Pith. sign in

REVIEW 3 major objections 4 minor 175 references

Contextual Deconvolution splits demand into baseline plus shocks; it cuts safety stock and order variance but under-provisions spikes, so total cost falls only when holding costs exceed about 20% of stockout costs.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 01:48 UTC pith:5JRDIBD7

load-bearing objection Honest negative result with a useful h/p crossover, but the cost claim rests on an underspecified shock-injection rule. the 3 major comments →

arxiv 2607.25664 v1 pith:5JRDIBD7 submitted 2026-07-28 cs.LG math.OCstat.ML

Contextual Deconvolution for Variance-Stable Demand Sensing: Kernel-Modulated Operators in Promotional Retail

Mohammad Forouhesh This is my paper
classification cs.LG math.OCstat.ML
keywords retail demand sensingpromotional carryoverkernel-modulated operatorconvex decompositionsafety stockbullwhip effectvariance ratiohierarchical partial pooling
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper tries to establish that much of retail demand volatility is not inherent noise but a recoverable mixture of smooth structural demand and sparse promotional shocks, and that recovering that mixture with a kernel-modulated operator—rather than letting a forecaster absorb it all—yields substantially more stable planning signals. On two large grocery catalogs, the method cuts forecast variance and safety stock by large factors, and its accuracy edge is reliability: far fewer SKUs land in the catastrophic-error tail compared with every baseline. The paper also pins down the trade-off: the smoothing under-provisions event spikes, so total inventory cost improves only when holding costs exceed about 20% of stockout costs, with a bootstrap confidence interval of [17%, 25%]. Below that threshold, the method is an operational-stability and inventory-capital layer, not an expected-cost minimizer. The paper is explicit that this is a conditional framework that requires a known future promotional calendar.

Core claim

On its own terms, the central claim is that demand sensing should be treated as an inverse problem: observed demand is a superposition of a smooth baseline and sparse event-driven shocks mapped through a causal carryover operator. Solving that convex decomposition produces forecasts that are dramatically smoother than machine-learning forecasts, cutting variance ratio and safety stock at scale, while accuracy improves mainly in the tail—fewer SKUs are mis-forecast by more than 200%. The load-bearing result is the honest cost accounting: the smoothing under-covers promotional spikes, so the method lowers total inventory cost only when the holding-to-stockout ratio h/p is roughly at least 0.20

What carries the argument

The central object is the kernel-modulated banded operator A_spec(theta): a lower-triangular banded matrix built from a causal kernel (impulse, exponential decay, or an optional surge-and-vacuum difference-of-exponentials) that maps sparse latent shocks into observed demand with multi-day carryover. It replaces the identity operator assumed by classical decomposition, remains invertible, and is estimated by a two-stage pipeline—robust detrending followed by ridge-based distributed-lag regression—with hierarchical partial pooling across SKUs. The operator is data-derived: it reduces to an impulse wherever promotional response is instantaneous, so the operational gains rest on the structural d

Load-bearing premise

The load-bearing premise is that the magnitudes of future promotional shocks injected by the method's deterministic event stream are recovered without bias from the historical distribution of recovered shocks conditioned on event type and promotional depth; the paper does not specify or validate that estimator, and if it is biased downward, the under-provisioning of spikes—and the h/p about 0.20 cost crossover derived from it—would be an artifact of the injection rule rather

What would settle it

Estimate true promotional lift out-of-sample on held-out event days (event-day demand minus model baseline), compare it with the magnitudes the method injects, and re-run the inventory simulation with unbiased injection magnitudes. If under-provisioning disappears and the method no longer shows higher stockout cost below h/p about 0.20, the central cost crossover is an artifact of the injection estimator, not of the decomposition.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • In holding-cost-dominated settings (roughly h/p >= 0.20), deploying this method as a planning layer should lower total inventory cost, not just smooth signals.
  • Where stockout costs dominate—the common grocery regime—the method raises expected cost, and its value is limited to stability and reduced capital tied up in buffer stock.
  • Catalog-scale deployment is feasible without per-SKU training: one category-level kernel applied across tens of thousands of SKUs matches per-SKU estimation.
  • The carryover kernel contributes only where genuine multi-day carryover exists; in impulsive-promotion regimes the decomposition alone does the work, so gains are structural rather than tuning artifacts.
  • Planners inherit a distinctive tail: far fewer SKUs with catastrophic forecast errors, at the price of systematically missing promotional-spike magnitudes.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Inference: The unspecified estimator for the deterministic event stream's injected shock magnitudes is the fragile link; replacing it with a validated, unbiased magnitude model could shift the h/p crossover and should be tested before treating 0.20 as a deployment rule.
  • Inference: The cross-sectional tail reliability suggests a natural ensemble role—using this method as a guardrail member to bound worst-SKU failures while a reactive forecaster covers spike timing and depth.
  • Inference: The non-normality analysis implies that dense learned operators spread impulse responses across lags; a testable corollary is that constraining any event-aware forecaster to a low-rank or banded response kernel would reduce its variance ratio without an accuracy collapse.
  • Inference: The measured 20% threshold comes from one base-stock simulation; real policies with lost-sales versus backordering, multi-echelon coupling, or service-level penalties could move the crossover, so the threshold is a diagnostic guide rather than a universal constant.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper introduces Contextual Deconvolution (CD), a two-stage convex decomposition that separates a smooth structural baseline from sparse, calendar-driven promotional shocks using a kernel-modulated banded operator, with hierarchical partial pooling for catalog-scale deployment. The method is evaluated strictly out-of-sample on 30,490 M5 SKUs and 2,845 Favorita items. The empirical claims are: CD dramatically reduces Variance Ratio and safety stock relative to ML baselines; it improves cross-sectional reliability (low dispersion and catastrophic-failure rate) without large median accuracy gains; and it under-provisions promotional spikes, so that a full inventory-cost accounting yields a crossover at holding-to-stockout ratio h/p ≈ 0.20 (95% CI [0.17,0.25]), below which CD is an operational-stability layer rather than an expected-cost minimizer. The authors are explicit that VR and safety stock are diagnostics, and they repeatedly frame the paper's contribution as a tradeoff analysis rather than a pure accuracy improvement.

Significance. If the claims hold, the paper provides a valuable, honest counterpoint to accuracy-centric forecasting: a scalable structural decomposition that trades expected cost for operational stability in stockout-dominated regimes, and a clear negative result about when smoothing should not be used. The reproducibility effort is a real strength (CPU-only pipeline, seeded draws, single-command regeneration), and the cross-sectional reliability result in Table 6 is a concrete, falsifiable empirical finding. The theoretical guarantees in Appendix D are carefully hedged and connected to measured proxies. However, the central quantitative conclusion — the h/p≈0.20 crossover — depends on an underspecified component in the Stream B shock-injection rule, so the significance is conditional on that component being specified and validated.

major comments (3)
  1. [Appendix P; §4.3 Finding 4b; Table 15; Fig. 2] Stream B's future shock magnitudes are described only as 'estimated from the empirical distribution of historically recovered shocks, conditioned on event type and promotional depth' (Appendix P). No estimator is specified (mean? median? quantile?), and no validation is reported. This is load-bearing: Finding 4b/Table 15 report CD's stockout cost at roughly 2.9× Tuned XGBoost's, and the h/p≈0.20 crossover follows from that under-provisioning. The recovered shocks feeding this distribution come from the ℓ1-regularized Problem (2)/Eq. (5); Theorem 1 certifies support recovery, not unbiased magnitudes, and ℓ1/ridge shrinkage biases nonzero coefficients toward zero. If the injected magnitudes inherit the shrinkage, CD's event-day forecasts are systematically low, inflating stockout cost, and the crossover is an artifact of the injection rule. Please specify the estimator, validate it on synt
  2. [§4.3 Finding 4c; §4.6; Table 5] The paper frames CD's operational advantage as coming from the structural decomposition rather than the ML paradigm, but no smoothing-only control is compared. Since VR and std-based safety stock are minimized by any sufficiently smooth forecast (as the paper itself admits in §4.6), a Tuned XGBoost with post-hoc temporal smoothing or output averaging could plausibly reproduce the low VR, low safety stock, and part of the order-variance gains without the sparse-shock decomposition. The ablation in Finding 4c varies the operator (learned vs. impulse) but keeps the two-stream forecaster; it does not isolate smoothness per se from the structural decomposition. Adding a 'smoothed ML' baseline to the head-to-head and the inventory-cost sweep is necessary to substantiate the claim that the decomposition, not just smoothing, carries the operational gains.
  3. [§4.3 Finding 4b; Appendix R.6] The inventory-cost simulation that generates the h/p≈0.20 crossover is underspecified. The text gives lead time (7 days), stockout cost ($10/unit), service level (0.95), and the h/p sweep, but does not state whether stockouts are lost sales or backorders, how base-stock levels are computed (demand distribution, forecast error model), the review period, or the exact safety-stock formula. These modeling choices can materially shift the crossover and the reported 95% CI [0.17,0.25]. Because this crossover is the paper's central negative result, the full protocol should be provided so the cost accounting is reproducible and verifiable.
minor comments (4)
  1. [Abstract; §1] Typesetting errors: missing spaces in 'we introduceContextual Deconvolution' (§1) and 'ContextualDeconvolution' (abstract). Please correct.
  2. [Table 5 and Table 6] Table 5 reports metrics for 200 SKUs, while Table 6 uses four draws of 2,000 SKUs. The captions say this, but the main text sometimes moves between samples without restating the sample size (e.g., the 200-SKU RMSSE comparison and the 2,000-SKU reliability table). Please make the sample size explicit at each point of comparison.
  3. [Appendix G] The SARIMAX comparison reports SD 0.647 vs 0.663 for ARIMAX variants against 0.113 for CD, but Table 6 reports CD's SD as 0.23±0.02 on 2,000 SKUs. Please clarify the sample size and whether the SARIMAX comparison is on a different subset, so the reader can reconcile the numbers.
  4. [Appendix R.4; §3.1 footnote] The footnote gives median non-zeros 95.6 for joint optimization vs 1.9 for two-stage, but Appendix R.4 reports this on 50 M5 SKUs. State the sample size explicitly in both places.

Circularity Check

0 steps flagged

No significant circularity: the paper explicitly disclaims the VR/safety-stock gains as diagnostics, and the h/p crossover, while dependent on an unvalidated Stream B injection rule, is an out-of-sample simulation rather than a fitted input recycled as a prediction.

full rationale

The only apparent self-referential element is the Variance Ratio / safety-stock result, but the manuscript explicitly disclaims it as an objective: "Because the Variance Ratio and std-based safety stock are both minimized by any sufficiently smooth forecast, we treat them as diagnostics rather than objectives" (Abstract; §4.6). Thus the low VR and safety stock are presented as properties of the smoothing objective, not as independent predictions, which removes the circularity. The load-bearing h/p ≈ 0.20 crossover is computed from a strictly out-of-sample inventory simulation against actual test demand, so it is not a fitted parameter renamed as a prediction. One substantive weakness is the Stream B injection rule: "we construct A_future_spec from these known calendars and inject expected shock magnitudes, estimated from the empirical distribution of historically recovered shocks" (Appendix P), with no estimator specified. If the recovered shocks inherit Lasso/ridge shrinkage, the injected magnitudes could be biased downward, which would bias the stockout cost and crossover. This is a validity/robustness risk, not a circularity: the forecast is a model input, and the cost is measured against external test data; the paper does not define its conclusion in terms of the injection rule. There are no load-bearing external self-citations (the references contain no prior work by the author; the theorems are proved in the appendices), the transfer-function framing is explicitly credited to classical econometrics, and the DoE kernel's vacuum component is explicitly tested and found inessential (Appendix T). The derivation chain is therefore self-contained rather than circular.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 0 invented entities

The central smoothing effect depends on λ_smooth and λ_sparse, the kernel bandwidth and kernel parameters, and the injection magnitudes; each is either hand-set or estimated from the same data. The theoretical guarantees require incoherence, restricted eigenvalue, beta-min, and a linear-detrender idealization, all acknowledged as idealized.

free parameters (6)
  • λ_smooth = 10.0
    Fixed default smoothness penalty in Problem (2); chosen by hand, never tuned, but the smoothing it controls is what produces low VR and safety stock.
  • λ_sparse = 8.0
    Sparsity penalty; hand-set default. Controls shock sparsity and therefore which events are provisioned.
  • Kernel bandwidth n = 14 days
    Data-derived from empirical impulse response estimated on 500 M5 SKUs (Table 2); not tuned to downstream metric, but still estimated from the evaluation dataset family.
  • DoE kernel parameters θ = Not reported numerically
    Surge/vacuum/decay rates fitted by nonlinear least squares against empirical per-category kernels (Stage 2).
  • Future shock magnitudes = Not reported
    Stream B injects expected magnitudes from the historical distribution of recovered shocks; estimator unspecified (Appendix P).
  • Huber δ = Auto-estimated from MAD
    Standard robust location parameter for the detrending stage.
axioms (6)
  • domain assumption Additive model b = x_base + A_spec x_spec + ε
    Eq (1); assumes demand is a superposition of smooth baseline, convolved sparse shocks, and Gaussian noise.
  • domain assumption Causal banded kernel operator
    Eq (3); assumes promotional effects are causal and confined to n lags.
  • ad hoc to paper Identifiability conditions (mutual incoherence, restricted eigenvalue, beta-min)
    Assumption 1 / Theorem 1; authors acknowledge these are demanding and only certify large shocks.
  • ad hoc to paper Two-stage heuristic approximates the joint nonconvex problem
    §3.4 explicitly calls the pipeline a heuristic approximation; Theorem 2 covers only a linear-detrender idealization.
  • domain assumption Known future promotional/SNAP calendar
    Conditional framework; when calendar unknown, CD reverts to baseline-only forecasting (Limitations).
  • domain assumption Base-stock inventory simulation parameters
    Lead time 7 days, p=$10/unit, α=0.95 fixed a priori; crossover estimate depends on these conventions.

pith-pipeline@v1.3.0-alltime-deepseek · 39711 in / 10506 out tokens · 110895 ms · 2026-08-01T01:48:10.488945+00:00 · methodology

0 comments
read the original abstract

Machine learning demand forecasts optimize statistical accuracy yet leave excess operational volatility that inflates safety stock and amplifies the Bullwhip effect. We introduce \textbf{Contextual Deconvolution} (CD), a two-stage estimator that reframes demand sensing as a convex decomposition: a kernel-modulated banded operator separates transient promotion-driven shocks from a smooth structural baseline, and hierarchical partial pooling enables catalog-scale deployment without per-SKU training. The operator is data-derived, not imposed---it reduces to the identity wherever the promotional response is impulsive (most of M5, all of Favorita) and contributes only where genuine multi-day carryover exists, so the gains rest on the structural decomposition itself. Evaluating strictly out-of-sample on 30,490 M5 SKUs and 2,845 Favorita items, with calendar-aware baselines given CD's identical future calendar, we anchor the contribution on a full inventory-cost accounting: CD lowers safety stock, holding cost, and order variance but under-provisions event spikes, reducing total cost only when holding costs exceed $\sim$20\% of stockout costs (95\% CI $[17\%,25\%]$); otherwise it is an operational-stability and inventory-capital layer, not an expected-cost minimizer. Its accuracy contribution is reliability rather than central tendency: across eleven baselines, CD attains the lowest cross-sectional dispersion of per-SKU error and mis-forecasts by more than 200\% on 0.8\% of SKUs versus 9.9--20.6\% for every baseline, ranking first on both in all four M5 draws. Because the Variance Ratio and std-based safety stock are minimized by any sufficiently smooth forecast, we treat them as diagnostics, not objectives. A supporting analysis shows the learned demand operators are non-normal, yet CD's compact parametric kernel matches their operational performance interpretably.

Figures

Figures reproduced from arXiv: 2607.25664 by Mohammad Forouhesh.

Figure 1
Figure 1. Figure 1: The statistical-operational tradeoff. Standard ML forecasts (red, dashed) optimize point-wise [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Cost-ratio sweep on 1,000 M5 SKUs (strictly out-of-sample): median [PITH_FULL_IMAGE:figures/full_fig_p014_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Left: Gamma kernel (always positive) fails to capture the post-event dip on synthetic signed ground-truth data. Right: Calendar corruption on 50 M5 SKUs (95% CI): CD wMAPE stays flat (1.019 → 1.039 across 0–50% corruption), VR near zero (0.002–0.008); Vanilla XGBoost, by contrast, has median VR two orders of magnitude higher even with the uncorrupted calendar (Finding 1). Scalability and Systems. The full … view at source ↗
Figure 4
Figure 4. Figure 4: Two-Stream Forecasting Architecture. Stream A extrapolates the smooth structural baseline; [PITH_FULL_IMAGE:figures/full_fig_p034_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Synthetic kernel recovery benchmark. (a) Ground-truth DoE kernel. (b) CCF with identity [PITH_FULL_IMAGE:figures/full_fig_p035_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Favorita Variance Ratio distribution (corrected live on-promotion and holiday channels), on the [PITH_FULL_IMAGE:figures/full_fig_p037_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Real kernel recovery on M5 (category-level empirical means with IQR bands). Blue solid: empirical [PITH_FULL_IMAGE:figures/full_fig_p038_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: EDMD vs. parametric kernel on 500 M5 SKUs. (a) Representative SKU: category-level EDMD [PITH_FULL_IMAGE:figures/full_fig_p040_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Pseudospectra of learned Koopman operators ( [PITH_FULL_IMAGE:figures/full_fig_p042_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Transient growth of the Koopman semigroup [PITH_FULL_IMAGE:figures/full_fig_p043_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Pseudospectra of learned Koopman operators on Favorita ( [PITH_FULL_IMAGE:figures/full_fig_p045_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Transient growth of the Koopman semigroup [PITH_FULL_IMAGE:figures/full_fig_p046_12.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

175 extracted references · 7 linked inside Pith

  1. [1]

    Sloan Management Review , volume=

    The Bullwhip Effect in Supply Chains , author=. Sloan Management Review , volume=. 1997 , publisher=

    1997

  2. [2]

    1961 , publisher=

    Industrial Dynamics , author=. 1961 , publisher=

    1961

  3. [3]

    Quantitative Models for Supply Chain Management , pages=

    The Bullwhip Effect: Managerial Insights on the Impact of Forecasting and Information on Variability in a Supply Chain , author=. Quantitative Models for Supply Chain Management , pages=. 2000 , publisher=

    2000

  4. [4]

    Managing Supply Chain Operations , year=

    The Bullwhip Effect: Managerial Insights on the Impact of Forecasting and Information on Variability in a Supply Chain , author=. Managing Supply Chain Operations , year=

  5. [5]

    Oreshkin, Boris N and Carpov, Dmitri and Chapados, Nicolas and Bengio, Yoshua , journal=

  6. [7]

    2020 , publisher=

    Salinas, David and Flunkert, Valentin and Gasthaus, Jan and Januschowski, Tim , journal=. 2020 , publisher=

    2020

  7. [8]

    Makridakis, Spyros and Spiliotis, Evangelos and Assimakopoulos, Vassilios , journal=. The. 2018 , publisher=

    2018

  8. [9]

    Cleveland, Robert B and Cleveland, William S and McRae, Jean E and Terpenning, Irma , journal=

  9. [10]

    2016 , publisher=

    Chen, Tianqi and Guestrin, Carlos , booktitle=. 2016 , publisher=

    2016

  10. [11]

    The American Statistician , volume=

    Forecasting at Scale , author=. The American Statistician , volume=. 2018 , publisher=

    2018

  11. [12]

    1970 , publisher=

    Time Series Analysis: Forecasting and Control , author=. 1970 , publisher=

    1970

  12. [13]

    2016 , publisher=

    Time Series Analysis: Forecasting and Control , author=. 2016 , publisher=

    2016

  13. [14]

    Makridakis, Spyros and Spiliotis, Evangelos and Assimakopoulos, Vassilios , journal=. The. 2020 , publisher=

    2020

  14. [15]

    Makridakis, Spyros and Spiliotis, Evangelos and Assimakopoulos, Vassilios , journal=. The. 2022 , publisher=

    2022

  15. [16]

    2018 , howpublished=

    Corporaci\'. 2018 , howpublished=

    2018

  16. [17]

    2020 , howpublished=

    2020

  17. [18]

    Automatic Time Series Forecasting: The Forecast Package for

    Hyndman, Rob J and Khandakar, Yeasmin , journal=. Automatic Time Series Forecasting: The Forecast Package for

  18. [19]

    Management Science , volume=

    Modeling Managerial Behavior: Misperceptions of Feedback in a Dynamic Decision Making Experiment , author=. Management Science , volume=. 1989 , publisher=

    1989

  19. [20]

    Operations Research , volume=

    The Stabilization Effect of Inventory in Supply Chains , author=. Operations Research , volume=. 1995 , publisher=

    1995

  20. [21]

    American Economic Review , volume=

    Inventories and the Volatility of Production , author=. American Economic Review , volume=

  21. [22]

    The Variability of Aggregate Demand with (

    Caplin, Andrew S , journal=. The Variability of Aggregate Demand with (

  22. [23]

    Manufacturing & Service Operations Management , volume=

    A Single-Item Inventory Model for a Nonstationary Demand Process , author=. Manufacturing & Service Operations Management , volume=. 1999 , publisher=

    1999

  23. [24]

    An Empirical Analysis of the Bullwhip Effect in the

    Gaur, Vishal and Fisher, Marshall L and Raman, Ananth , journal=. An Empirical Analysis of the Bullwhip Effect in the. 2005 , publisher=

    2005

  24. [25]

    Interfaces , volume=

    How Do You Win the Capstone Forecasting Competition? , author=. Interfaces , volume=

  25. [26]

    Journal of Business & Economic Statistics , volume=

    Structural Demand Information and the Improving Model-Data Fit of Dynamic Econometric Models , author=. Journal of Business & Economic Statistics , volume=

  26. [27]

    OR Spectrum , volume=

    Planning Safety Stocks in Material Requirements Planning Systems , author=. OR Spectrum , volume=

  27. [28]

    Naval Research Logistics , volume=

    Demand Estimation in Lost Sales Inventory Systems , author=. Naval Research Logistics , volume=

  28. [29]

    Sloan Management Review , volume=

    Traps in Supply Chain Management , author=. Sloan Management Review , volume=

  29. [30]

    Management Science , volume=

    The Value of Fast Fashion: Quick Response, Enhanced Design, and Strategic Consumer Behavior , author=. Management Science , volume=

  30. [31]

    Harvard Business Review , volume=

    What is the Right Supply Chain for Your Product? , author=. Harvard Business Review , volume=

  31. [32]

    Demand Management: The Evaluation of Demand Management Practices , author=

  32. [33]

    2000 , publisher=

    Designing and Managing the Supply Chain: Concepts, Strategies, and Case Studies , author=. 2000 , publisher=

    2000

  33. [34]

    Journal of the Royal Statistical Society: Series B (Methodological) , volume=

    Regression Shrinkage and Selection via the Lasso , author=. Journal of the Royal Statistical Society: Series B (Methodological) , volume=

  34. [35]

    Journal of Fourier Analysis and Applications , volume=

    Enhancing Sparsity by Reweighted _1 Minimization , author=. Journal of Fourier Analysis and Applications , volume=

  35. [36]

    2004 , publisher=

    Convex Optimization , author=. 2004 , publisher=

    2004

  36. [37]

    SIAM Journal on Imaging Sciences , volume=

    A Fast Iterative Shrinkage-Thresholding Algorithm for Linear Inverse Problems , author=. SIAM Journal on Imaging Sciences , volume=

  37. [38]

    The Annals of Mathematical Statistics , volume=

    Robust Estimation of a Location Parameter , author=. The Annals of Mathematical Statistics , volume=

  38. [39]

    2018 , publisher=

    Forecasting: Principles and Practice , author=. 2018 , publisher=

    2018

  39. [40]

    Journal of the Operational Research Society , volume=

    Stock Control with Sporadic and Slow-Moving Demand , author=. Journal of the Operational Research Society , volume=

  40. [41]

    European Journal of Operational Research , volume=

    Optimal Inventory Policies under Demand Forecasts , author=. European Journal of Operational Research , volume=

  41. [42]

    Journal of Business & Economic Statistics , year=

    Machine Learning and Causal Inference for Demand Forecasting , author=. Journal of Business & Economic Statistics , year=

  42. [43]

    Computers & Industrial Engineering , volume=

    Demand Forecasting in the Era of Machine Learning , author=. Computers & Industrial Engineering , volume=

  43. [44]

    Interfaces , volume=

    Zara Uses Operations Research to Reengineer Its Global Distribution Process , author=. Interfaces , volume=

  44. [45]

    Supply Chain Management Review , year=

    Achieving Accurate and Order-of-Magnitude Inventory Reductions , author=. Supply Chain Management Review , year=

  45. [46]

    Management Science , volume=

    Information Sharing and Order Variability in a Supply Chain , author=. Management Science , volume=

  46. [47]

    Manufacturing & Service Operations Management , volume=

    Inventory Record Inaccuracy in Retail: An Exploratory Analysis , author=. Manufacturing & Service Operations Management , volume=

  47. [48]

    2012 , publisher=

    Handbook of Newsvendor Problems: Models, Extensions and Applications , author=. 2012 , publisher=

    2012

  48. [49]

    A Tutorial on Supply Chain Management , author=

  49. [50]

    Econometrica , volume=

    A Formal Theory of the Employment Relationship , author=. Econometrica , volume=

  50. [51]

    2009 , edition=

    The Elements of Statistical Learning: Data Mining, Inference, and Prediction , author=. 2009 , edition=

    2009

  51. [52]

    International Journal of Forecasting , volume=

    Temporal Fusion Transformers for Interpretable Multi-horizon Time Series Forecasting , author=. International Journal of Forecasting , volume=. 2021 , publisher=

    2021

  52. [53]

    Challu, Cristian and Olivares, Kin G and Winkler, Max and Wang, Tian and Garza, Federico and Mergenthaler, Tim and Woen, Tian , journal=

  53. [54]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  54. [55]

    Advances in Neural Information Processing Systems , volume=

    Autoformer: Decomposition Transformers with Auto-Correlation for Long-Term Series Forecasting , author=. Advances in Neural Information Processing Systems , volume=

  55. [56]

    arXiv preprint arXiv:2403.07815 , year=

    Chronos: Learning the Language of Time Series , author=. arXiv preprint arXiv:2403.07815 , year=

    Pith/arXiv arXiv

  56. [58]

    arXiv preprint arXiv:2402.02592 , year=

    Moirai: A Time Series Foundation Model for Universal Forecasting , author=. arXiv preprint arXiv:2402.02592 , year=

    Pith/arXiv arXiv

  57. [59]

    arXiv preprint arXiv:2310.08278 , year=

    Lag-Llama: Towards Foundation Models for Time Series Forecasting , author=. arXiv preprint arXiv:2310.08278 , year=

    Pith/arXiv arXiv

  58. [60]

    arXiv preprint arXiv:2310.03589 , year=

    TimeGPT-1 , author=. arXiv preprint arXiv:2310.03589 , year=

    Pith/arXiv arXiv

  59. [61]

    Proceedings of the Twenty-Ninth International Joint Conference on Artificial Intelligence , pages=

    Transformers in Time Series: A Survey , author=. Proceedings of the Twenty-Ninth International Joint Conference on Artificial Intelligence , pages=

  60. [62]

    Proceedings of the 11th International Conference on Learning Representations , year=

    Koopman Neural Operator Forecaster for Time-Series with Temporal Distributional Shifts , author=. Proceedings of the 11th International Conference on Learning Representations , year=

  61. [63]

    International Journal of Forecasting , volume=

    Deep Learning for Time Series Forecasting: The Electric Angle , author=. International Journal of Forecasting , volume=

  62. [64]

    ACM Computing Surveys , year=

    A Survey on Time-Series Forecasting with Deep Learning: Applications, Models, and Challenges , author=. ACM Computing Surveys , year=

  63. [65]

    ACM Computing Surveys , year=

    Deep Learning for Time Series Forecasting: A Tutorial and Literature Survey , author=. ACM Computing Surveys , year=

  64. [66]

    Proceedings of the National Academy of Sciences , volume=

    Operational Methods in Quantum Mechanics , author=. Proceedings of the National Academy of Sciences , volume=

  65. [67]

    PLOS ONE , volume=

    Koopman Invariant Subspaces and Finite Linear Representations of Nonlinear Dynamical Systems for Control , author=. PLOS ONE , volume=

  66. [68]

    Chaos: An Interdisciplinary Journal of Nonlinear Science , volume=

    Extended Dynamic Mode Decomposition with Dictionary Learning: A Data-Driven Adaptive Spectral Decomposition of the Koopman Operator , author=. Chaos: An Interdisciplinary Journal of Nonlinear Science , volume=

  67. [69]

    Nature Communications , volume=

    VAMPnets for Deep Learning of Molecular Kinetics , author=. Nature Communications , volume=

  68. [70]

    Journal of the ACM , volume=

    Robust Principal Component Analysis? , author=. Journal of the ACM , volume=

  69. [71]

    SIAM Journal on Optimization , volume=

    Rank-Sparsity Incoherence for Matrix Decomposition , author=. SIAM Journal on Optimization , volume=

  70. [72]

    SIAM Review , volume=

    \ _1 Trend Filtering , author=. SIAM Review , volume=

  71. [73]

    The Annals of Statistics , volume=

    Adaptive Piecewise Polynomial Estimation via Trend Filtering , author=. The Annals of Statistics , volume=

  72. [74]

    Journal of Retailing and Consumer Services , volume=

    Estimation of Promotion Effects in Scanner Data using a Robust Baseline Approach , author=. Journal of Retailing and Consumer Services , volume=

  73. [75]

    Marketing Science , volume=

    Pull-Through and the Dynamic Effects of Promotions on Brand Choice , author=. Marketing Science , volume=

  74. [76]

    Marketing Science , volume=

    Promotion Dynamics in the Marketing Mix , author=. Marketing Science , volume=

  75. [77]

    Handbook of Marketing , year=

    Sales Promotion , author=. Handbook of Marketing , year=

  76. [78]

    Handbook of Marketing , year=

    Sales Promotion Models: A Review , author=. Handbook of Marketing , year=

  77. [79]

    Marketing Science , volume=

    Variety-Seeking, Purchase Timing, and the Lightening Effect , author=. Marketing Science , volume=

  78. [80]

    Marketing Science , volume=

    The Shopping Incidence Model: A Model of Purchase Event Frequency , author=. Marketing Science , volume=

  79. [81]

    Journal of Econometrics , volume=

    Generalized Autoregressive Conditional Heteroskedasticity , author=. Journal of Econometrics , volume=

  80. [82]

    Econometrica , volume=

    Autoregressive Conditional Heteroscedasticity with Estimates of the Variance of United Kingdom Inflation , author=. Econometrica , volume=

Showing first 80 references.