acceptodds
Under review as a conference paper at ICLR 2027

Lift Out, Kick In: Decoupling Event Effects for Time Series Forecasting

Abstract

Real-world time series are often shaped by exogenous events whose effects are superimposed on the system’s underlying temporal dynamics. Historical observations therefore entangle persistent dynamics with event-induced deviations, making it difficult to determine which patterns should be extrapolated and which should be attributed to specific event contexts. Existing event-aware forecasting methods can incorporate event information or model event-related variation, but typically treat historical event attribution and future event generation separately. We propose Loki (Lift Out, Kick In), a general event-aware forecasting framework built on a shared Event Response Function (ERF). The ERF models event effects according to event type and relative temporal position and is shared across the historical and forecast windows. Loki lifts out estimated event-induced deviations from historical observations before forecasting, then uses the same ERF to kick in responses to known future events. This formulation separates event-response modeling from temporal forecasting while coupling historical attribution and future generation through a shared mechanism. Across five forecasting benchmarks, Loki achieves the best standardized MAE on all five datasets, reducing error by approximately 2.8%–6.0% over the strongest non-Loki baseline. Ablations and controlled analyses further support the roles of Lift Out, Kick In, shared response modeling, and compatibility across forecasting backbones.

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

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_RybnOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates how to disentangle event effects from underlying temporal dynamics in event-aware time series forecasting, while also predicting the impact of known future events. The motivation stems from the fact that transient changes triggered by historical events may be erroneously propagated into the future by a model, a form of confusion that cannot be easily resolved by merely incorporating event information as input. To address this, we employ a shared event-response function to strip away the effects of historical events prior to forecasting, and subsequently inject the responses of future events into the baseline predictions. Experimental results demonstrate that this approach achieves the lowest normalized mean absolute error across five benchmark datasets.

Critical strengths

1.The study addresses the issue of the erroneous extrapolation of historical event effects, with a clear research motivation.

2.It employs a unified approach, using shared event responses for both historical disentanglement and future injection resulting in a concise structure.

3.The effectiveness of the method is validated across five benchmarks, and key design components are examined through ablation analysis.

Critical weaknesses
  1. Existing research has already addressed issues such as the erroneous extrapolation of historical event effects and the necessity of disentangling routine dynamics from event-driven effects. The present paper fails to clearly articulate what novel contributions it offers beyond these established approaches. Furthermore, the authors do not explicitly demonstrate the systematic prediction biases that arise in existing event-conditioned models following the conclusion of an event, nor do they adequately distinguish between genuine "tail effects" (lingering impacts) of an event and the erroneous extrapolation of event effects. Consequently, a mere reduction in overall prediction error is insufficient to justify the motivation behind the proposed method.

  2. The fact that historical and future events belong to the same underlying process does not automatically imply that they share the same response structure. Moreover, the relative impact of an event may fluctuate based on demand conditions, weather, or operational circumstances; however, the response functions in this paper rely primarily on event type, intensity, relative timing, and the specific entity involved, offering limited capacity to capture context-dependency or synergistic interactions between multiple events. Additionally, because the model is trained solely based on the final prediction loss, the backbone and the event-specific branches can compensate for one another, which does not guarantee that the historical adjustments actually correspond to the intended disentanglement of event effects.

  3. By halting gradients when an event is active, the model effectively forces the event-specific branches to absorb prediction errors, errors that may, in fact, stem from poorly modeled "routine" (non-event) dynamics. The current experiments do not sufficiently disentangle the respective contributions of the shared structural components versus the gradient-halting mechanism. Furthermore, while the method accounts for "pre-event" effects and "post-event" tail effects, the gradient-halting mechanism is triggered strictly by the binary status of the event indicator.

  4. The primary experiments focus on comparisons with general-purpose predictors, lacking benchmarks against the most relevant existing methods, namely, those designed for event decomposition or historical filtering, as well as a baseline that employs a simple response lookup table (using the same backbone and training protocol). It is therefore unclear whether the observed performance gains stem from the decomposition structure itself or from the complex parameterization of the response functions. Additionally, existing comparisons between shared and independent response components fail to adequately control for differences in the number of parameters or the regularization strategies employed.

  5. The semi-synthetic experiments rely on favorable conditions, such as stable multiplicative responses, independent scheduling, and non-overlapping events, thereby failing to rigorously test the model’s performance under scenarios involving context-dependency, overlapping events, or mismatches in the assumed response structure. Moreover, the analysis of event contexts defines windows based solely on whether an event indicator is active; however, since "pre-event" or "post-event" (tail) responses can persist even when the indicator is zero, this approach cannot accurately isolate genuine event effects from erroneous extrapolations.

Questions for authors

see weaknesses

8 comments
AuthorsOpenReview ↗

Rebuttal Part 1

W1. Existing research has already addressed issues ..

(i) Technical advance and practical benefits.

Loki is a simple, jointly trainable event pre- and post-processing module: the backbone learns routine dynamics from event-adjusted history, and responses are restored from future event schedules. It is compatible with different backbones and, with a lightweight structure, achieves the lowest MAE on all five benchmarks; on Ride-Hailing it uses only 0.27M parameters, about 86% below the median of the 15 neural baselines (Appendix C.1). It addresses three connected needs:

  • Historical adjustment should serve forecasting. PHR and SPADE remove historical events by predefined rules, and CEDAR trains its event branch in a separate stage, so none adjusts history by its effect on forecasting. Loki learns the adjustment from the final forecasting error, so it is optimized for its effect on forecasting. Future-only and Direct Fusion are worse on all five datasets (Table 2).
  • Responses should vary with event conditions. Dynamic regression, Prophet and NeuralProphet fit one coefficient per event and offset, so responses do not vary with event intensity. Loki's ERF generates responses from intensity, relative position and entity through multiscale temporal bases, covering anticipation and carryover. An event × offset lookup table is worse in every seed on four datasets (Table 2).
  • The two sides should constrain each other. NBEATSx generates backcast and forecast coefficients separately, so the two sides can explain event effects independently. Loki shares one ERF: the rule learned from observed future effects also sets the historical adjustment, which has no direct supervision since historical event effects are never observed. In semi-synthetic experiments, sharing lowers historical MAE from 0.0294 to 0.0207 (Ride-Hailing) and from 0.0272 to 0.0118 (M5) (Appendix E.4).

Loki also has lower MAE than the added SPADE, CEDAR, AutoARIMA-X, log-linear, Prophet, NeuralProphet and NBEATSx baselines on all five datasets (Table 1).

(ii) Erroneous extrapolation in existing event-conditioned models.

  • Separating genuine tails from erroneous extrapolation. Event end does not imply effect end. We select windows where a historical event has ended and demand has largely recovered: relative to a non-event reference for the same entity and time of week, the post-event observed deviation is at most half the historical deviation. This allows residual changes; we count forecasts that extend the historical anomaly beyond the observed residual. Agreement in at least 4/5 seeds is required for stochastic models; deterministic models use their single result.

  • Rate in model outputs. We apply the same screening to all five datasets. M5 has no test window with the relevant event ending before the forecast origin; Seoul Bike has only one eligible sample. We therefore report the other three datasets. All rates and replacement counts: Appendix D.10.

DatasetLokiMedian of the other 18 modelsLowest other model
Ride-Hailing14.4%20.7%7.7% (TFT)
TaxiBJ13.9%94.9%0.1% (AutoARIMA-X)
Rossmann8.0%10.2%5.1% (CEDAR)

On TaxiBJ, most neural models, including TimeXer, TiDE and DAG, exceed 93%. All diagnostic windows come from one post-Qingming forecast date (2016-04-05).

  • Is the bias related to historical event inputs? Other factors may also cause this bias. We freeze the models and all future inputs, replacing only historical event-period demand with a normal reference built from pre-origin observations. Two replacements at non-event positions serve as controls. A sample supports erroneous continuation if same-direction bias, MAE and MSE decrease in at least 4/5 seeds and the MAE gain exceeds both controls. Counts of qualifying window–target pairs:
Dataset (window–target pairs)DAGTiDETFTLoki
Ride-Hailing (67)6102
TaxiBJ (63)181500
Rossmann (32)0200

For qualifying samples, replacing event-period observations reduces bias more effectively than replacing non-event positions, supporting historical event inputs as a contributor to erroneous extrapolation. This evidence comes from controlled input checks, not overall accuracy differences; the normal reference is not causal ground truth for event effects.

(iii) Loki under the same diagnostic. On TaxiBJ, Loki's rate (13.9%) is far below most neural models and no pair meets the replacement criterion. On Ride-Hailing and Rossmann, Loki ranks 2nd and 5th of 19 models, comparable to the better models, so it reduces but does not eliminate the problem. AutoARIMA-X and Prophet show low rates on TaxiBJ but much higher overall error (0.1718 and 0.2040 vs 0.1128). Post-event 24-hour MAE ranks Loki 1st on Ride-Hailing and Rossmann and 3rd on TaxiBJ; Five-dataset MAE and ranks: Appendix D.10.

AuthorsOpenReview ↗

Rebuttal Part 2

W2. The fact that historical and future … does not automatically imply that they share the same response structure … limited capacity to capture context-dependency or synergistic interactions between multiple events … the backbone and the event-specific branches can compensate for one another..

Thank you for raising these concerns.

(i) Sharing a response function, not a fixed response. The two sides share one mapping from event conditions to responses rather than one response value: the ERF varies with event intensity, relative timing and entity, so each occurrence of the same event type can have a different response. If historical and future responses followed systematically different rules, learning separate ERFs for the two sides should perform better; the experiments show otherwise:

Data typeDatasetMetricSharedIndependentHalf-independent
Real dataRide-HailingPrediction MAE0.156960.157170.15691
Real dataM5Prediction MAE0.095830.098250.09845
Semi-syntheticRide-HailingHistorical MAE after adjustment0.02070.02940.0299
Semi-syntheticM5Historical MAE after adjustment0.01180.02720.0277

Half-independent matches the parameter count of Shared. Sharing gives the best prediction on M5 and is on par on Ride-Hailing, and it clearly improves historical adjustment on both semi-synthetic datasets. Real-data parameter controls: Appendix D.2; semi-synthetic recovery: E.4.

(ii) Variation of relative impact with context. Continuous environmental variables, such as temperature, modulate the baseline through FiLM; localized weather events, such as rain and snow, enter the ERF as event inputs. The ERF does not additionally condition its response on recent demand levels, limiting its ability to represent this form of state dependence. In semi-synthetic experiments where the injected response depends on the recent demand level, Loki reduces future effect error relative to the zero-response reference by 70% (Ride-Hailing) and 79% (M5), compared with 93% and 94% under stable effects. We will state this limitation in the revised paper and leave context-conditioned response modeling to future work. Injection and full results: Appendix E.5.

(iii) Interaction among events. Concurrent events add in log space, which captures compounding impacts. Semi-synthetic future effect MAE for the combined response:

DatasetOverlap, no interactionOverlap, with interactionNo interaction, training-RMS matched
Ride-Hailing0.02470.04380.0253
M50.01640.03390.0167

An explicit interaction term raises the error, and the gap remains after matching the training perturbation RMS, so the multiplicative combination captures only part of the interaction. Injection and training-magnitude control: Appendix E.5.

(iv) Compensation between backbone and ERF.

  • Gradient mechanism. The same forecasting loss backpropagates along three paths:
PathBackpropagation order
Future ERFLoss → final forecast → future response
BackboneLoss → final forecast → baseline forecast → Backbone
Historical ERFLoss → final forecast → baseline forecast → Backbone → adjusted history → historical response

The future-ERF path is the shortest, so the ERF learns the response rule mainly on the future side, and sharing applies it to the historical adjustment. The backbone keeps its gradients at all positions, so it learns routine dynamics from errors at event positions instead of leaving them to the ERF.

  • Input constraint. The ERF acts only within the event-response support, including anticipation and carryover. It generates responses solely from event information, limiting its freedom to compensate for the backbone's errors on routine variation.
  • Experimental evidence. On real data, the counterfactual without the event is unobservable, so we injected known event effects into real series:

Historical MAE follows before → after adjustment; future effect MAE follows zero response → multiplicative structure.

DatasetInjection conditionHistorical MAEFuture effect MAE
Ride-HailingMultiplicative effects0.3333 → 0.02070.3284 → 0.0228
Ride-HailingEvents scheduled, no effect injected0 → 0.00970 → 0.0082
M5Multiplicative effects0.1927 → 0.01180.2493 → 0.0155
M5Events scheduled, no effect injected0 → 0.00520 → 0.0062

With an injected effect, Lift Out reduces the historical error by 93.8% (Ride-Hailing) and 93.9% (M5), and Kick In reduces the future-increment error by 93.1% and 93.8%, so the effect is captured by the ERF rather than absorbed by the backbone. In the no-effect control, both outputs remain close to zero, so the ERF rarely attributes routine dynamics to events. Settings, metrics and full results: Appendix E.1–E.4.

AuthorsOpenReview ↗

Rebuttal Part 3

W3.By halting gradients when an event is active, the model effectively forces the event-specific branches to absorb prediction errors, errors that may, in fact, stem from poorly modeled "routine" (non-event) dynamics. The current experiments do not sufficiently disentangle the respective contributions of the shared structural components versus the gradient-halting mechanism. Furthermore, while the method accounts for "pre-event" effects and "post-event" tail effects, the gradient-halting mechanism is triggered strictly by the binary status of the event indicator.

Thank you for pointing out that SG may force the event branch to absorb routine errors, that its contribution was not separated from the shared structure, and that it is triggered by binary event indicators. We have removed SG from the revised Loki; below, we compare SG on and off to explain this revision and examine whether the structural gains remain without SG.

(i) Why SG can make the event branch absorb routine errors. SG was intended to let the backbone learn routine variation and the ERF learn event responses. However, event positions may also contain baseline forecasting errors; SG prevents the backbone from correcting them while still letting the ERF update on them, so the event response can absorb baseline errors, as the reviewer points out. The same forecasting loss backpropagates along three paths; the last column shows the change when SG is enabled at event-active forecast positions in the submitted version:

PathBackpropagation orderWith SG
Future ERFLoss → final forecast → future responseKept
BackboneLoss → final forecast → baseline forecast → BackboneBlocked
Historical ERFLoss → final forecast → baseline forecast → Backbone → adjusted history → historical responseBlocked

Because the blocking is set by the binary event indicator, it also does not cover anticipation and carryover positions. Without SG, the backbone and the ERF learn from the same loss at all positions.

(ii) Basis for removing SG. The table reports standardized prediction MAE (multi-seed mean ± sample SD) with SG off and on:

DatasetSG offSG on
Ride-Hailing0.1570 ± 0.00030.1573 ± 0.0003
TaxiBJ0.1128 ± 0.00160.1120 ± 0.0017
M50.0958 ± 0.00110.0961 ± 0.0010
Rossmann0.1365 ± 0.00280.1538 ± 0.0036
Seoul Bike0.3238 ± 0.02080.3228 ± 0.0257

Removing SG lowers MAE on Ride-Hailing, M5 and Rossmann, while the TaxiBJ and Seoul Bike differences are within one seed SD. In semi-synthetic experiments under matched multiplicative injection, we fixed the shared multiplicative architecture, injected known multiplicative effects and varied only SG. Prediction and recovery metrics use the same valid positions within response support, including anticipation and carryover. Enabling SG does not improve the decomposition either:

All metrics are standardized MAE; each cell follows the order SG off → on.

DatasetResponse-support prediction MAEHistorical MAE after adjustmentFuture effect MAE
Ride-Hailing0.1694 → 0.16970.0207 → 0.02080.0228 → 0.0230
M50.1128 → 0.11570.0118 → 0.01340.0155 → 0.0179

(iii) Structural gains do not depend on SG. In the final version without SG, full Loki still outperforms Direct Fusion, which feeds events directly into the forecast (Ride-Hailing 0.15696 vs 0.16124, M5 0.09583 vs 0.10257), and the shared ERF still outperforms separate ERFs for history and future (M5 0.09583 vs 0.09825). The gains of Lift Out/Kick In and sharing therefore do not come from SG.

The complete structural comparisons without SG are reported in revised Table 2; parameter-budget comparisons of shared and independent ERFs are provided in Appendix D.2.

Reviewer_yGkmOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Loki, an event-aware forecasting framework that uses a shared Event Response Function (ERF) to model historical and future event effects. It removes the estimated event response from the observed history before forecasting and applies the same response model to known future events afterward. The ERF represents event type, relative position, temporal anticipation/carryover, and unit-specific modulation. Across five datasets, Loki obtains the best MAE among the baselines. The paper frames the learned decomposition as predictive attribution, not causal identification.

Critical strengths
  1. The paper uses a shared ERF to remove historical event effects and add responses to known future events.
  2. Ablations show that jointly modeling historical and future event effects performs better than direct event fusion or separate ERFs.
  3. The paper improves MAE across diverse datasets, and semi-synthetic tests show that it can recover injected event responses with little false attribution.
Critical weaknesses
  • The information sets of the baselines are not fully specified, it is unclear whether all future-exogenous baselines receive the same event, calendar, and environmental covariates as Loki.
  • The multiplicative response model encodes proportional effects, but its behavior for zero-valued or signed targets is not discussed.
  • The stop-gradient rule at event-active forecast positions is a training heuristic, it is not isolated in the ablation study.
  • The compared baselines are dominated by neural forecasting models. Comparisons with representative classical dynamic-regression or holiday-effect baselines would strengthen the empirical positioning.
Questions for authors
  1. The ablation should examine how performance changes when ShistS^{\mathrm{hist}} is computed excluding contributions from events occurring after the forecast origin.
  2. The authors should specify which datasets contain zero or non-positive target values and investigate the behavior of the method when the baseline or target values are zero.
  3. How sensitive is Loki to missing events or errors in event start and end times?
6 comments
AuthorsOpenReview ↗

Rebuttal Part 1

W1. The information sets of the baselines are not fully specified, it is unclear whether all future-exogenous baselines receive the same event, calendar, and environmental covariates as Loki.

Thank you for pointing out that the original manuscript did not sufficiently specify the baselines' inputs. The revised Table 1 distinguishes models using only target history, historical covariates, and known future covariates. The table below lists the common variable sources used by the known-future-covariate group, covering both the historical and forecast windows.

DatasetEvent informationCommon calendar informationEnvironmental information
Ride-HailingWeather-related events, social events, and holidaysDay of week, time of dayNo separately designated environmental fields
TaxiBJHoliday typeDay of week, time of dayTemperature, wind speed, and weather type
M5SNAP and event categoriesDay of weekPrice index and promotion share
RossmannPromotions, public holidays, and school holidaysDay of weekOpening status
Seoul BikeHolidaysDay of week, time of dayWeather and operating status

The models retain the following differences in how these inputs are handled:

  • Additional calendar information: some baselines retain within-year information.
  • Constant and collinear variables: AutoARIMA-X filters constant and linearly dependent columns; Log-linear assigns zero coefficients to columns that are constant during fitting. Loki and CEDAR omit TaxiBJ's constant weather_07 column.
  • Categories and encoding: Prophet constructs categorical regressors from categories observed during training; each method retains its own encoding scheme.

Appendix B.2, Tables 9 and 10 lists the covariates and their historical/forecast usage, including model-specific calendar inputs and preprocessing differences. Loki receives no exclusive event or environmental observations. Among the 11 models supporting future covariates, Loki achieves the lowest MAE on all five datasets, showing that its forecasting advantage also holds among models with access to the common event, calendar, and environmental information.

W2. The multiplicative response model encodes proportional effects, but its behavior for zero-valued or signed targets is not discussed.

Thank you for pointing out this applicability boundary. Loki operates in the original target space: historical observations are divided by a positive multiplier, and the backbone forecast is multiplied by a positive multiplier. Both operations remain defined when the historical observation or predicted baseline is zero or negative.

However, the multiplicative response preserves zeros and signs. For a fixed baseline, it cannot generate a nonzero event increment from an exactly zero baseline or represent an event-induced sign reversal; a finite positive multiplier also cannot map a nonzero baseline exactly to zero. A zero historical observation does not force a zero future forecast, since the backbone can predict a nonzero baseline from context. Q2 and Appendix D.7 report the per-dataset zero-value statistics and zero-target prediction results. We will clarify these applicability limits in the revised manuscript.

AuthorsOpenReview ↗

Rebuttal Part 2

W3. The stop-gradient rule at event-active forecast positions is a training heuristic, it is not isolated in the ablation study.

Thank you for pointing out the missing isolated ablation. We have held the architecture and other training settings fixed, changed only the SG setting, and evaluated both prediction accuracy and effect recovery.

(i) Real-data ablation. Standardized prediction MAE (multi-seed mean ± sample SD) on all five datasets is reported below:

DatasetSG offSG on
Ride-Hailing0.1570 ± 0.00030.1573 ± 0.0003
TaxiBJ0.1128 ± 0.00160.1120 ± 0.0017
M50.0958 ± 0.00110.0961 ± 0.0010
Rossmann0.1365 ± 0.00280.1538 ± 0.0036
Seoul Bike0.3238 ± 0.02080.3228 ± 0.0257

Removing SG lowers MAE on Ride-Hailing, M5 and Rossmann, while the TaxiBJ and Seoul Bike differences are within one seed standard deviation.

(ii) Semi-synthetic experiments. SG improves neither historical adjustment nor future effect recovery on either dataset under matched multiplicative injection. We fixed the shared multiplicative architecture and varied only SG. Prediction and recovery metrics use the same valid positions within response support, including anticipation and carryover.

All metrics are standardized MAE; each cell follows the order SG off → on.

DatasetResponse-support prediction MAEHistorical MAE after adjustmentFuture effect MAE
Ride-Hailing0.1694 → 0.16970.0207 → 0.02080.0228 → 0.0230
M50.1128 → 0.11570.0118 → 0.01340.0155 → 0.0179

(iii) Analysis and revision. At event-active forecast positions, SG detached the original-scale backbone output before Kick In, changing the gradient paths as follows:

PathBackpropagation order without SGWith SG
Future ERFLoss → final prediction → future responseRetained
BackboneLoss → final prediction → baseline prediction → backboneBlocked
Historical ERFLoss → final prediction → baseline prediction → backbone → adjusted history → historical responseBlocked

Non-active positions are unaffected; the shared ERF can still receive updates through the future path.

SG was intended to let the backbone learn routine variation and the ERF learn event responses. SG blocks backbone updates from errors at event-active positions while retaining updates to the future ERF, increasing the possibility that the ERF compensates for baseline errors.

We have therefore removed SG. Without it, Full Loki still outperforms Direct Fusion and Future-only, and the shared ERF still has lower mean prediction MAE than Independent, on both datasets. The revised paper will describe the final joint training procedure and use the final model for structural ablations. The SG analysis and on/off experiments above are reported in this response.

W4. The compared baselines are dominated by neural forecasting models. Comparisons with representative classical dynamic-regression or holiday-effect baselines would strengthen the empirical positioning.

Thank you for the suggestion. We have expanded the comparison to classical regression and component models, including:

  • AutoARIMA-X: Exogenous regression with ARIMA errors, representing classical dynamic regression.
  • SCAN*PRO-style log-linear regression: Regression in log space to represent multiplicative event effects.
  • Prophet: Trend and seasonal components, with holidays and other information supplied through exogenous regressors.

The table reports standardized prediction MAE: multi-seed means for Loki and deterministic-fit results for the three classical baselines.

ModelRide-HailingTaxiBJM5RossmannSeoul Bike
AutoARIMA-X0.46070.17180.11790.19270.5843
SCAN*PRO-style log-linear0.20980.12140.14500.17790.3585
Prophet0.26020.20400.10830.19200.7736
Loki0.15700.11280.09580.13650.3238

Loki has lower prediction MAE than these methods on all five datasets, with reductions of approximately 7.1%–25.2% relative to the best of these classical baselines on each dataset. This extends the empirical comparison to established event-modeling approaches. Revised Table 1 will report the full results; the appendix will document inputs and fitting configurations and include a method comparison table detailing differences in historical decomposition, forecast reconstruction and parameter learning.

AuthorsOpenReview ↗

Rebuttal Part 3

Q1. The ablation should examine how performance changes when is computed excluding contributions from events occurring after the forecast origin.

Thank you for this suggestion. Schedule statistics across the five datasets show substantial differences in how often future new events can contribute to historical adjustment; Ride-Hailing and Rossmann have the highest coverage of relevant test windows. A relevant window contains a new event starting at or after the forecast origin, within the ERF support that can reach historical positions.

DatasetRelevant test windows (%)Relevant test windows excluding weather (%)
Ride-Hailing95.39%40.38%
TaxiBJ5.00%5.00%
M50.00%0.00%
Rossmann67.89%67.89%
Seoul Bike4.11%4.11%

We therefore added the ablation on Ride-Hailing and Rossmann: after removing future new events' contributions to the historical response and retraining, average accuracy remains largely unchanged, both overall and on event-relevant windows. We retain the original complete schedule encoding and mask these new events' contributions only when computing the historical ERF response. Historical events continuing across the origin and the future ERF response remain unchanged. The Ride-Hailing intervention includes weather events.

The metric is standardized MAE over the full forecast horizon. Relative changes are computed for matched seeds and then averaged.

DatasetEvaluation windowsFull LokiAblatedRelative MAE change (%)
Ride-HailingAll windows0.1569560.157156+0.128%
Ride-HailingEvent-relevant windows0.1575650.157767+0.128%
Ride-HailingNon-weather new-event windows0.1707300.171032+0.177%
RossmannAll windows0.1364920.136680+0.136%
RossmannEvent-relevant windows0.1359780.136138+0.117%

These results indicate a small average forecasting benefit from this path within the evaluated scopes.

Appendix D.8 reports the coverage statistics, intervention protocol, and results.

Q2. The authors should specify which datasets contain zero or non-positive target values and investigate the behavior of the method when the baseline or target values are zero.

Thank you for the question.

(i) How does the model handle zero historical observations or backbone forecasts? Loki adjusts history as h = x · exp(−s_hist) and restores future responses as ŷ = y₀ · exp(s_future), where x is a historical observation and y₀ is the backbone forecast before restoration. The model neither takes the logarithm of the target nor divides by the target or backbone forecast, so these operations remain defined when historical observations or backbone forecasts are zero.

A zero historical observation remains zero after adjustment, but the backbone can use other historical positions and covariates to predict a nonzero future value. If the backbone forecast y₀ is zero, a finite multiplicative response cannot make it nonzero; a finite positive multiplier also cannot map a nonzero backbone forecast exactly to zero. This is a representational limit of the ERF for a given forecast; the full model can still handle series containing zero observations.

In testing, all exactly zero backbone forecasts correspond to zero true targets; no cases were observed in which the backbone forecast was zero but the target was nonzero. Positive multipliers also preserve the sign of the backbone forecast; since none of the five datasets contains negative targets in the original target space, sign-changing responses are outside the current empirical validation. We will clarify these boundaries in the revised manuscript.

(ii) How accurately does the model predict zero targets? All five datasets contain zero targets (true target = 0) and no negative targets in the original target space. The table, also included in Appendix D.7, reports the fraction of zero targets in the full data and compares prediction errors before and after ERF restoration at zero-target positions in the test set. Errors are mean standardized MAE across multiple seeds.

DatasetZero targets in full dataBackbone prediction MAE at zero targetsFinal prediction MAE at zero targets
Ride-Hailing0.0106%0.06380.0639
TaxiBJ5.1286%0.00490.0047
M50.2626%No zero targets in the test setNo zero targets in the test set
Rossmann16.9946%0.02190.0209
Seoul Bike3.3676%0.43100.4196

At zero-target positions, mean error decreases after ERF restoration in TaxiBJ, Rossmann and Seoul Bike, while Ride-Hailing is approximately unchanged; exactly zero final forecasts are not guaranteed. In Seoul Bike, zero targets reflect non-operation, whose status is supplied to the backbone. Errors at these positions do not directly measure ERF event-effect recovery (Appendix B.1).

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

This paper proposes Loki, an event-aware forecasting framework that separates event-associated variation from the underlying time-series trajectory. A shared Event Response Function (ERF) estimates multiplicative event responses over both the historical and forecast windows. Historical responses are removed before forecasting, and future responses are applied to the backbone prediction. The ERF and forecasting backbone are trained jointly using the final forecasting loss.

The method is evaluated on five benchmarks and achieves the lowest standardized MAE among the compared models on all five. Additional experiments include component ablations, comparisons across forecasting backbones, event-regime analyses, and semi-synthetic event-response recovery.

Critical strengths
  1. The framework provides a clear and coherent way to incorporate known events into forecasting. Sharing the response function across historical adjustment and future prediction gives these operations a consistent parameterization and allows the approach to be combined with different forecasting backbones.

  2. The empirical results demonstrate practical promise. Loki improves standardized MAE across five benchmarks, and the ablations and backbone experiments provide useful evidence that the event-response component contributes to forecasting performance. The controlled recovery experiments also show that the model can recover injected responses when the data-generating mechanism matches its assumptions.

Critical weaknesses
  1. Limited methodological originality and insufficient comparison with the closest approaches.

Separating external event effects from underlying dynamics, forecasting the adjusted trajectory, and restoring future event effects is a natural modeling strategy. Using the same response mechanism on both sides of the forecasting boundary is likewise a reasonable extension of this idea. While the paper provides a concrete implementation, it does not sufficiently establish what nontrivial methodological advance the shared ERF offers beyond this basic formulation. The absence of experimental comparisons with the closely related event-decomposition approaches discussed in Section 2.2 further limits the evidence for a substantive advance. My concern is about the originality and significance of the central mechanism, rather than the use of an existing forecasting backbone.

  1. Insufficient evaluation of the central modeling assumptions.

The paper adopts a multiplicative decomposition without comparing it against an additive alternative under an otherwise comparable shared-response architecture. The semi-synthetic experiments inject multiplicative responses, so they demonstrate recovery under a matched assumption rather than robustness to alternative event mechanisms. Although the real-world benchmarks span different applications, they primarily involve demand, traffic, and sales with calendar-related structure. The experiments do not establish robustness to unstable periodicity, state-dependent responses, or changing event effects. The lack of an ablation of the event-active gradient-blocking strategy also makes it difficult to determine how much of the reported benefit comes from the decomposition itself versus the training procedure.

Questions for authors
  1. What is the substantive technical advance of the shared ERF over existing event-decomposition methods, particularly those discussed in Section 2.2? Please clarify the nontrivial modeling or learning problem addressed by the proposed formulation and provide a direct comparison where feasible.

  2. Why is multiplicative decomposition preferred over additive decomposition? How does an additive variant perform when using the same backbone, shared-response structure, and comparable training protocol? Does multiplicative Loki remain effective when injected event effects are additive or depend on the underlying system state?

  3. Where exactly is the stop-gradient operation applied at event-active forecast positions? In particular, does it also block the gradient reaching the historical ERF through the forecasting backbone? Without this operation, do the full model and the shared ERF retain their advantages?

  4. How are the seasonal lengths selected for the default backbone and the ERF response support? How sensitive is the method to these choices, particularly when periodicity is misspecified, drifting, or irregular? How does the architecture balance recent and distant historical information beyond including a fixed-length recent-state branch?

  5. The model directly predicts an entire future window. How does its error vary with forecast distance, and what evidence supports temporal consistency across predicted steps? A horizon-wise analysis would help assess whether aggregate metrics conceal deterioration at longer horizons.

  6. Have the authors considered event-conditioned learnable transformations and inverse transformations, rather than restricting event removal and restoration to multiplication and division? Such a formulation might capture more general nonlinear or state-dependent responses. A discussion of the tradeoffs would help clarify the scope of the current design.

  7. Adaptive temporal weighting, learned or adaptive seasonal scales, and recursive forecasting are additional alternatives relevant to the temporal-modeling assumptions above. I do not assume that these alternatives would necessarily outperform Loki, but a discussion of their relationship to the proposed framework would help distinguish deliberate design choices from limitations of the current implementation.

9 comments
AuthorsOpenReview ↗

Rebuttal Part 1

W1. Limited methodological originality and insufficient comparison with the closest approaches.

Separating external event effects from underlying dynamics, forecasting the adjusted trajectory, and restoring future event effects is a natural modeling strategy. Using the same response mechanism on both sides of the forecasting boundary is likewise a reasonable extension of this idea. While the paper provides a concrete implementation, it does not sufficiently establish what nontrivial methodological advance the shared ERF offers beyond this basic formulation. The absence of experimental comparisons with the closely related event-decomposition approaches discussed in Section 2.2 further limits the evidence for a substantive advance. My concern is about the originality and significance of the central mechanism, rather than the use of an existing forecasting backbone.

Thank you for pointing out that the advance of the shared ERF beyond the basic decomposition idea was not established, and that comparisons with the closest event-decomposition methods were missing.

(i) Technical advance. We agree that removing event effects from history and restoring them in the forecast is a natural strategy, and we do not claim the decomposition itself as new. Loki's advance is a lightweight event pre- and post-processing module that, through a shared event-response function, is jointly trainable and makes this strategy learnable: the backbone learns routine dynamics from event-adjusted history, and responses are restored from future event schedules. It is compatible with different backbones and, with a lightweight structure, achieves the lowest MAE on all five benchmarks; on Ride-Hailing it uses only 0.27M parameters, about 86% below the median of the 15 neural baselines (Appendix C.1). It addresses three connected needs:

  • Historical adjustment should serve subsequent forecasting. PHR and SPADE process history using predefined rules, while CEDAR trains its event branch in stages. Loki learns historical adjustment from the final forecasting error, jointly with future responses and the backbone, allowing the model to adjust which variation to remove and by how much. With the same backbone, Loki outperforms Direct Fusion, which feeds events directly into the forecast, and Future-only, which restores only future responses, on all five datasets (revised Table 2).
  • Responses should vary with event conditions. Dynamic regression and Prophet already learn and share regression coefficients or components; NeuralProphet also uses shared components to adjust history and trains them jointly with an autoregressive network. Loki's advance lies in its multiscale response parameterization combining event intensity, relative position and entity information, and in applying responses in paired operations to historical observations and the full backbone forecast, representing anticipation, carryover and variation across event conditions. Reducing the multiscale temporal bases to the impulse and the slowest forward/backward pair raises MAE on four of the five datasets, for example from 0.1365 to 0.1448 on Rossmann (Appendix D.11).
  • Historical adjustment and future responses should constrain each other. NBEATSx shares its block network but generates backcast and forecast coefficients separately. Loki shares an event response function, so patterns learned from future response prediction directly inform historical adjustment and are optimized according to how that adjustment affects forecasting, reducing the freedom to explain event effects separately on the two sides. In semi-synthetic experiments with known injected effects, sharing reduces historical MAE after adjustment relative to separate ERFs from 0.0294 to 0.0207 on Ride-Hailing and from 0.0272 to 0.0118 on M5 (Appendix E.4).

(ii) Comparison with the closest methods. We added SPADE and CEDAR, reimplemented from their papers since neither has a public implementation, together with AutoARIMA-X, SCAN*PRO-style log-linear regression, Prophet, NeuralProphet and NBEATSx. Loki has lower MAE than all of them on all five datasets (revised Table 1). The largest margin is on Rossmann, where the strongest added baseline, CEDAR, has MAE 0.1545 versus 0.1365 for Loki. Appendix F compares how each method handles historical decomposition, forecast reconstruction, parameter sharing and parameter learning, and Q1 reports the controlled comparisons in detail.

We will revise Section 2.2 and the contribution statement to present the shared ERF as the mechanism that makes historical adjustment learnable through future-side supervision.

AuthorsOpenReview ↗

Rebuttal Part 2

W2. Insufficient evaluation of the central modeling assumptions.

The paper adopts a multiplicative decomposition without comparing it against an additive alternative under an otherwise comparable shared-response architecture. The semi-synthetic experiments inject multiplicative responses, so they demonstrate recovery under a matched assumption rather than robustness to alternative event mechanisms. Although the real-world benchmarks span different applications, they primarily involve demand, traffic, and sales with calendar-related structure. The experiments do not establish robustness to unstable periodicity, state-dependent responses, or changing event effects. The lack of an ablation of the event-active gradient-blocking strategy also makes it difficult to determine how much of the reported benefit comes from the decomposition itself versus the training procedure.

Thank you for pointing out the need to evaluate the modeling assumptions more thoroughly. We have added relevant comparisons and will clarify the scope of the paper.

(i) Rationale for multiplicative decomposition. The multiplicative form lets the ERF learn relative changes induced by events, so the absolute change scales with the underlying level. The additive comparison in the ablation study uses the same backbone, shared ERF and training protocol; the multiplicative variant has lower mean prediction MAE on all five datasets. See Q2 for the rationale and detailed results.

(ii) Performance when the assumptions do not fully hold. We have added semi-synthetic experiments with additive effects and state-dependent responses. Multiplicative Loki still recovers part of the additive effects, although less accurately than the additive architecture. When responses vary with state, historical adjustment and future effect recovery also remain effective, with recovery error remaining. See Q2 for the detailed results and limitations.

(iii) Does SG affect these conclusions? Gradient analysis, ablation experiments and semi-synthetic experiments support removing SG; the structural benefits of the full model remain after its removal. The final Loki therefore no longer uses this operation. See Q3 for the gradient paths, SG on/off comparisons and structural comparisons.

(iv) Scope and limitations. This work focuses on local responses triggered by known event schedules in time series with a usable periodic background, including anticipatory, concurrent and carryover effects. The current study does not cover unstable periodicity or persistently drifting event-response mechanisms. We also deliberately do not feed the recent underlying-series level directly into the ERF, to limit its freedom to compensate for forecasting errors in regular dynamics; this design also limits its capacity to represent state-dependent responses. Appendices D.5 and E.5 discuss temporal-scale sensitivity and the state-dependence limitation, respectively.

AuthorsOpenReview ↗

Rebuttal Part 3

Q1. What is the substantive technical advance of the shared ERF over existing event-decomposition methods, particularly those discussed in Section 2.2? Please clarify the nontrivial modeling or learning problem addressed by the proposed formulation and provide a direct comparison where feasible.

Thank you for the question. Loki models event effects as local proportional responses acting on the underlying series. The ERF combines event type, intensity, relative position and entity information through multiscale bidirectional temporal bases. Historical inputs are divided by the response multiplier, and backbone forecasts are multiplied by the future response multiplier. Both paths update the ERF through the final forecasting loss, so patterns learned from future response prediction inform historical adjustment and are optimized according to how that adjustment affects forecasting.

Specific differences from the closest methods. They correspond to three design choices:

(1) Historical adjustment serves forecasting: PHR and SPADE process history using predefined rules, and CEDAR trains its event branch in stages, whereas Loki learns the adjustment from the final forecasting error, jointly with future responses and the backbone.

(2) Responses vary with event conditions: dynamic regression and Prophet learn and share regression coefficients or components, and NeuralProphet also uses shared components to adjust history and trains them jointly with an autoregressive network; Loki's ERF instead generates responses from these event conditions through multiscale bidirectional temporal bases and applies them in paired operations to historical observations and the full backbone forecast.

(3) The two sides constrain each other: NBEATSx shares its block network but generates backcast and forecast coefficients separately, whereas Loki uses the same ERF on both sides. Appendix F compares historical decomposition, forecast reconstruction, parameter sharing and parameter learning across all methods.

Controlled evaluation of the technical advance. We separately test the forecasting benefits of decomposition and sharing, and whether sharing improves historical adjustment. All models use the same forecasting backbone. The semi-synthetic experiments inject known multiplicative effects; all metrics in the table are standardized MAE.

ComparisonModel orderMetricRide-HailingM5
DecompositionDirect Fusion → Future-only → LokiPrediction MAE0.16124 → 0.15778 → 0.156960.10257 → 0.09848 → 0.09583
SharingShared → IndependentPrediction MAE0.15696 → 0.157170.09583 → 0.09825
Semi-synthetic historical recoveryShared → IndependentHistorical MAE after adjustment0.0207 → 0.02940.0118 → 0.0272
  • Decomposition: Loki outperforms Direct Fusion and Future-only on both datasets.
  • Sharing: The shared ERF lowers mean prediction MAE on M5; on Ride-Hailing the difference is within one seed SD.
  • Recovery: In semi-synthetic experiments, sharing improves historical adjustment on both datasets. A parameter-matched Half-independent variant gives 0.0299 and 0.0277, so the gain comes from sharing rather than capacity (Appendix E.4).

Direct comparisons with related methods. We added SPADE and CEDAR. Neither has a public implementation, so we reimplemented both following their papers. We also added AutoARIMA-X, SCAN*PRO-style log-linear regression, Prophet, NeuralProphet and NBEATSx. Loki has lower prediction MAE than these added methods on all five datasets. Revised Table 1 reports the full results by input availability; Appendix C.2, Table 12 documents SPADE's backbone adaptation and CEDAR's structured-event substitution.

We will revise Section 2.2 and the statement of contributions accordingly.

Comments on the paper

4 comments
AuthorsOpenReview ↗

Rebuttal Part 1

Thank you for your constructive suggestions.

W1. The closest competing methods are absent from the empirical comparison. The related work discusses explicit event-modeling approaches that Loki positions itself against, yet the main comparison covers only generic forecasting backbones and covariate-conditioning models, so the claim that a shared response function improves on existing decomposition-style event handling is argued rather than tested. The main table also mixes models with different levels of information access, which makes the headline ranking partly a cross-regime statement. Including the closest event-aware baselines, or at least reporting rankings within each information regime, would substantiate the positioning.

Thank you for pointing out the missing comparison with the closest methods and the cross-regime ranking in the main table. We added relevant baselines and report comparisons separately by information regime.

(i) Comparisons with the closest methods. We added SPADE and CEDAR. Neither has a public implementation, so we reimplemented both following their papers. We also added AutoARIMA-X, SCAN*PRO-style log-linear regression, Prophet, NeuralProphet and NBEATSx. Loki has lower prediction MAE than these added methods on all five datasets. The full results will appear in revised Table 1.

(ii) Comparisons within the same information regime. The main table groups models by input availability:

  • Pure endogenous: DLinear, PatchTST, SparseTSF, TimeMixer, iTransformer.
  • Historical exogenous: TimeXer, CrossLinear, XLinear.
  • Future exogenous: TiDE, DAG, TFT, AutoARIMA-X, log-linear regression, Prophet, NeuralProphet, NBEATSx, SPADE, CEDAR and Loki.

Multi-seed mean standardized prediction MAE of the strongest model in each regime:

DatasetPure endogenousHistorical exogenousFuture exogenous (excl. Loki)Loki
Ride-Hailing0.1618 (PatchTST)0.1649 (CrossLinear)0.1625 (TiDE)0.1570
TaxiBJ0.1264 (DLinear)0.1276 (CrossLinear)0.1191 (TFT)0.1128
M50.1022 (PatchTST)0.1019 (XLinear)0.1032 (NeuralProphet)0.0958
Rossmann0.2482 (PatchTST)0.2518 (XLinear)0.1545 (CEDAR)0.1365
Seoul Bike0.5047 (TimeMixer)0.5050 (CrossLinear)0.3310 (TFT)0.3238

Within the future-exogenous regime, Loki has the lowest MAE on all five datasets, improving on the strongest other model by 3.4% (Ride-Hailing), 5.3% (TaxiBJ), 7.2% (M5), 11.7% (Rossmann) and 2.2% (Seoul Bike).

The future-exogenous group uses common event, calendar and environmental information. Models retain their own encodings and constant-column handling, and some baselines also use additional calendar features. Loki receives no exclusive event or environmental observations. Full input configurations are provided in revised Appendix B.2, Tables 9 and 10.

AuthorsOpenReview ↗

Rebuttal Part 2

W2.The decomposition is only weakly identified … none is ablated, and no parameter or compute comparison rules out added capacity as a confounder …

Thank you. We added effect-recovery tests, design ablations, and parameter and computational-cost comparisons.

(i) Attribution of event effects. To test whether the ERF captures injected event effects, we compare its estimates with known effects added to real series:

Historical MAE follows before → after adjustment; future effect MAE follows zero response → multiplicative structure.

DatasetInjection conditionHistorical MAEFuture effect MAE
Ride-HailingMultiplicative effects0.3333 → 0.02070.3284 → 0.0228
Ride-HailingEvents scheduled, no effect injected0 → 0.00970 → 0.0082
M5Multiplicative effects0.1927 → 0.01180.2493 → 0.0155
M5Events scheduled, no effect injected0 → 0.00520 → 0.0062

With an injected effect, Lift Out reduces the historical error by 93.8% (Ride-Hailing) and 93.9% (M5), and Kick In reduces the future-increment error by 93.1% and 93.8%, so the effect is captured by the ERF rather than absorbed by the backbone. In the control experiment where events are scheduled but no effect is injected (truth: 0), both operations still produce small outputs, but these remain close to zero, indicating that the ERF rarely attributes routine dynamics to events. Settings and full recovery results: revised Appendices E.2–E.4.

(ii) Additional ablations. We ablated stop-gradient (SG), the clipping range, initialization and the temporal bases.

  • SG. The table reports standardized prediction MAE (multi-seed mean ± sample SD) with SG off and on:
DatasetSG offSG on
Ride-Hailing0.1570 ± 0.00030.1573 ± 0.0003
TaxiBJ0.1128 ± 0.00160.1120 ± 0.0017
M50.0958 ± 0.00110.0961 ± 0.0010
Rossmann0.1365 ± 0.00280.1538 ± 0.0036
Seoul Bike0.3238 ± 0.02080.3228 ± 0.0257

Removing SG lowers MAE on Ride-Hailing, M5 and Rossmann, while the TaxiBJ and Seoul Bike differences are within one seed SD.

  • Clipping. Total-response clipping at ±8, ±16 or none gives identical prediction metrics.

The final Loki therefore uses neither SG nor response clipping. The initialization and temporal-basis comparisons below use this version and report multi-seed mean standardized prediction MAE:

VariantRide-HailingM5
Loki0.15696 ± 0.000310.09583 ± 0.00109
Random response initialization0.15689 ± 0.000350.09621 ± 0.00172
Simplified temporal bases0.15782 ± 0.000120.09823 ± 0.00230
  • Initialization. Random initialization of the response weights, N(0, 0.01²), gives results close to zero initialization. We keep zero initialization so that training starts with no event effect, where both event operations are identities.
  • Temporal bases. Reducing the temporal bank to the impulse and the slowest forward/backward pair (19 → 3 bases on Ride-Hailing, 7 → 3 on M5, same support range) increases MAE on both datasets, supporting the richer basis set. These comparisons are reported in revised Appendix D.11.

(iii) Capacity and computation. The gains do not come from added capacity.

  • Parameters and cost. Loki has 0.27M trainable parameters, fewer than every neural baseline except SparseTSF, NeuralProphet and SPADE. Cost on Ride-Hailing, measured on the same RTX 6000D GPU:
ModelTrainable parametersTraining (min)Inference (ms/window)
DLinear0.90M2.30.07
PatchTST1.66M11.60.07
SparseTSF2453.10.07
TimeMixer10.97M30.10.38
iTransformer0.48M5.80.06
TimeXer3.02M24.80.14
CrossLinear2.01M3.50.08
XLinear3.12M2.50.07
TiDE5.45M8.50.06
DAG17.21M46.90.31
TFT0.95M175.80.73
NBEATSx1.98M10.50.08
NeuralProphet0.11M6.90.07
SPADE0.23M110.10.40
CEDAR4.78M90.315.83
Loki0.27M18.10.20

Training time is the measured per-epoch time multiplied by the number of epochs actually run; inference covers the full test split. AutoARIMA-X, log-linear regression and Prophet are fitted per series on CPU and are therefore not included. Parameter counts, training and inference costs are reported in revised Appendix C.1.

  • Same-capacity control. To isolate capacity, we also compared Loki with a same-backbone variant that has the same number of parameters and the same event information, but feeds the events directly into the forecast instead of removing and restoring event responses. If the gains came from capacity, this variant would match Loki; instead, its MAE is 0.16124 versus 0.15696 on Ride-Hailing and 0.10257 versus 0.09583 on M5. Full structural comparisons: revised Table 2.
AuthorsOpenReview ↗

Rebuttal Part 3

W3. The supporting analyses cover a narrow slice of the benchmarks under favorable conditions … the recovery study relies on injected events that are well separated … the margins over the strongest baselines are modest, statistical significance should be reported. The acknowledged sensitivity … to imperfect event schedules is likewise never quantified.

Thank you for raising dataset coverage, the conditions of the semi-synthetic study, statistical significance and schedule errors.

We selected Ride-Hailing and M5 for in-depth analysis because they cover different domains and temporal resolutions: Ride-Hailing provides large-scale, multi-entity, multi-target demand at 30-minute intervals, while M5 provides daily retail sales across stores and departments. This lets us examine event-response recovery and backbone compatibility against different temporal backgrounds. To broaden the evaluation, revised Tables 1 and 2 report the main results and internal structural ablations across all five datasets.

(i) Overlapping and interacting events in semi-synthetic experiments.

Overlap. We fix event count, type, duration and dose, shifting times to create overlap without interaction. Future effect MAE:

DatasetSeparated → overlapping
Ride-Hailing0.0228 → 0.0261
M50.0177 → 0.0194

Error increases slightly on both datasets; Loki still recovers responses following its multiplicative composition.

Interaction. We add an interaction to fixed overlapping schedules and individual responses. We also match the no-interaction control's training perturbation strength to check whether larger effects alone explain the error increase. Future effect MAE:

DatasetNo interactionNo interaction, training-strength matchedWith interaction
Ride-Hailing0.02470.02530.0438
M50.01640.01670.0339

Interaction increases future effect MAE on both datasets; the error remains higher after matching training perturbation strength.

Both evaluate total event effects. Settings and results: Appendix E.5.

(ii) Statistical significance. For each dataset, we compared Loki with the strongest baseline using a Diebold–Mariano (DM) test over test forecast origins, with HAC standard errors that account for overlapping forecast windows.

Prediction MAE is standardized and follows baseline → Loki; reductions are relative to the baseline.

DatasetBest baselinePrediction MAEMAE reduction (%)Test originsOrigins won by Loki (%)DM p-valueSeeds won by Loki
Ride-HailingPatchTST0.1618 → 0.15703.0%17880%0.0105/5
TaxiBJTFT0.1191 → 0.11285.3%2080%0.204/5
M5XLinear0.1019 → 0.09586.0%1—n/a5/5
RossmannCEDAR0.1545 → 0.136511.7%16298%< 0.0015/5
Seoul BikeTFT0.3310 → 0.32382.2%7359%0.853/5

The advantage is significant on Ride-Hailing and Rossmann. TaxiBJ and Seoul Bike have few test origins, and their differences are not significant; M5 has only one test forecast origin, so the DM test across forecast origins is not applicable. The test procedure and full results are in revised Appendix D.12.

(iii) Effect of inaccurate schedules. Loki targets settings in which future events are known, such as public holidays, promotions and planned activities; baselines that use known-future covariates, including TiDE, TFT, DAG, SPADE and CEDAR, rely on the same assumption. To quantify the effect of schedule errors, we keep the trained models fixed and perturb only the test-time event schedule: we either randomly drop 30% of events or displace event starts, ends or whole events at the predefined middle severity level. Event-specific displacement magnitudes and constraints are given in the full perturbation settings. Each condition uses three random schedules, averaged within each model seed and then across model seeds. The table reports the percentage change in prediction MAE relative to the original schedule:

Dataset30% droppedStart displacedEnd displacedWhole event shifted
Ride-Hailing+2.7%+1.3%+2.7%+2.0%
TaxiBJ+1.0%−0.9%+9.7%+9.3%
M5+10.9%+13.4%+9.7%+18.0%
Rossmann+52.3%+85.2%+28.8%+124.8%
Seoul Bike+0.4%+0.5%−0.1%+0.4%

Event-schedule errors usually reduce prediction accuracy, but their impact depends on the dataset and error type. Mean errors increase with the omission rate on all five datasets, whereas timing errors have asymmetric effects. Rossmann is the most sensitive to schedule errors, while changes on Ride-Hailing and Seoul Bike are smaller under the tested perturbations. Seoul Bike has limited holiday coverage (Appendix B.1). All three severity levels and affected-event counts are reported in revised Appendix D.9.

AuthorsOpenReview ↗

Summary of Revisions and Responses

Dear AC and Reviewers,

We thank all four reviewers for their thorough and constructive reviews. Their questions led us to test Loki's central mechanism more directly. The reviewers recognized the following strengths:

  1. The paper addresses erroneous extrapolation of historical event effects "with a clear research motivation" (Reviewer Rybn) and turns it "into a well-defined decomposition" of what to extrapolate and what to attribute to events (Reviewer MBaU).
  2. Sharing one response function across history and forecast "gives these operations a consistent parameterization" (Reviewer PHUU), "resulting in a concise structure" (Reviewer Rybn), and "directly addresses the weakness the paper identifies in prior work" (Reviewer MBaU).
  3. The approach can "be combined with different forecasting backbones" (Reviewer PHUU) and "brings consistent gains when attached to different forecasting backbones, indicating that the contribution does not depend on one particular architecture" (Reviewer MBaU).
  4. The empirical design "tests the mechanism rather than only aggregate accuracy" (Reviewer MBaU), and semi-synthetic tests show recovery of injected responses "with little false attribution" (Reviewer yGkm).

The reviewers also raised the following main questions; the corresponding evidence is summarized below.

1. How does Loki differ from the closest event-modeling methods? (PHUU, Rybn, MBaU, yGkm)

  • We clarified that Loki is a lightweight, jointly trainable event pre- and post-processing module: one shared ERF removes event effects from history and restores them in the forecast, around a replaceable backbone. We also added a detailed comparison with the closest methods in Section 2.2 and Appendix F. (PHUU, Rybn)
  • We added SPADE, CEDAR, NeuralProphet, NBEATSx, Prophet, AutoARIMA-X and log-linear regression. Loki has the lowest MAE on all five datasets, also among models given the same known-future covariates (Table 1, Appendix B.2). (MBaU, PHUU, Rybn, yGkm)

2. Do the gains come from the decomposition itself, rather than the training procedure or added capacity? (MBaU, Rybn, yGkm, PHUU)

  • Training procedure: ablations on all five datasets and semi-synthetic recovery tests showed no consistent benefit from the event-active stop-gradient. We therefore revised the model to remove it, together with response clipping, and report the main results and structural ablations with the revised model. Reviewer Rybn confirmed that the model remains effective without it. (MBaU, Rybn, yGkm, PHUU)
  • Capacity: a same-backbone Direct Fusion variant with the same parameters and event inputs, which feeds events directly into the forecast, is worse on all five datasets (Table 2). (MBaU, Rybn)
  • Parameterization: replacing the ERF with an event lookup table in the same structure is worse in every seed on four datasets (Table 2). (Rybn)

3. Does the ERF attribute event effects rather than routine dynamics? (MBaU, Rybn, PHUU)

  • With known effects, including anticipatory and carryover tails, injected into real series, Lift Out and Kick In each reduce the recovery error by over 93%; with events scheduled but no effect injected, their outputs stay near zero (Appendices E.1–E.4). (MBaU, Rybn)
  • On real data, where event effects are unobservable, a controlled input-replacement diagnostic supports historical event inputs as a contributor to erroneous extrapolation; Loki reduces, though does not eliminate, this error (Appendix D.10). (Rybn)
  • In a follow-up experiment, Loki has the lowest MAE among the structural variants on Ride-Hailing, both where post-event demand stayed elevated and where it had recovered; with injected no-, short- and long-tail effects, the ERF recovers the tail of each event type. (Rybn)

4. How reliable are the results, and where do they apply? (MBaU, PHUU, yGkm, Rybn)

  • We added Diebold–Mariano significance tests against the strongest baseline; the differences are significant on Ride-Hailing and Rossmann, the two datasets with the most test origins (Appendix D.12). (MBaU)
  • We extended the semi-synthetic experiments to additive, state-dependent, overlapping and interacting event effects, and examined how well Loki recovers the injected responses in each scenario (Appendices E.2–E.5). (PHUU, Rybn)
  • We report sensitivity to event-schedule errors (D.9), horizon-wise errors (D.6) and zero-valued targets (D.7), and state the resulting scope in Section 5. Reviewer yGkm noted that their main concerns "have been addressed" and that the remaining limitations "are now clearly characterized." (MBaU, PHUU, yGkm)

Together, these results indicate that both the decomposition structure and the ERF parameterization contribute to Loki's gains, and show that the learned responses track known event effects within the stated scope.

We hope these revisions can be taken into account in the final assessments, and we would be glad to clarify any remaining questions.

Sincerely, The Authors of Submission 44522

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.