. The reported overall MSE improvements are 2.7% for CMoS, 0.7% for iTransformer, 0.4% for CycleNet, and 0.2% for DLinear. Component ablations on ETTh1 provide additional evidence about normalization and feature injection.

Critical strengths
  1. A simple and practically relevant intervention. Equation (1) explicitly prevents imputed values from entering the normalization statistics. This modification requires no additional normalization parameters or learned imputer and can be incorporated into existing normalization-based backbones without changing the forecasting objective.

  2. Broad evaluation with controlled input pairing. The study spans multiple architectures, datasets, horizons, and missingness patterns. Providing the same masks and completed inputs to each off/full pair makes these comparisons useful for assessing the overall modification under the evaluated conditions.

  3. Useful ablations and transparent reporting. The paper includes normalization-only and injection-only ablations, fully observed controls, absolute MSE/MAE results, mask-seed variability, and negative cases. It also explicitly acknowledges that the MASK token does not contribute to the reported gains. These disclosures make the evidence and its limitations inspectable.

Critical weaknesses
  1. The P2 mask statistics require reconciliation with the stated 30% missingness. Table 16 reports mean longest missing runs of 53.83–53.86 steps on ETT and 92.33 on Weather, with a look-back of 96. Under the table's stated definition, the missing count must be at least the longest missing run, implying average missing fractions of at least approximately 56.1% and 96.2%, respectively. Although Sections 3–4 describe 0.3 as a target, the achieved rates and the reason for this discrepancy are not explained. Consequently, the cross-protocol results do not isolate missingness structure at a common realized severity. This issue does not by itself invalidate within-protocol off/full comparisons using shared masks.

  2. The controls do not fully isolate the proposed mechanism from other changes. The clean-data controls show sizeable CMoS off/full differences, including 4.8% on ETTh1 and 4.0% on ETTm1 (Section 5). These differences cannot establish a bound on the effects of the wrapper's other changes after training on corrupted inputs. The parameter-matched normalization-only ablation is useful; however, the manuscript does not establish that epsilon is matched, despite specifying 1e-5 for the mask-aware path versus 1e-10 for native CMoS. The full wrapper also adds a trainable sequence projection. An epsilon-matched normalization comparison and an equally parameterized mask-blind wrapper on the same corrupted inputs would more directly isolate the contributions of observed-only statistics and mask information.

  3. The specified summaries do not establish the claimed recovery of temporal gap locations. Section 3 defines the embedding inputs as missing rate, longest missing-run length, and a fully-missing indicator. These summaries alone cannot distinguish equal-length missing blocks at different temporal positions. This leaves the positional explanations in Sections 5–6, particularly the explanation of the P1 injection gain, insufficiently substantiated. The concern applies to the explicit summary embedding: the input signal, full-mask normalization, or a known tail protocol may still provide positional cues. The paper should clarify how those information sources support the proposed explanation. An explicit residual-injection equation is also needed to verify the claimed identity initialization.

  4. The incremental value of the particular gating and injection design remains unresolved. The off/full comparisons quantify changes relative to native mask-blind backbones, but do not establish an advantage over simpler mask-aware conditioning with matched normalization and comparable capacity. The component ablation evaluates the inclusion of the injection module, while the threshold sweep evaluates sensitivity to tau. Neither directly isolates the contribution of gating or the architecture-specific injection choices. A simpler mask-aware control, such as an ungated variant or direct mask conditioning, would clarify which aspects of the proposed design account for the improvement.

  5. The uncertainty analysis does not characterize robustness across training realizations. The three mask seeds measure sensitivity to sampled missingness under a fixed nominal training seed. Appendix G additionally states that wrapper construction changes the dropout random stream between the two arms. The reported variability therefore does not establish stability across independent training initializations. This matters particularly for the 0.2–0.7% aggregate improvements on the non-CMoS backbones. Different dropout draws do not invalidate a comparison, but training variability should be characterized before drawing strong robustness conclusions from these small average gains.

Questions for authors
  1. Protocol clarification: What are the realized missing fractions for each dataset and protocol, and how does the nominal ratio determine the masks? Please reconcile Table 16's P2 run lengths with the 30% framing and report the achieved rates for the missingness-strength sweep.

  2. Causal attribution: Are epsilon values identical in the normalization-only and baseline arms? What evidence separates observed-only statistics and informative mask features from numerical stabilization and the additional trainable projection? Matched controls on the same corrupted inputs would be particularly informative.

  3. Method clarification: Does the embedding receive positional or raw-mask information beyond the three stated summaries? If not, how should the positional explanation of the injection gains be interpreted? Please also provide the exact residual-injection equation and initialization so that the claimed identity property can be verified.

  4. Incremental contribution: How does MAFF compare with simpler mask-aware conditioning or an ungated injection variant under matched normalization and comparable capacity? Which results support the specific gating and architecture-adapted injection choices?

  5. Training robustness: Do the small non-CMoS improvements persist across independent training seeds? How are backbone initialization and stochastic training controlled, and can training variability be characterized separately from mask variability?

3 comments

We thank the reviewer for a careful check. The inference from Table 16 agrees with our implementation: the reproduced mean longest missing runs are 53.5 and 92.5 steps, against the 53.86 and 92.33 in the table. We re-audited the mask generator and ran 564 matched controls in one environment.

Q1 Realized rates, and how the nominal ratio determines them.

ProtocolETT realizedWeather realizedLongest run (ETT / Weather)
P1 point-wise random30.1%30.0%3.5 / 3.5
P2 temporal blocks85.4%96.9%53.5 / 92.5
P3 correlated groups17.0%18.8%16.3 / 18.1
P4 window tail30.2%30.2%29.0 / 29.0

Both causes are in the generator. P2 sets the block count to L⋅C⋅r/ℓL \cdot C \cdot r / \ell with ℓ=0.1L\ell = 0.1L, but each block zeroes every channel of that time band, so the realized coverage depends on CC times the nominal ratio rather than on the ratio alone: it follows 1−exp⁡(−Cr)1 - \exp(-C r), giving 85% at 7 channels and 97% at 21. This explains why one nominal ratio produces two very different severities on the two datasets.

P3 fixes the intra-group block length at 20% of the window, and the ratio only enters the per-group count and the per-group length, both clipped to a column region that is already fully zeroed, so the ratio has no further effect on the mask and the realized fraction stays at 17% to 19% whatever value is passed in.

The fix is implemented as a separate protocol and leaves P1 to P4 bit-identical, so the masks of the main experiments remain reproducible. After the fix, P2 realizes 28.0% at a nominal 30% and no longer scales with the channel count (28.1% and 27.3% at 321 channels), and P3 realizes 30.4% and varies monotonically with the ratio (9.8%, 30.4%, 50.0%). Because both arms share the same masks, the within-protocol off/full comparisons and all significance statements are unaffected, as the reviewer notes; what is affected is the statement that all four protocols target 30%, and the comparison of severity across protocols.

Q2 Epsilon and causal attribution. The reviewer lists four candidate explanations (observed-only statistics, mask features, numerical stabilization and the projection) and we address each in turn.

Begin with the component decomposition on clean data (protocol none), ETTh1, three seeds and horizons 96 and 336, MSE.

Backboneoffnormalization onlyinjection onlyfull
CMoS0.54080.54080.51070.5107
iTransformer0.43910.43910.43860.4386
CycleNet0.47790.47790.47350.4735
DLinear0.44180.44180.44350.4435

Numerical stabilization is not the explanation. Native CMoS uses 1e-10 and our path uses 1e-5, five orders of magnitude apart, while in this table the normalization-only arm differs from the baseline on clean data by a relative 1.03e-7, which is floating-point noise; the revision will align the two constants.

The additional projection is fully visible on clean data: the offset comes entirely from the injection branch and changes direction, CMoS −5.56%, CycleNet −0.90%, iTransformer −0.10%, DLinear +0.38%, so this term is tied to how a backbone handles its input rather than to missingness. The 60 of 80 cells the reviewer counted come from this branch.

For the mask features we ran the same-parameter control suggested in W2: the full wrapper, the same parameter count and normalization, with the summaries zeroed so the injection path carries no missingness information.

ProtocolFull MAFFZeroed featuresDifference
P10.51750.5177−0.03%
P20.83830.8386−0.03%
P30.54340.5436−0.03%
P40.58060.5808−0.03%

The paired difference is 0.03% (24 pairs, all with the same sign, paired standard deviation 0.010%). The path therefore has two separable effects. The injection branch is a 96×96 learnable linear map, 10,120 parameters active in the forward pass, plus one scalar per channel; its effect on clean data is the injection-only column above. The mask summaries add a 0.03% refinement on top. The split is guaranteed by construction: the summary input is identically zero when nothing is missing, so the mask branch is exactly inert there, and one path switches automatically between the clean and the corrupted case.

What remains is the observed-only statistic itself, bit-identical to the baseline on clean data. DLinear provides the decisive control: no instance normalization, nothing for the repair arm to act on, difference 0.0000 under all four protocols. This also rules out parameters, initialization, dropout streams and data order, since both arms run the same definition and forward path and differ only in where the statistics come from.

The strength sweep and its prediction, then Q3 to Q5, follow in the next comment.

On magnitude. The gain against gap structure follows, ETTh1, three seeds.

BackboneProtocolr=0.1r=0.3r=0.5
CMoSshort blocks+0.02%−0.19%−0.60%
CMoScorrelated long blocks−0.11%−2.45%−6.09%
CMoStail−0.72%−5.64%−12.32%
iTransformershort blocks−0.38%−1.10%−0.41%
iTransformercorrelated long blocks−0.36%−0.59%−0.37%
iTransformertail−0.17%−0.99%−1.46%
CycleNetshort blocks+0.34%+0.33%−0.06%
CycleNetcorrelated long blocks+0.23%−1.41%−3.38%
CycleNettail−0.26%−1.53%−4.07%

On the CMoS tail at r=0.5 the paired difference averages −12.32% with a paired standard deviation of 0.97%, and all six pairs (two horizons by three seeds) share a sign; at r=0.3 it averages −5.64% with 0.36%, again six of six. Short blocks give +0.02% and −0.19% with mixed signs, and −0.60% at r=0.5 with six of six. Point-wise missingness is discussed under Q5.

The shape is predictable in advance. Interpolation makes the values inside a gap a linear combination of the two observed endpoints, so the variance inside the gap is close to zero, and the longer the gap the more the window variance is compressed and the mean pulled to the endpoints. A tail gap has no right anchor, so the fill degenerates into constant extrapolation and both moments are biased; for point-wise missingness the fill comes from adjacent observations and the bias is negligible. The expected order is tail, then long internal gaps, then short gaps and point-wise missingness, rising monotonically in length and strength. The measurements match this prediction, which is why we treat the part as a mechanism rather than a tuning result.

Q3 Positional information and injection equation. The three summaries (missing rate, longest contiguous missing run and fully-missing indicator) give exactly the same input for gaps of equal length at different positions, so they carry no position. Positional cues come from two places: the completed input, where gap positions are visible in the shape of the values, and the per-position values that are normalized. Section 5 attributes the P4 gain directly to the summaries without separating these layers, and the revision corrects that.

The injection is x←x+α(Proj([x,m])−x)x \leftarrow x + \alpha(\mathrm{Proj}([x, m]) - x), where m=g⊙[miss rate,longest run/L,fully-missing]m = g \odot [\text{miss rate}, \text{longest run}/L, \text{fully-missing}], Proj:RL+d→RL\mathrm{Proj}: \mathbb{R}^{L+d} \to \mathbb{R}^L maps the concatenated features back to the sequence length, and α\alpha a learnable scalar per channel initialized small; backbones with aligned tokens project the embeddings at the same position. Proj\mathrm{Proj} is initialized to the identity on the zz block and to zero on the mm block and the bias, so the starting point of training is equivalent to no injection, and no raw mask or explicit positional feature enters this path beyond the three summaries.

Q4 Necessity of the gate. Across τ∈{1,4,16}\tau \in \{1, 4, 16\} the error changes by less than 2e-3%, so the gate is not a sensitive hyperparameter. With the control in Q2, the body of the gain comes from the residual, and the gate and the summaries are corrections of the same small order. The revision describes it as an initialization detail rather than a design element that needs separate justification.

Q5 Training seeds. Three training seeds (2021, 7, 13), mask seed fixed at 2024, ETTh1, three backbones, four protocols, two horizons and two arms, 144 runs. For a fixed configuration the training seed moves the error by 0.02% to 2.29%, median 0.12%. Paired difference of the normalization-only arm from the baseline follows (%, negative means lower MSE; each cell gives horizon 96 then horizon 336).

BackboneP1 point-wiseP2F short blocksP3F correlated groupsP4 tail
CMoS+0.54 / +0.46−0.26 / +0.02−2.32 / −1.68−4.79 / −5.04
iTransformer+0.35 / +0.24−0.90 / −1.83−0.78 / −0.23−2.24 / −1.43
CycleNet+0.87 / +0.62+0.20 / +0.26−1.07 / −0.94−1.72 / −1.09

21 of 24 cells share a sign across the three seeds, and all seven cells above 1.5% do, with a maximum seed standard deviation of 0.92%. The three with mixed signs are CMoS short blocks at 336 (+0.02%, seeds −0.07, +0.04, +0.07), iTransformer correlated groups at 336 (−0.23%, −0.61, −0.41, +0.33) and iTransformer tail at 336 (−1.43%, −3.72, −2.26, +1.70); there the seed spread exceeds the effect and we do not rely on that cell.

Point-wise missingness is consistently positive across three backbones and three seeds, from +0.24% to +0.87%: the fill comes from adjacent observations while dropping them reduces the effective sample size, and the two offset. The direction is reproducible rather than seed noise.

We hope these clarifications address your concerns. We would be glad to provide further details if needed.

Thank you for the clarifications and additional controls, which resolve the injection-identity ambiguity and strengthen the evidence for observed-only normalization across training seeds on ETTh1. However, the confirmed P2/P3 generator issues substantially change the interpretation of the original cross-protocol comparisons and P3 strength sweep, while the corrected-protocol forecasting results presented remain limited to ETTh1. Moreover, the matched control in Q2 shows only approximately 0.03% incremental improvement from the mask summaries beyond the same normalization/projection wrapper, leaving the practical contribution of the proposed gated mask-conditioning design limited. I therefore maintain my score.

Reviewer_JBZ2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

MAFF addresses forecasting with missing values through observed-only normalization, gated missingness features, and a learnable token for fully missing channels. It aims to prevent imputed values from contaminating normalization statistics while informing the model about missingness.

Critical strengths

The paper identifies normalization statistics as a concrete pathway through which imputation affects forecasts and proposes a simple, parameter-free modification to that computation. Shared masks, component ablations, and reporting of negative results help characterize when mask-aware treatment is beneficial and where its effectiveness is limited.

Critical weaknesses
  1. Although the paper discusses missingness-aware forecasting methods such as S4M, GinAR, and CRIB, none is included in the evaluation.
  2. The evaluation is restricted to four ETT variants and Weather, providing limited domain diversity and relatively few channels. Missingness is artificially generated, without evaluation on naturally occurring gaps.
Questions for authors
  1. Could you justify restricting the evaluation to ETT and Weather, which have relatively few channels and limited domain diversity? Or could you provide results on larger or higher-dimensional benchmarks to strengthen the generality of the findings?

  2. Comparisons with specialized missingness-aware methods would help establish MAFF’s incremental contribution. Could you provide additional experimental comparisons with methods specifically designed for forecasting with missing values?

  3. The evaluation covers several missingness patterns, but the missing-ratio sweep is limited to ETTh1 with CMoS. Could you broaden this analysis?

2 comments

We thank the reviewer. The first and third points are addressed with new experiments; the second we do not attempt, and we explain why.

Q1 Domain and dimensionality. We ran the same component comparison on ECL (321 channels, hourly) with CMoS, three mask seeds and horizons 96 and 336, both arms sharing the same masks.

ProtocolNormalization-only arm vs baseline (%, negative means lower MSE)
P1 point-wise random+1.10%
P2F short blocks−0.36%
P3F correlated channel groups−2.95%
P4 window tail−6.68%

The ordering nearly coincides with 7-channel ETTh1: correlated groups −2.95% against −2.45%, tail −6.68% against −5.64%, point-wise slightly positive, short blocks within noise. This is not a coincidence. The instance-normalization statistics of a channel depend only on its own 96 time steps, so the channel count does not change how the contamination forms; it only changes how many channels a given mask covers. The mechanism should therefore transfer along the channel dimension, and ECL tests that prediction rather than providing one more sample.

ECL uses the corrected block and correlated-group protocols. The block budget of the original P2 grows with the channel count, and at 321 channels the realized missing fraction is 99.0% at horizon 96 and 99.7% at horizon 336, which leaves the window essentially empty; extending to higher dimensionality therefore requires the protocol fix first. On ETT, before the fix, they are 30.1%, 85.4%, 17.0% and 30.2% for P1 to P4. The scope of this paper is synthetically generated missingness.

Q2 Comparison with specialized methods. We do not run end-to-end comparisons against S4M, GinAR or CRIB, because the methods sit at different levels. Our claim concerns the normalization statistics: it adds no parameters, replaces no backbone and introduces no imputer, while S4M, GinAR and CRIB are complete models with their own architectures and training objectives. Placing them beside our four backbones in one table would mix the architectural difference with the wrapper difference and could not answer how large the part we claim is.

We instead provide four matched controls that answer exactly that question.

ControlPurposeResult
Normalization-only arm vs baselineidentical parameter count (937 against 937)bit-identical on clean data, maximum absolute difference 5.96e-8
DLinear negative controlthat backbone has no instance normalization, so the arm has nothing to act ondifference 0.0000 under all four protocols
Clean-data decompositionseparates the effect of the injection branchCMoS −5.56%, CycleNet −0.90%, iTransformer −0.10%, DLinear +0.38%
Zeroed mask featuressame parameters, same normalization, mask information removedall four protocols differ from the full model by 0.03%

These four controls give a definite decomposition. The injection branch adds a 96×96 learnable linear map (10,120 parameters active in the forward pass) and one scalar per channel to the normalized output; on clean data it reverses direction across backbones and crosses zero. It is not a general gain, and we do not count it as missingness awareness. The mask information contributes 0.03%, isolated by the zeroed-feature control, and the normalization-only arm adds no parameters and is bit-identical to the baseline on clean data. An end-to-end comparison is not equivalent to this decomposition, but it answers where the gain comes from, and the revision will state the boundary.

Q3 Scope of the strength sweep. The analysis now covers three backbones, three protocols and three ratios. The sweep in the submitted version reports a constant value for P3 because the ratio never reached mask generation. The corrected curves follow (ETTh1, three seeds, negative means lower MSE).

BackboneProtocolr=0.1r=0.3r=0.5
CMoSshort blocks+0.02%−0.19%−0.60%
CMoScorrelated long blocks−0.11%−2.45%−6.09%
CMoStail−0.72%−5.64%−12.32%
iTransformershort blocks−0.38%−1.10%−0.41%
iTransformercorrelated long blocks−0.36%−0.59%−0.37%
iTransformertail−0.17%−0.99%−1.46%
CycleNetshort blocks+0.34%+0.33%−0.06%
CycleNetcorrelated long blocks+0.23%−1.41%−3.38%
CycleNettail−0.26%−1.53%−4.07%

The three backbones share one shape: short gaps show no response, long gaps rise monotonically with strength, and the tail is strongest. The revision will replace the sweep of the submitted version with this table.

We hope these clarifications address your concerns. We would be glad to provide further details if needed.

Thank you for the detailed response and additional experiments. I appreciate the the efforts to address my concerns.

However, I still believe that the evaluation remains insufficiently comprehensive in terms of dataset diversity. While the additional experiments on ECL provide useful evidence, the evaluation is still limited to a relatively narrow range of datasets, making it difficult to assess the generalizability of the proposed method across different domains and real-world missingness scenarios.

Therefore, I will maintain my original rating.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.