acceptodds
Under review as a conference paper at ICLR 2027

Estimate What Repeats, Learn What Remains: Post-Hoc Correction of Frozen Time Series Forecasters

Abstract

Deep forecasters and pretrained foundation models have made time-series fore- casting far more accurate, yet the errors of a trained model are not unstructured noise. Whether a forecaster overpredicts or underpredicts often depends on where the target falls within the daily cycle and on how many steps ahead the forecast is made, and the same dependence is already present on the training data. Exist- ing post-hoc correctors lower the average error but leave much of this recurring pattern in place. We introduce LIRA (Lookup and Input-conditioned Residual Ad- justment), a correction framework for frozen forecasters that first estimates what repeats and then learns what remains. A recurring error lookup pools historical residuals by target cycle position and forecast lead, and a channel-shared linear head is trained on the residual left after the lookup. Because the lookup stores statistics rather than learned weights, newly observed forecast errors update the lookup alone, so online correction needs neither additional backbone calls nor gradient optimization. Across five task-specific backbones and four foundation models, online LIRA lowers horizon-averaged MSE in every backbone–dataset pair, by 8.77% and 17.64% on average, and matches or outperforms the post-hoc correctors δ-Adapter and PIR in most task-specific pairs, with lower correction computation and storage costs.

Primary area:
learning on time series and dynamical systems
Keywords:
Time Series, Forecasting

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_tkJcOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes LIRA, a post-hoc correction method for frozen time-series forecasting models. The key observation is that forecasting residuals contain recurring patterns associated with target cycle positions and forecast leads, which can persist after backbone training and transfer to future predictions. LIRA corrects forecasts through two components. A recurring error lookup estimates systematic residual patterns. A channel-shared linear head then predicts the remaining residuals from the observed history and original forecasts. The method also supports online adaptation by updating lookup statistics using newly observed residuals without modifying the backbone or remainder head. Experiments with five task-specific models and four pretrained foundation models report consistent forecasting improvements.

Critical strengths
  • Consistent empirical improvements across different models, including backbones & FMs.
  • A simple framework for post-hoc correction. LIRA separates residual correction into recurring error estimation and input-conditioned remainder prediction. It can be applied to both frozen backbones and FMs, and its online adaptation only requires updating historical statistics without gradient optimization.
Critical weaknesses

The insights of the two components are not sufficiently established.

The core idea of LIRA is that forecasting residuals remains learnable structures, and that these structures include two parts, one can be captured by recurring errors determined by channel and cycle position, the other one can be learned from the input window and the prediction. However, neither theoretical nor empirical explanations are well established for: why such structures can be learned by post-hoc designs but cannot be learned by the model itself. The paper tries to use Figure 6 to make the explanations, but the experimental design and analysis are not sufficient or clear. As for the lookup component, it mainly brings two sources of new information, channel id & calendar position. Appendix C.10 shows that the improvement of LIRA is much less for calendar-aware backbones, which partially support the importance of calendar position. More experiments are needed to confirm which design contributes to the improvements on these backbones. As we go deeper in Figure 6a, "lookup + raw head" is a weired setting, where the target of raw head is real error ( Y−Y^Y - \hat{Y} without lookup) but inference adds the output of raw head and lookup. The training output space and inference output space are different. More surprisingly, it brings a huge improvement over backbone, lookup only and raw head, which means it does not ruin the prediction, but combines the benefit of the two designs. And the gap to the full LIRA is quite small. Such results need explanation. The only valid explanation that occurs to me is that, the outputs / residual error space of lookup and raw head are orthogonal, which means the outputs will not interfere with each other. As for the head component, it needs explanation about why head component can learn information that the backbone itself fails to learn. The head component uses only the original input the backbone forecast, introducing no additional predictive information. Figure 6b further strengthen my question, as more than half the MSE gain can be recovered by adding a scalar bias learned by head component.

Questions for authors
  • Statistics on the learned α\alpha and γ\gamma.
  • Statistics on how orthogonal are between the output of lookup and the output of raw head / head component, as well as the scalar expectation of the two.
  • A series experiments with 3×2×23\times2\times2 settings.
    • Head input (3): input only, forecast only, input + forecast.
    • head output (2): vector, scalar offset only
    • training recipe (2): frozen backbone + learnable linear head, joint fine-tuning. For joint fine-tuning, the data flow keeps the same. Take forecast only + joint fine-tuning as an example, it is the same as appending a residual linear layer.
  • Analysis on the aforementioned results.
6 comments
AuthorsOpenReview ↗

Response to Reviewer tkJc (1/4)

Thank you for your thoughtful feedback and recognition of our consistent gains, simple post-hoc design, and online adaptation. We have added coefficient and output statistics and comparisons of head inputs and training strategies.

w3, q1, q2 — Combining the lookup and head, and remainder learning (1/2)

You asked why a raw-residual head works well with the lookup and why this combination performs nearly as well as LIRA. We should have more clearly distinguished why the corrections work together from what remainder learning adds.

1. Understanding the combined gains and the remaining performance gap

As you noted, the raw head is trained on r=Y−Y^r=Y-\hat Y, although inference combines it with the lookup. In Lookup + raw head, the channel–lead-group coefficients γ\gamma are fitted and weight decay is selected using validation error after lookup correction; the head's training target remains the raw residual. Even the naive sum, which keeps the head and coefficients from raw-head-only unchanged, improves on either correction alone.

1.1. Where does the performance difference arise?
MethodMean MSE gain (%)
Lookup only2.89
Raw head only4.17
Naive sum6.01
Lookup + raw head6.36
LIRA6.78

Let BB be the lookup correction, u=γPu=\gamma P the applied head correction, and GG the MSE gain (%) over the backbone.

G(B+u)=G(B)+G(u)−2⟨B,u⟩∥r∥2×100⏟C.G(B+u)=G(B)+G(u)-\underbrace{\frac{2\langle B,u\rangle}{\|r\|^2}\times100}_{C}.

All rows below share the lookup. G(u)G(u) applies each method's calibrated head alone, without refitting; CC is the interaction term in percentage points (pp). A positive CC subtracts from the sum of the standalone gains, while a negative CC adds to it. The inner product and norm are computed over all test origins, leads, and channels for each seed and setting then averaged.

MethodG(B)G(B) (%)G(u)G(u) (%)CC (pp)G(B+u)G(B+u) (%)
Naive sum2.8884.168+1.0436.013
Lookup + raw head2.8883.910+0.4396.359
LIRA2.8883.576−0.3186.782

Exact orthogonality in every run would yield 2.89 + 4.17 = 7.06% for the unchanged corrections. The naive sum retains 85% of this additive benchmark; the remaining 15% (about 1.04 pp) is the interaction penalty. It therefore still improves on either correction alone.

Post-lookup γ fitting and weight-decay selection reduce this penalty from 1.043 to 0.439 pp, raising the combined gain from 6.013% to 6.359%. With LIRA, CC becomes −0.318 pp. Compared with Lookup + raw head, its standalone head gain falls by 0.334 pp, but the interaction contribution improves by 0.757 pp, yielding 0.423 pp more combined gain. The interaction improvement outweighs the standalone loss. This is consistent with training the head for its role after lookup correction; the controlled comparison in Part 2 evaluates the target change.

1.2. What do the two corrections address, and why does the raw head also combine effectively with the lookup?

To examine your orthogonality hypothesis, we decompose each correction into its mean over the HH leads within each window and channel (offset) and the remainder (shape). These components are orthogonal by construction, so the lookup–head interaction splits exactly into offset and shape contributions.

StatisticLookup BBRaw head (γP\gamma P)Remainder head (γP\gamma P)
Offset share of correction energy (median, %)76964
Offset share of head gain (ratio of mean gains, %)–8675
Mean lookup–head cosine: offset / shape–+0.05 / +0.13−0.02 / −0.09

The lookup uses centered patterns and contributes mainly to shape, while most of the head's gain comes from window-specific offsets. These results support the view that the two corrections primarily address different error components.

q2. Output means and alignment with the lookup. We compute means and cosines over all test origins, leads, and channels for each seed and setting, then average across seeds and settings. The raw head is the model selected for Lookup + raw head.

StatisticRaw headRemainder head (LIRA)
Mean PP (before γ)0.0491480.050591
Mean γP\gamma P0.0080860.009772
Interaction term CC+0.44−0.32
cos⁡(B,P)\cos(B,P) before γ+0.12−0.12
Settings with ∣cos⁡(B,P)∣<0.1\lvert\cos(B,P)\rvert<0.1 / total60/10068/100

The lookup mean is approximately 0.000012; output means use standardized forecast units. The counts indicate weak alignment in many settings. (∣cos⁡(B,P)∣<0.1\lvert\cos(B,P)\rvert<0.1 in 60/100 and 68/100). Together with the offset/shape results, this helps explain why a raw head already captures much of the combined benefit. However, weak alignment can nevertheless leave a meaningful interaction contribution, whose effect on performance is quantified using the applied correction u=γPu=\gamma P above.

AuthorsOpenReview ↗

Response to Reviewer tkJc (2/4)

w3, q1, q2 — Combining the lookup and head, the performance gap, and remainder learning (2/2)

2. Remainder learning improves the combined shape correction

The raw target includes recurring errors that the lookup also corrects, whereas LIRA trains the head on the remainder, r−Boofr-B^{\mathrm{oof}}. To locate the additional gain, we decompose the difference in combined MSE gain into offset and shape contributions.

Entries are differences between LIRA and Lookup + raw head, in percentage points normalized by the original backbone MSE.

AggregationOffset gain difference (pp)Shape gain difference (pp)
Mean over 100 settings−0.06+0.49

The additional gain is concentrated in shape, with a small reduction in offset gain.

To separate the training-target effect from weight-decay selection, we compared the targets at each fixed weight decay, holding the backbone forecasts, lookup, α, initialization, batch order, and calibration procedure fixed. The table reports the mean relative MSE reduction over Lookup + raw head, using its MSE as the denominator. The last column selects weight decay separately for each target on validation data.

Weight decay00.0010.010.10.3Selected on validation
LIRA MSE gain (%)0.650.510.440.350.210.45

The mean gain is positive at every tested weight decay and decreases as regularization strengthens. Thus, the improvement cannot be attributed solely to weight-decay selection; changing the training target also contributes.

3. q1. Applied α and γ coefficients

The comparisons above use calibrated lookup and head corrections. To show how strongly each component is applied, we report the lookup coefficients α and head coefficients γ below.

These coefficients are fitted and shrunk on validation data, then fixed for testing. They are defined per channel and lead group, with upper bounds of 1.5 for α\alpha and 1 for γ\gamma.

CoefficientMeanMedian25th–75th percentilesZero coefficients (%)
α cycle0.810.850.56–1.095.1
α lead0.760.850.30–1.1011.9
α interaction0.860.900.57–1.207.1
γ raw head0.630.770.22–1.0010.2
γ remainder head0.630.760.25–1.007.3

The coefficients allow correction strength to vary across groups. A zero coefficient disables the corresponding component within that group.

Thank you for your careful review. We have revised RQ2 accordingly in the revised paper.

AuthorsOpenReview ↗

Response to Reviewer tkJc (3/4)

w1, w4, q3, q4 — What the head learns, and whether the backbone needs further training (1/2)

You asked why a lightweight head can recover errors that a more expressive backbone has not learned. Our experiments show that a head can correct predictable residual error with the backbone fixed, with much of its MSE gain coming from window-specific level corrections.

Why a post-hoc correction can help. Let XX denote the backbone's complete prediction-time input, Y^=f(X)\hat Y=f(X), and r=Y−Y^r=Y-\hat Y. For a correction u=u(Z)u=u(Z) based on prediction-time information ZZ, define m(Z)=E[r∣Z]m(Z)=\mathbb E[r\mid Z]. For each fixed channel and lead, expectations are over forecast windows with all fitted parameters held fixed:

$ \mathrm{MSE}(r)-\mathrm{MSE}(r-u) = \mathbb E[m(Z)^2]-\mathbb E[(u-m(Z))^2].$

The first term is the potential gain from predictable residual error; the second is the error of the learned correction relative to the ideal correction m(Z)m(Z).

  • Lookup. With ZZ comprising channel, cycle position, and lead, the lookup can provide information beyond XX when cycle position is not determined by XX. Even a predictor optimal for XX alone can then leave phase-dependent errors.
  • Head. With Z=(X,Y^)Z=(X,\hat Y) and Y^=f(X)\hat Y=f(X), the head adds no information beyond XX, but m(Z)=E[Y∣X]−f(X)m(Z)=\mathbb E[Y\mid X]-f(X) can remain nonzero. If f(X)=E[Y∣X]f(X)=\mathbb E[Y\mid X], no correction using these inputs can reduce expected MSE. A positive expected head-only gain therefore requires input-predictable residual error.

Even when a backbone can represent such corrections, training on finite data does not guarantee that its fitted forecasts equal E[Y∣X]\mathbb E[Y\mid X] on future windows. The observed head-only gains provide empirical evidence that some remaining error is recoverable with a frozen backbone, although they do not isolate its specific cause.

For the head's additional gain after lookup correction BB, the same identity applies to r−Br-B, with mB(Z)=E[r−B∣Z]m_B(Z)=\mathbb E[r-B\mid Z].

The role of the scalar offset. The scalar component in Figure 6b (revised Table 7) shifts all leads by an offset that varies across windows and channels. Its contribution indicates that much of the head's MSE gain comes from correcting forecast levels. Part 2 separately evaluates heads trained directly to predict these offsets.

Further backbone training. Part 2 isolates the head's contribution and compares frozen with joint training. For DLinear, an input-only linear head with channel-specific calibration also improves forecasts. Here, we separately compare complete LIRA with backbone-only fine-tuning of the same checkpoints using training and validation targets. The evaluation covers 35 dataset–backbone combinations. Entries report mean MSE gains (%) relative to the original backbone; † denotes online updates.

BackboneFine-tuningFine-tuning†LIRALIRA†
PatchTST0.700.566.157.77
iTransformer0.390.645.366.70
DLinear5.937.6610.4511.72
TimeMixer0.45−0.824.005.80
TimeBase0.200.2610.7611.86
Mean1.531.667.348.77

Under these procedures, complete LIRA yields larger average gains than backbone-only fine-tuning for every listed backbone.

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

The paper proposes LIRA, a post-hoc corrector for frozen point forecasters. It contains two parts: Error lookup and Remainder head. In online mode, the lookup statistics are updated by EMA once a forecast's full target window has been observed. The backbone, head and coefficients stay fixed, so no gradients or extra backbone calls are needed.

Critical strengths

Simple, cheap and black-box, with consistent gains and careful handling of leakage.

The diagnostic analyses go beyond leaderboard numbers. The half-day-shift control (Fig. 4) shows the gain really depends on correct cycle alignment. The offset/shape split (Fig. 6b) and the remainder-target ablation (Fig. 6a) make clear what each component contributes.

Critical weaknesses

Much of the gain likely comes from the experimental setup, not from removing residual structure in general:

All experiments use L = 96. For 10-min data (Weather, Solar) one day is S = 144 steps, so the backbone cannot see a full daily cycle. For 15-min data (ETTm) the lookback covers exactly one day. In these cases the cycle-position lookup gives the backbone information it does not have, i.e. absolute time of day. This would also explain why a model trained to minimize training loss still shows cycle-dependent bias on the training split, which the paper uses as evidence that the pattern is not caused by distribution shift.

The largest gains are on Solar (up to 31%), ETTm1 and Weather. Gains on hourly ETTh1 (S = 24 < L) and ECL are small.

Gains also shrink with a longer lookback: average 6.78% → 4.63% at L = 336, and on Weather 10.7% → 4.8%. They are small for CycleNet on ETTh1 and ECL (0.8%, 1.7%).

The foundation-model result looks like the same effect, only larger. Uncorrected ETTm1 MSE is 0.616 (Chronos-2), 0.834 (Toto-2.0) and 0.875 (TimesFM-2.5), far worse than the ≈0.38–0.43 of the task-specific backbones. These models are normally run with much longer contexts, and with a 96-step (one-day) context they are clearly handicapped. The 17.64% average and the claim that corrected foundation models beat uncorrected task-specific backbones therefore mainly show that an underconfigured setup can be repaired. They say little about how useful LIRA is for properly configured models.

Conceptually, the lookup is a lead- and time-of-day-dependent bias correction, very close to classical MOS/NWP post-processing, and the head is a linear residual model. The paper does not test how much LIRA adds once the backbone has enough context or calendar information.

Questions for authors

Please report gains as a function of L/S. Specifically:(Could change my assessment.) Weather and Solar with L ≥ 2S; foundation models at their recommended context lengths (e.g. ≥ 512–2048). If LIRA's gains stay substantial when the backbone can see several cycles, my main concern would be largely resolved.

Please add stronger simple baselines to the main tables:(Could change my assessment.) a cycle×lead mean-residual subtraction with no calibration; the "joint calendar + context" linear head from Fig. 6a; for task-specific backbones, the same backbone retrained with time-of-day inputs, beyond TiDE/TimeXer on three datasets.

Why do foundation models score so poorly on ETTm1/ETTm2 here? How does this compare with results these models report in their own papers?

6 comments
AuthorsOpenReview ↗

Response to Reviewer gZKS (1/5)

Thank you for your careful review and constructive feedback. We appreciate your positive assessment of the simple, low-cost black-box correction, the evaluation without leakage, and the component analyses. To address your concerns about context length and temporal information, we added longer-lookback experiments and comparisons with simple baselines.

w1–w4, q1, q3 — Gains across L/S and longer FM contexts (1/2)

We evaluated the longer lookbacks you requested. Average MSE gains persist as context increases for task-specific backbones on Weather and Solar, and for foundation models on ETTm1, ETTm2, Weather, and Solar.

Task-specific backbones: Weather and Solar. We retrained the backbones with L=336 and 720 and gave the LIRA head the same observation length. The validation and test target intervals were unchanged. Scores are averaged over horizons and then backbones; gains in parentheses are computed against the backbone at the same L before averaging.

DatasetLL/SBackbone MSELookup-only† MSE (gain %)LIRA MSE (gain %)LIRA† MSE (gain %)Backbone MAELIRA MAE (gain %)LIRA† MAE (gain %)
Weather960.670.2560.247 (3.37)0.240 (6.44)0.233 (8.97)0.2860.283 (0.89)0.273 (4.44)
Weather3362.330.2340.232 (0.70)0.225 (3.62)0.224 (4.18)0.2740.271 (1.04)0.269 (1.96)
Weather7205.000.2360.235 (0.23)0.231 (2.13)0.230 (2.31)0.2780.277 (0.44)0.276 (0.86)
Solar960.670.2890.250 (12.42)0.249 (11.84)0.237 (16.14)0.3280.303 (7.00)0.291 (10.75)
Solar3362.330.2380.231 (3.07)0.228 (4.08)0.227 (4.44)0.2880.275 (4.39)0.273 (4.97)
Solar7205.000.2150.210 (2.18)0.209 (2.82)0.209 (2.94)0.2750.263 (4.43)0.262 (4.73)

Even at L≥2S, horizon-averaged MSE improves for every evaluated backbone on Weather and Solar.

AuthorsOpenReview ↗

Response to Reviewer gZKS (2/5)

w1–w4, q1, q3 — Gains across L/S and longer FM contexts (2/2)

Foundation models with longer lookbacks. We extended the context to L=512, 1024, and 2048 while keeping the other inference conditions fixed. Each cell reports horizon-averaged MSE as Backbone → Backbone + LIRA†, with the gain relative to the backbone at the same context length in parentheses.

Dataset (S)BackboneL=96: Backbone MSE → LIRA† MSE (gain %)L=512: Backbone MSE → LIRA† MSE (gain %)L=1024: Backbone MSE → LIRA† MSE (gain %)L=2048: Backbone MSE → LIRA† MSE (gain %)
ETTm1 (96)Chronos-20.616 → 0.375 (+39.0)0.387 → 0.349 (+9.8)0.382 → 0.351 (+8.3)0.350 → 0.334 (+4.6)
Sundial0.538 → 0.364 (+32.2)0.366 → 0.353 (+3.6)0.356 → 0.349 (+1.8)0.329 → 0.331 (−0.5)
TimesFM-2.50.875 → 0.409 (+53.3)0.381 → 0.349 (+8.5)0.390 → 0.361 (+7.3)0.386 → 0.358 (+7.1)
Toto-2.00.834 → 0.406 (+51.3)0.406 → 0.355 (+12.7)0.404 → 0.355 (+12.2)0.373 → 0.350 (+6.1)
ETTm2 (96)Chronos-20.318 → 0.274 (+13.7)0.280 → 0.264 (+5.7)0.261 → 0.251 (+3.5)0.252 → 0.247 (+1.8)
Sundial0.304 → 0.272 (+10.5)0.275 → 0.257 (+6.5)0.261 → 0.250 (+4.6)0.240 → 0.236 (+1.4)
TimesFM-2.50.372 → 0.276 (+25.7)0.291 → 0.266 (+8.5)0.279 → 0.264 (+5.3)0.276 → 0.261 (+5.3)
Toto-2.00.332 → 0.270 (+18.7)0.285 → 0.261 (+8.5)0.271 → 0.259 (+4.3)0.255 → 0.250 (+2.1)
Weather (144)Chronos-20.307 → 0.244 (+20.4)0.258 → 0.224 (+13.1)0.236 → 0.222 (+6.1)0.226 → 0.214 (+5.5)
Sundial0.285 → 0.247 (+13.2)0.244 → 0.221 (+9.7)0.234 → 0.220 (+5.8)0.221 → 0.215 (+2.4)
TimesFM-2.50.316 → 0.244 (+22.6)0.263 → 0.221 (+16.0)0.248 → 0.219 (+11.7)0.244 → 0.217 (+11.0)
Toto-2.00.316 → 0.251 (+20.6)0.246 → 0.219 (+10.9)0.232 → 0.219 (+5.8)0.219 → 0.211 (+3.3)
Solar (144)Chronos-20.529 → 0.285 (+46.1)0.187 → 0.180 (+3.6)0.178 → 0.170 (+4.3)0.171 → 0.165 (+3.5)
Sundial0.666 → 0.247 (+62.9)0.229 → 0.217 (+5.6)0.182 → 0.180 (+0.9)0.173 → 0.172 (+0.7)
TimesFM-2.50.913 → 0.215 (+76.4)0.281 → 0.234 (+16.7)0.250 → 0.223 (+10.7)0.234 → 0.218 (+6.8)
Toto-2.01.090 → 0.269 (+75.3)0.318 → 0.225 (+29.2)0.257 → 0.212 (+17.5)0.222 → 0.205 (+7.5)

Average MSE gains remain on each of the four datasets even at the longest context. Thus, correction remains useful after the backbone has observed multiple cycles.

The tendency for gains to decrease with longer context is consistent with the backbone using more periodic information and leaving less error to correct. The variation across datasets and the CycleNet results also suggest that the remaining recurring error determines how much additional gain is available.

We agree that, with short contexts, the lookup can supply temporal information unavailable to the backbone. Providing calendar inputs does not, however, eliminate its contribution. On TiDE and TimeXer, which receive calendar features, a lookup estimated from training residuals still lowers test MSE (see our response to tkJc (4/4), w2). Its contribution becomes much smaller when the backbone directly learns channel-specific periodic output corrections, as in CycleNet. LIRA still provides additional gains in those comparisons.

Bias already present during training. We agree that missing phase information can account for some of this bias when L≤S. Yet recurring residuals also remain in backbones with calendar inputs and carry over to the test interval. Lookup gains persist on Solar at L=720, where the backbone observes several cycles. These observations indicate that neither distribution shift nor short context alone accounts for the residual structure.

Thank you for your careful review. We have included the results in the main section of the revised paper.

AuthorsOpenReview ↗

Response to Reviewer gZKS (3/5)

q3, w4 — FM performance and scope of the claims

q3. FM performance on ETTm1 and ETTm2. Short context is a major contributor. Extending the context alone, with the same checkpoints and inference procedure, substantially lowers the uncorrected MSE. For Sundial-Base, whose published benchmark permits a direct comparison, the resulting MSE is below the reported values. The other three models report different benchmarks or metrics, so we could not make the same comparison.

EvaluationETTm1 MSEETTm2 MSEWeather MSE
Sundial-Base, Table 1 of its paper0.3360.2580.234
Sundial-Base, our evaluation at L=20480.3290.2400.221

w4. Scope of the claims. L=96 was a common setting aligned with the task-specific evaluation protocol. We now state that the reported performance is based on a 96-step look-back, and we added the results across L/S to the main text (revised Tables 2 and 3).

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

The paper proposes LIRA, a post-hoc correction method for frozen time-series forecasters. It first estimates recurring errors by target cycle position and forecast lead, then trains a channel-shared linear head to predict the remaining residual from the observation history and original forecast. Head training uses out-of-fold lookup corrections with overlapping targets excluded. Online adaptation updates only residual statistics after complete target windows become available. Experiments cover five task-specific backbones on seven datasets and four foundation models on six datasets, reporting average horizon-averaged MSE reductions of 8.77% and 17.64%, respectively. Ablations examine lookup transfer, remainder learning, history requirements, and online updates.

Critical strengths

LIRA provides a simple, interpretable way to exploit systematic forecast errors. Indexing residuals by both cycle position and forecast lead distinguishes predictions that target the same time but have different biases. The forecast-only interface and gradient-free online updates make the method useful when the underlying model cannot be modified.

The evaluation addresses several plausible explanations for the gains. The phase-shift control supports the importance of cycle alignment, while experiments with longer lookbacks, CycleNet, and calendar-informed backbones show that improvements extend beyond models lacking temporal information. The consistency across backbone families strengthens the practical contribution.

Critical weaknesses

The benefit of staged estimation is not fully separated from differences in model capacity. LIRA uses channel-specific cycle–lead tables and calibration coefficients, but Appendix D.2 does not establish whether the joint calendar-plus-context baseline has equivalent channel-specific calendar parameters and calibration. If its calendar mapping is shared across channels, it cannot represent the same residual patterns. Figure 6 therefore does not conclusively establish an advantage over a comparably expressive joint model.

The foundation-model gains need stronger attribution. The linear head receives both observations and forecasts and can learn a substantial supervised predictor of its own. For example, TimesFM-2.5 on ETTm1 improves from 0.875 to 0.409 MSE, while uncorrected DLinear achieves 0.404. This does not invalidate the improvement, but it leaves unclear how much comes from exploiting foundation-model predictions versus the added supervised model. The component ablations are limited to task-specific backbones.

Questions for authors

Does the joint calendar-plus-context baseline have channel-specific calendar parameters and the same calibration and selection procedure as LIRA? Does the two-stage approach retain its advantage when these factors are matched?

On the foundation-model benchmarks, how does full LIRA compare with lookup-only correction, head-only correction, and a linear predictor trained directly from the observation history using the same data? These comparisons would clarify the contribution of both the lookup and the frozen foundation model.

3 comments
AuthorsOpenReview ↗

Response to Reviewer hQyu (1/2)

Thank you for your careful reading and helpful suggestions. We appreciate your recognition of the simple, interpretable use of recurring errors and the ability to update the correction without modifying the model or computing gradients.

w1, q1 — Joint baseline capacity, calibration, and selection

We agree that the original joint baseline was not matched to LIRA's capacity: its calendar weights were shared across channels. Figure 6a (first version paper) therefore did not establish an advantage over a joint model with the same capacity. Calibration and validation selection followed the same procedure as the LIRA head, with no test data used for selection.

For a stronger comparison, we jointly trained the head and a channel-specific cycle–lead table with the same indexing as LIRA's lookup, using the same training and model-selection procedure as the LIRA head. We compared a shared calibration coefficient with separate coefficients for the calendar and head components. The separate coefficients were fitted together on the same validation data used by LIRA.

All methods below were evaluated on the same 100 settings. Only rows marked † use online updates.

MethodMean MSE gain (%)Mean MAE gain (%)
Joint, shared calendar across channels (Figure 6a)4.752.35
Joint, channel-specific calendar, shared γ (same calibration as Figure 6a)6.123.13
Joint, channel-specific calendar, separate calendar/context coefficients (revised Table 10)6.532.63
LIRA6.783.38
LIRA†8.014.84

LIRA retains larger mean MSE and MAE gains against the joint model with matched capacity. Updating the lookup with new observations further lowers the error without gradient updates.

We added the stronger joint comparison to “Separate vs. joint training” in the revised RQ2 (Table 10).

AuthorsOpenReview ↗

Response to Reviewer hQyu (2/2)

w2, q2 — Contributions of LIRA's components and the foundation-model forecasts

We compared lookup-only, head-only, and full LIRA on the same frozen foundation-model forecasts. The head-only variant learns the original residual without a lookup; † denotes online lookup updates. The table uses L=96, the setting of the foundation-model results, and averages test MSE over four horizons and then across the six datasets.

ModelBackbone MSELookup-only† MSEHead-only MSEFull LIRA† MSE
Chronos-20.3950.3590.3310.323
Sundial0.3670.3410.3270.318
TimesFM-2.50.4500.3780.3320.328
Toto-2.00.4420.3650.3310.329
Overall (24 combinations)0.4140.3610.3300.324

Full LIRA has lower mean MSE than either component alone for all four models.

Contribution of the foundation-model forecast: comparison with direct linear prediction. We also trained a linear layer to predict the target directly from the observed history, using the same training procedure, validation selection, and data cutoffs. At L=96 the foundation models are constrained by the short context, so we compare at L=512, closer to their recommended length. The table reports MSE averaged over four horizons; values below the direct predictor are bold.

Dataset (L=512)Direct linear MSE (same training procedure)Chronos-2 MSE (uncorrected → + LIRA†)Sundial MSE (uncorrected → + LIRA†)TimesFM-2.5 MSE (uncorrected → + LIRA†)Toto-2.0 MSE (uncorrected → + LIRA†)
ETTm10.3580.387 → 0.3490.366 → 0.3530.381 → 0.3490.406 → 0.355
ETTm20.2670.280 → 0.2640.275 → 0.2570.291 → 0.2660.285 → 0.261
Weather0.2400.258 → 0.2240.244 → 0.2210.263 → 0.2210.246 → 0.219

With LIRA, every foundation model in the table outperforms direct linear prediction on each dataset. At L=96, FM + LIRA† is below the direct predictor in 20 of 24 dataset–model pairs. Together with the component comparison, this supports using foundation-model forecasts with the lookup and remainder head.

Thank you again for your thoughtful and valuable feedback.

As the review deadline approaches, we would be grateful if you could revisit our response. We have also revised the main paper to reflect your comments and appreciate your consideration of these updates.

Please let us know if you have any further questions or concerns. We would be happy to provide any additional clarification.

Thank you for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.