Event-preserving feature engineering for intermittent demand forecasting using SHOS Authors: B Sendhil Nathan, Veera Siva Reddy B, C Chandrasekhara Sastry, Sachin Salunkhe, Robert Cep Published: 2026 ABSTRACT AbstractIntermittent demand forecasting remains a persistent challenge in large-scale supply chains due to extreme sparsity, irregular demand occurrence, and highly variable demand magnitudes. While recent studies have increasingly emphasized architectural complexity to address these challenges, comparatively less attention has been directed toward the role of statistically grounded feature representation under sparse-demand conditions. This study introduces the Smoothed Hybrid Occurrence-Size (SHOS) framework as an event-preserving feature-engineering approach for intermittent-demand forecasting. SHOS decomposes demand into latent occurrence probability and conditional demand size components and generates adaptive, series-specific statistical representations using sparsity-aware exponential smoothing. Rather than functioning as a standalone forecasting model, SHOS operates as a feature-generation mechanism that embeds occurrence-size structure directly into supervised learning pipelines. The proposed framework was evaluated using a large-scale automotive aftermarket dataset consisting of approximately 56,000 dealer-part time series and approximately 1.4 million monthly observations generated after preprocessing and zero-padding. Models were evaluated using rolling-window cross-validation, signal-preservation analysis, and robustness studies across heterogeneous demand segments. Under the evaluated sparse-demand setting and model configurations, SHOS-augmented tree-based models substantially improved forecasting accuracy and stability relative to raw-feature baselines. In particular, the SHOS-augmented LightGBM configuration reduced the mean absolute error by approximately 46.3% relative to the baseline model, while the weighted mean absolute percentage error improved by more than 40%. Beyond conventional forecasting accuracy, the study additionally demonstrates that SHOS preserves event timing, peak structure, and frequency-domain characteristics that are often distorted by conventional smoothing approaches. The results further suggest that statistically grounded feature representations can substantially improve sparse-demand learning behaviour under the evaluated automotive aftermarket forecasting environment. Rather than establishing a universal hierarchy between representation and architectural complexity, the findings demonstrate the effectiveness of event-preserving representation-oriented forecasting under highly intermittent demand conditions.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1038/s41598-026-57129-6.Keywords: Event preservation, Feature engineering, Intermittent demand forecasting, Machine learning, SHOS, Sparse time seriesSubject terms: Engineering, Mathematics and computing Introduction Intermittent demand forecasting remains a major challenge in supply chain and inventory management, due to long periods of zero demand interspersed with irregular, low-frequency demand events. This problem is especially critical in automotive aftermarket and spare-parts systems, where organizations manage large populations of slow-moving stock-keeping units (SKUs) across distributed dealer networks1–3. Under such conditions, inaccurate forecasts can simultaneously increase inventory holding costs, obsolescence risk, and service-level failures. Classical intermittent-demand forecasting methods such as Croston’s method, the Syntetos-Boylan Approximation (SBA), and the Teunter-Syntetos-Babai (TSB) framework address this challenge by decomposing demand into occurrence and conditional size components4–6. These approaches remain widely used because they provide stable and interpretable forecasting behaviour under sparse-demand conditions. However, they operate primarily as univariate forecasting procedures and cannot directly exploit auxiliary variables or cross-sectional information increasingly available in modern enterprise systems. Recent forecasting research has therefore increasingly explored machine-learning and deep-learning approaches capable of learning nonlinear relationships from large multivariate datasets7–10. Tree-based learners such as LightGBM and XGBoost have demonstrated strong predictive performance in tabular forecasting problems, while deep learning architectures aim to capture complex temporal structure7,8. However, intermittent-demand environments remain challenging for such models due to extreme sparsity, irregular demand occurrences, and weak temporal continuity, which limit the availability of stable learning signals. At the same time, forecasting competitions and hybrid forecasting studies increasingly suggest that statistically meaningful feature representations may substantially improve forecasting performance even without highly specialized architectures11,12. This raises an important question in intermittent-demand forecasting: can statistically grounded feature representations improve sparse-demand learning behaviour more effectively than increasing architectural complexity alone. To address this problem, this study introduces the Smoothed Hybrid Occurrence-Size (SHOS) framework as an event-preserving feature-engineering approach for intermittent-demand forecasting. SHOS decomposes intermittent demand into a latent-occurrence probability and a conditional-demand-size component via sparsity-aware adaptive smoothing. Rather than functioning as a standalone forecasting model, SHOS serves as a representation-generation framework whose outputs are directly incorporated into supervised learning pipelines. The proposed framework is evaluated using a large-scale automotive aftermarket dataset containing approximately 56,000 dealer-part time series and approximately 1.4 million monthly observations after preprocessing and zero-padding. Forecasting models are evaluated using rolling-window cross-validation together with signal-preservation and robustness analyses under highly intermittent demand conditions. In the present study, the term “event-preserving” refers to the ability of the SHOS representation framework to retain meaningful demand-event characteristics, including occurrence timing, peak magnitude behaviour, and frequency-domain structure, while reducing instability arising from extreme sparsity. The results demonstrate that SHOS-derived representations substantially improve forecasting accuracy and stability under the evaluated sparse-demand setting while preserving event timing and demand-structure characteristics that are often distorted by conventional smoothing approaches. Rather than proposing a universal replacement for complex architectures, the study investigates the role of statistically grounded feature representation within the evaluated automotive aftermarket forecasting environment. Literature review Classical intermittent-demand forecasting Intermittent-demand forecasting methods are fundamentally based on the principle that demand occurrence and demand magnitude should be modelled separately. Croston’s method introduced this decomposition by independently smoothing inter-demand intervals and positive demand sizes4. Subsequent studies demonstrated that Croston’s estimator exhibits positive statistical bias, motivating several refinements13. The Syntetos-Boylan Approximation (SBA) introduced a debiasing correction that improved forecast accuracy and inventory performance under intermittent demand5,14. Later, the Teunter-Syntetos-Babai (TSB) method replaced interval smoothing with direct smoothing of demand occurrence probability, enabling improved responsiveness to changes in demand frequency6,15. These approaches remain foundational because they explicitly preserve the occurrence-size structure characteristic of intermittent demand. Despite their interpretability and robustness, classical methods remain limited by their primarily univariate formulation. They cannot naturally incorporate rich auxiliary information, such as product attributes, dealer behaviour, or cross-sectional demand patterns, which are increasingly available in modern supply-chain systems. Machine learning and deep learning approaches Machine-learning and deep-learning methods have increasingly been adopted for forecasting due to their ability to learn nonlinear relationships from large-scale multivariate data7–10. Tree-based ensemble learners such as LightGBM and XGBoost have demonstrated strong predictive performance in structured forecasting environments7,8, while neural network architectures aim to capture complex temporal dynamics16. However, highly intermittent demand presents substantial challenges for purely data-driven learning approaches. Long sequences of zero demand reduce signal availability, increase noise sensitivity, and may destabilize sequential learning behaviour9,10. As a result, machine-learning performance under extreme sparsity often depends strongly on the quality of the underlying feature representation rather than solely on architectural complexity. Hybrid forecasting and representation-learning gap Recent forecasting research increasingly emphasizes hybridization strategies that combine statistical forecasting structure with machine-learning flexibility11,12,17,18. Existing hybrid intermittent-demand approaches generally integrate statistical estimators with machine-learning or deep-learning models through multi-stage, ensemble, or hurdle-based architectures. Examples include Croston-LSTM hybrids, SBA-based feature integration within tree-based learners, and two-stage occurrence-size forecasting frameworks. While these methods often improve forecasting performance, they typically rely on architectural decomposition or model stacking to capture the structure of intermittent demand. More recently, feature-based forecasting approaches have explored embedding statistical information directly into machine-learning pipelines as informative representations rather than treating it as a separate forecasting stage18. Nevertheless, there remains limited research on adaptive, statistically grounded feature representations specifically designed for highly sparse intermittent-demand environments. The SHOS framework proposed in this study addresses this gap by generating adaptive occurrence-size representations that are incorporated directly into supervised learning models as input features. Unlike multi-stage hybrid architectures, SHOS performs hybridization at the representation level while preserving computational simplicity and forecasting scalability under sparse-demand conditions. SHOS framework and theoretical formulation Problem formulation The foundational work on intermittent demand forecasting4,6, a time series of demand is modelled as the product of two latent processes: an occurrence process and a size process. Let denote the observed demand quantity in period t, with indicating whether a demand event occurred, and representing the conditional size. The forecasting objective is to estimate the conditional expectation: where is the information set available up to time t. This work focuses on generating robust, series-specific estimates of these two components and incorporating them as features into machine learning models, rather than directly optimizing a standalone forecasting equation. In this formulation, the forecasting objective is to estimate the conditional expectation of demand given the information set , which includes all historical demand observations and available covariates up to time t-1. Theoretical basis of SHOS Intermittent demand processes are fundamentally characterized by two distinct sources of uncertainty: whether demand occurs in a given period, and the magnitude of demand conditional on occurrence. This structure is naturally represented as a point process, where observed demand is generated by the interaction of a binary occurrence process and a conditional size process. Classical intermittent-demand models such as Croston-type estimators are built on this decomposition, reflecting the renewal nature of demand events separated by irregular intervals of inactivity. The SHOS framework adopts this same conceptual foundation, explicitly treating demand occurrence and demand size as separate latent components that evolve over time. SHOS suppresses stochastic sparsity-induced noise without introducing substantial phase lag, peak attenuation, or distortion of intermittent event structure. Although SHOS is conceptually motivated by the occurrence-size decomposition principle introduced in Croston-type intermittent-demand methods, the SHOS framework does not directly reuse Croston forecast outputs as replacement predictions or embedded forecast estimators. Instead, SHOS independently generates adaptive latent representations of occurrence probability and conditional demand size through sparsity-aware smoothing and recency-adaptive updating mechanisms operating on the zero-padded demand series. A central challenge in intermittent demand forecasting arises from the amplification of variance under extreme sparsity. When long sequences of zero demand dominate the series, direct smoothing or regression on raw demand values conflates non-occurrence with small magnitudes, leading to unstable estimates and high noise sensitivity. SHOS addresses this issue by smoothing the occurrence process independently of the size process. By updating the probability of demand occurrence at every time step, including zero-demand periods, SHOS stabilizes frequency estimation and prevents isolated demand events from exerting disproportionate influence. Conditional demand size, in contrast, is updated only when demand occurs, preserving meaningful magnitude information without contamination from inactivity periods. This separation acts as a variance-stabilizing mechanism, particularly effective in highly sparse regimes. The SHOS Decomposition and representation are illustrated in Fig. 1. Illustration of the SHOS framework decomposing raw demand into latent components that preserve event timing and peak magnitude without phase lag. Importantly, SHOS does not perform smoothing via convolution or moving-average operations on the raw demand series. Instead, it applies event-preserving smoothing at the level of latent demand components. The timing of demand events is retained through the occurrence process, while magnitude information is preserved through conditional size updates. As a result, SHOS suppresses stochastic noise arising from sparsity without introducing phase lag or attenuating genuine demand bursts. This distinction explains why SHOS fundamentally differs from low-pass filtering approaches and why it can retain cyclical and burst-driven structure while improving statistical stability. These theoretical properties motivate the use of SHOS-derived occurrence and size estimates as statistically grounded input features for downstream machine learning models. Unlike classical Croston-type estimators that explicitly model inter-demand intervals as renewal durations, the SHOS framework operates by adaptively smoothing latent occurrence probability and conditional demand size representations over a fixed temporal grid. Consequently, SHOS is less dependent on exact interval estimation and instead focuses on generating statistically stable occurrence-size representations suitable for supervised sparse-demand learning. The SHOS algorithm: adaptive statistical priors The SHOS (Smoothed Hybrid Occurrence-Size) algorithm is an adaptive statistical feature generation framework designed to operate robustly under extreme demand intermittency, where time series are dominated by long sequences of zero demand punctuated by sporadic, irregular non-zero observations. The algorithm decomposes intermittent demand into two latent components estimated at each time step: the probability of demand occurrence and the conditional demand size given that demand occurs. This formulation allows SHOS to update continuously over time, including during prolonged zero-demand periods, which is essential for maintaining stable estimates in highly sparse demand environments. The overall logical flow of the SHOS algorithm is illustrated in Fig. 2. The SHOS algorithm explicitly leverages zero-filled observations as informative indicators of demand non-occurrence, enabling continuous updating of occurrence probability during inactivity periods and ensuring statistically consistent estimation under extreme sparsity19. Algorithmic flow of the SHOS framework, illustrating adaptive smoothing, recency adjustment, and occurrence size updates. Algorithm. SHOS feature generation procedure. Note: In the rolling window cross-validation protocol, all statistics required for initialization (e.g., global mean of positive demands) are computed using only the training window of the corresponding fold to avoid leakage. Let denote the observed demand quantity in the time period t, and let be a binary indicator that equals 1 when demand occurs and 0 otherwise. SHOS maintains a smoothed estimate of the probability of demand occurrence, denoted by , and a smoothed estimate of the conditional demand size, denoted by . The SHOS point forecast of expected demand is given by Smoothed occurrence probability and conditional size In highly intermittent demand series, naive estimators of occurrence probability and demand magnitude are often unstable or unresponsive due to the dominance of zero observations. SHOS mitigates this issue by explicitly updating the demand occurrence probability at every time step, including periods with zero demand. This ensures that the estimated occurrence probability decays gradually during inactivity rather than remaining fixed, thereby preventing persistent overestimation following isolated demand events. The occurrence probability is updated using exponential smoothing as where is the smoothing parameter controls the responsiveness of the occurrence estimate. The conditional demand size is updated only when positive demand is observed, thereby avoiding contamination of magnitude estimates by zero demand periods. Specifically, the update rule is defined as where is the smoothing parameter for the conditional size process. By separating the update mechanisms for occurrence and size, SHOS ensures that magnitude estimates remain statistically meaningful even when non-zero observations are rare. Adaptive smoothing, sparsity handling, and initialization To further stabilize estimation under anomalous intermittent demand patterns, SHOS employs sparsity adaptive smoothing. Series-specific sparsity is quantified using the proportion of zero demand periods over the observed horizon T, defined as The effective smoothing parameters are dynamically adjusted as a function of this sparsity measure according to where and are base smoothing values and k is a sparsity-scaling factor that controls the strength of adaptation. In this study, and k=0.1, which provides stronger smoothing for highly sparse series while preserving responsiveness for less intermittent ones. To improve robustness in short or extremely sparse series, SHOS employs a conservative initialization strategy. Initial estimates and are computed from the first three observed periods when sufficient positive demand exists; otherwise, they default to the global mean of positive demand across the dataset. In addition, a recency-aware adjustment mechanism temporarily increases the effective smoothing weights (subject to an upper bound of 0.9) when demand occurs after an extended period of zero demand. This allows SHOS to respond rapidly to genuine demand reactivation without destabilizing long-term estimates20. Sensitivity of SHOS to the sparsity scaling parameter The sparsity-scaling parameter (k) controls the extent to which SHOS adapts its smoothing behaviour as intermittency increases. To evaluate whether this parameter requires retuning across different demand regimes, a sensitivity analysis was conducted across representative SKU categories defined by demand movement intensity, intermittency, and lifecycle stage. In addition to adaptive SHOS configurations, the analysis also included the case (k = 0), corresponding to static SHOS without sparsity-adaptive smoothing. For fast-moving items, characterized by frequent demand occurrences and relatively low sparsity, forecasting performance varies gradually with respect to the sparsity-scaling parameter. As shown in Table 1; Fig. 3, MAE and WMAPE remain relatively stable for k ∈ [0,0.2], with only mild degradation observed at larger values. Static SHOS ((k = 0)) already produces competitive performance for dense demand regimes, indicating that the primary forecasting gains arise from the underlying occurrence-size representation rather than from aggressive sparsity adaptation. Increasing (k) beyond moderate levels introduces additional smoothing, which slightly reduces responsiveness to frequent demand fluctuations21. Table 1Sensitivity of SHOS forecasting performance to the sparsity-scaling parameter k across SKU categories defined by demand value, intermittency, and lifecycle stage.CategorykMAERMSEWMAPE (%)MAPE (%)R²Fast-movers0.00 StaticSHOS0.38152.554027.8233.790.75650.010.38282.563427.9933.980.75530.050.38602.597128.2234.510.74880.100.38942.606028.4734.900.74710.200.39692.576229.0235.300.75280.500.41922.636030.6538.030.7412Slow-movers0.00 StaticSHOS0.000640.00821.521.490.99830.010.000590.00771.401.370.99850.050.000610.00801.461.440.99840.100.000520.00691.251.220.99880.200.000420.00541.000.970.99930.500.000080.00310.190.090.9998Lifecycle: new0.00 StaticSHOS0.03760.691019.1814.920.72150.010.03810.677719.5415.110.72480.050.03960.647420.3015.450.74890.100.03990.652620.4615.680.74480.200.04120.689821.1416.580.71490.500.04550.665723.3618.060.7344Lifecycle: mature0.00 StaticSHOS1.02484.487026.8635.020.78080.011.02684.502426.9435.170.77980.051.02944.526827.0135.430.77740.101.03284.500227.1035.480.78000.201.04614.573827.4436.130.77270.501.10904.527229.1038.700.7773Lifecycle: declined-No SKU found----The case k = 0 corresponds to static SHOS without sparsity-adaptive smoothing. Sensitivity of SHOS forecasting performance to the sparsity-scaling parameter k across SKU categories defined by demand value, intermittency, and lifecycle stage. 0.00 Static SHOS 0.00 Static SHOS 0.00 Static SHOS 0.00 Static SHOS The case k = 0 corresponds to static SHOS without sparsity-adaptive smoothing. Sensitivity of SHOS forecasting error (WMAPE) to the sparsity scaling parameter across SKU categories, the point corresponds to static SHOS (no sparsity adaptation). For slow-moving items exhibiting extreme intermittency and prolonged zero-demand intervals, adaptive sparsity scaling provides a stabilizing effect. Increasing values of (k) progressively improve MAE, WMAPE, and R2, with the strongest performance observed at higher smoothing levels. In this regime, adaptive smoothing reduces estimation instability caused by sparse demand occurrence while preserving event-level responsiveness when demand reappears. The comparison between static SHOS and adaptive SHOS shows that sparsity-aware scaling provides incremental robustness under severe intermittency conditions, rather than serving as the sole source of forecasting improvement. A similar trend is observed across lifecycle-based SKU segmentation. Newly introduced parts show relatively modest variation in forecasting accuracy across different values of (k), remaining within a narrow error band despite changes in smoothing intensity. Mature lifecycle parts exhibit near-flat forecasting behaviour for k ∈ [0,0.2], followed by gradual degradation at larger values. These observations indicate that SHOS generalizes consistently across different lifecycle stages without requiring aggressive lifecycle-specific parameter tuning22. No SKUs satisfied the minimum activity criteria for the declined lifecycle category during the evaluation window, so this segment was excluded from the sensitivity analysis. Overall, the sensitivity analysis demonstrates that SHOS performance varies smoothly rather than abruptly with respect to the sparsity-scaling parameter across heterogeneous SKU groups. The results additionally indicate that adaptive sparsity scaling functions primarily as an incremental stabilization mechanism, while the dominant forecasting gains arise from the combined occurrence-size decomposition and event-preserving feature representation framework. Based on the balance between stability under sparse demand conditions and responsiveness under dense demand regimes, a fixed value of (k = 0.1) was adopted throughout this study as a stable and operationally practical default configuration. SHOS versus TSB: bias mitigation and adaptivity Although SHOS and the Teunter-Syntetos-Babai (TSB) method share a common conceptual foundation by modelling demand as the product of an occurrence probability and a conditional size, their estimation mechanisms differ fundamentally. In TSB, both components are updated using fixed exponential smoothing parameters, implicitly assuming homogeneous intermittency and stationarity across all series. While effective under moderate intermittency, this assumption can lead to delayed responsiveness and variance-induced bias in highly sparse or non-stationary demand environments. SHOS addresses these limitations by introducing sparsity-adaptive smoothing, in which the degree of smoothing is dynamically scaled to series-specific intermittency. This reduces overreaction to isolated demand events in highly sparse series while maintaining responsiveness in less intermittent ones. Furthermore, the recency-sensitive adjustment mechanism enables SHOS to adapt more rapidly than TSB when demand reappears after prolonged inactivity, mitigating the delayed-response bias commonly observed in fixed-parameter smoothing approaches23. Finally, the robust initialization strategy employed by SHOS stabilizes early estimates in short or highly intermittent series, further reducing estimation bias relative to classical methods. Collectively, these mechanisms enable SHOS to generate stable, informative, and adaptive estimates of demand occurrence probability and conditional size, making it particularly well suited for feature generation in anomalous intermittent demand scenarios. Experimental study design This section describes the methodological framework for studying event-preserving feature representation for intermittent demand forecasting. Rather than emphasizing model architecture, the experimental design is structured to isolate the effect of feature representation on learning behaviour under sparsity. Figure 4 provides a conceptual overview of the proposed SHOS feature representation. Starting from raw intermittent demand, the figure illustrates how conventional smoothing distorts demand events through averaging and phase lag, whereas SHOS instead decomposes demand into a smoothed occurrence probability and a conditional demand size, preserving event structure. These SHOS-derived signals are subsequently used as input features within a single-stage, architecture-agnostic supervised learning model. SHOS feature representation logic (event-preserving decomposition and integration). The experimental workflow described in the following subsections operationalizes this representation. Raw transactional data are aggregated to a monthly frequency and zero-padded to establish a consistent temporal grid. A comprehensive feature engineering process then generates a rich set of predictors, including SHOS-derived statistical features. Forecasting models are trained under a rolling-window cross-validation protocol that respects temporal ordering, enabling controlled evaluation of representation effects across heterogeneous demand regimes. Dataset description The experiments use a large-scale automotive aftermarket intermittent-demand dataset derived from dealer-level spare-parts sales transactions collected from a single national automotive aftermarket network. The original transactional records span the period from August 2023 to August 2025 and include approximately 59 dealers and 7,702 unique spare parts. Following preprocessing and dealer-part filtering, the dataset yielded approximately 56,000 active dealer-part demand series for forecasting analysis. Because raw transactional demand is extremely sparse and irregular, the original daily sales records were aggregated to a monthly frequency prior to forecasting analysis. Monthly aggregation was selected to reduce excessive day-level fragmentation while preserving the underlying intermittency structure relevant to spare-parts demand forecasting. To establish a consistent temporal representation across all dealer-part combinations, the monthly demand series were zero-padded over the complete aggregation horizon. This preprocessing step produced a final forecasting dataset containing approximately 1.4 million monthly observations. All forecasting experiments, SHOS feature-generation procedures, rolling-window validation analyses, significance testing, and model comparisons reported in this study were conducted exclusively on this final processed monthly-aggregated dataset. The zero-padding operation is statistically meaningful in the context of intermittent-demand forecasting because zero-valued observations explicitly represent periods of demand non-occurrence rather than missing information. These inactive periods contain essential occurrence-level information required for modelling renewal behaviour, demand sparsity, and event reactivation dynamics. Treating such periods as missing values would distort occurrence probabilities, alter sparsity distributions, and weaken the statistical representation of intermittent demand processes. The present study interprets zero-valued periods as meaningful non-occurrence events rather than missing observations. This distinction is important because zero-demand intervals carry essential information regarding sparsity structure, renewal behaviour, and demand reactivation dynamics. Consequently, zero-padding was retained intentionally to preserve occurrence-level information required for SHOS-based feature generation and comparative sparse-demand learning24. The processed dataset preserves the extreme sparsity characteristics of the original aftermarket demand environment. Across the active dealer-part series, the median Average Demand Interval (ADI) was approximately 25, while the median squared coefficient of variation (CV2) remained approximately 24, indicating highly intermittent and burst-driven demand behaviour. Figure 5 presents the empirical distribution of ADI and CV2 across all evaluated dealer-part series. The forecasting target variable is the total monthly quantity sold for each dealer-part combination (PartQty). Additional covariates include static part-level attributes such as part price, weight, and volumetric cube dimensions, which were incorporated as auxiliary explanatory variables during feature generation and supervised learning. Empirical distribution of intermittency characteristics (ADI and CV²) across 56,000 active dealer-part time series. This work additionally acknowledges that temporal aggregation may influence the observed renewal structure of intermittent demand. Finer-grained aggregation frequencies, such as weekly or daily representations, may preserve additional short-term event dynamics and inter-arrival variability. However, under the present large-scale dealer-part forecasting environment, such finer temporal resolutions substantially increased zero inflation, sparsity fragmentation, and noise sensitivity during preliminary exploratory analysis. The monthly aggregation adopted in this study, therefore, reflects an operationally motivated representation scale appropriate for the evaluated automotive aftermarket setting rather than a universally optimal temporal discretization strategy. Because the dataset originates from a single national automotive aftermarket ecosystem, the reported findings should be interpreted within the context of the evaluated market structure, dealer behaviour, and sparsity characteristics. Extension of the SHOS framework to alternative industrial sectors, geographical markets, or multi-echelon supply-chain environments remains an important direction for future validation. Monthly aggregation was selected as a practical compromise between temporal fidelity and statistical stability under highly sparse dealer-part demand conditions. At the raw daily transactional level, a substantial proportion of dealer-part series contained isolated low-volume events separated by extended inactive periods, resulting in highly fragmented temporal sequences with weak statistical learning signals. Aggregation to monthly frequency reduced excessive day-level fragmentation while preserving the broader intermittency structure required for representation-oriented forecasting and comparative model evaluation. The present study interprets zero-valued periods as meaningful non-occurrence events rather than missing observations. This distinction is important because zero-demand intervals carry essential information regarding sparsity structure, renewal behaviour, and demand reactivation dynamics. No artificial imputation procedure was used to replace missing demand observations. Zero-padding was applied exclusively to represent valid inactive dealer-part periods within the aggregation horizon. Preprocessing and feature engineering A structured feature engineering pipeline transforms the padded monthly data into a high-dimensional feature set. The pipeline includes: Temporal features: Lagged values of sales quantity up to 12 months, rolling sums and means (3- and 6-month windows), count of non-zero periods in the last 6 months, and time elapsed since the last sale. Temporal features: Lagged values of sales quantity up to 12 months, rolling sums and means (3- and 6-month windows), count of non-zero periods in the last 6 months, and time elapsed since the last sale. Algorithmic features: Forecasts from Croston’s method4 and the three core outputs of the SHOS algorithm: smoothed occurrence probability () smoothed conditional size (), and the SHOS point forecast. Crucially, SHOS features are computed using an adaptive, sparsity-aware smoothing mechanism designed to preserve demand events while stabilizing occurrence and size estimates with sparsity scaling factor k = 0.1, and an improved initialization that uses the first three months of data or falls back to the global mean of positive demands. Algorithmic features: Forecasts from Croston’s method4 and the three core outputs of the SHOS algorithm: smoothed occurrence probability () smoothed conditional size (), and the SHOS point forecast. Crucially, SHOS features are computed using an adaptive, sparsity-aware smoothing mechanism designed to preserve demand events while stabilizing occurrence and size estimates with sparsity scaling factor k = 0.1, and an improved initialization that uses the first three months of data or falls back to the global mean of positive demands. Global and latent features: Dealer- and part-level historical averages together with low-rank latent embeddings derived from the dealer-part activity matrix were incorporated to capture cross-sectional interaction structure within the sparse-demand environment. The latent embeddings were intended to summarize broad dealer purchasing behaviour, part-level co-occurrence structure, and recurring interaction patterns across dealer-part combinations rather than to function as direct probabilistic demand estimators. To generate these representations, Truncated SVD with 12 latent dimensions was used due to its computational efficiency and numerical stability with large sparse matrices. Global and latent features: Dealer- and part-level historical averages together with low-rank latent embeddings derived from the dealer-part activity matrix were incorporated to capture cross-sectional interaction structure within the sparse-demand environment. The latent embeddings were intended to summarize broad dealer purchasing behaviour, part-level co-occurrence structure, and recurring interaction patterns across dealer-part combinations rather than to function as direct probabilistic demand estimators. To generate these representations, Truncated SVD with 12 latent dimensions was used due to its computational efficiency and numerical stability with large sparse matrices. To strictly preserve temporal integrity and prevent data leakage, the Truncated SVD factorization was recomputed independently within each rolling-window cross-validation fold. Specifically, the dealer-part activity matrix was constructed exclusively from the training subset corresponding to the current fold, and the learned low-rank projection was subsequently applied to the associated test observations. No future demand information from the test horizon was used during latent embedding construction. The present study additionally acknowledges that alternative sparse-factorization approaches such as zero-inflated probabilistic matrix factorization, Poisson factorization, or nonnegative matrix factorization (NMF) may provide stronger probabilistic alignment with highly sparse count-based intermittent-demand data. However, the objective of the current work was to evaluate the contribution of SHOS-derived event-preserving statistical representations under a computationally scalable and controlled forecasting framework rather than to optimize latent factorization methodology itself. To isolate the effect of feature representation, baseline ML models are trained on base features only (temporal, global, and latent), while SHOS features are reserved for ablation studies. Forecasting models and comparative architectures To evaluate the effectiveness of SHOS-derived feature representations under intermittent demand, a diverse set of statistical, machine-learning, hybrid, and deep-learning forecasting configurations was examined under a unified experimental framework. All models were trained and evaluated using identical preprocessing procedures, rolling-window cross-validation folds, feature-generation protocols, and evaluation metrics to ensure fair and controlled comparison across architectures. The forecasting target for all supervised learning configurations was the monthly dealer-part demand quantity (PartQty). Unless otherwise stated, all models operated on the same zero-padded monthly-aggregated dataset described in section “Dataset description”. Classical intermittent-demand methods were incorporated into the study under three distinct roles. First, Croston and TSB methods were evaluated independently as standalone statistical forecasting baselines. Second, Croston- and SBA-derived signals were incorporated within selected machine-learning hybrids as feature-level integration benchmarks. Third, SHOS-derived occurrence and conditional-size representations were generated independently through adaptive smoothing procedures and evaluated separately within the supervised learning framework. This separation allows the contribution of SHOS-derived representations to be distinguished from the influence of externally generated Croston-type statistical features. The machine-learning (ML) models evaluated in this study include: LightGBM, XGBoost, Random Forest, Ridge Regression, and ElasticNet. In addition, deep-learning (DL) architectures were evaluated to examine the impact of sequential learning complexity under highly sparse demand conditions. These configurations include: Long Short-Term Memory (LSTM) networks, Temporal Convolutional Networks (TCN), and Transformer-based sequence models. For each architecture, two forecasting configurations were evaluated: Base model: trained using temporal lag variables, intermittency indicators, latent embeddings, and static dealer-part attributes. Base model: trained using temporal lag variables, intermittency indicators, latent embeddings, and static dealer-part attributes. Model + SHOS: trained using the identical feature set augmented with SHOS-derived occurrence probability and conditional demand size representations. Model + SHOS: trained using the identical feature set augmented with SHOS-derived occurrence probability and conditional demand size representations. The SHOS framework, therefore, operates as an independently generated representation-learning mechanism rather than as a direct extension or embedded reuse of Croston forecast outputs. This framework operates exclusively as a feature-generation mechanism and does not alter the prediction target, loss function, or forecasting objective. This design ensures that any observed performance improvement can be attributed specifically to representation-level augmentation rather than architectural modification or hybrid forecast blending25. To further evaluate whether explicit architectural decomposition provides additional benefit beyond feature-level decomposition, a two-stage hurdle forecasting framework was also implemented. The hurdle configuration separates demand occurrence and conditional demand magnitude into two sequential learning stages: Stage 1 (Occurrence model): A LightGBM classifier estimates the probability of observing non-zero demand, P(yt > 0), using the base feature set augmented with the SHOS occurrence representation . Stage 1 (Occurrence model): A LightGBM classifier estimates the probability of observing non-zero demand, P(yt > 0), using the base feature set augmented with the SHOS occurrence representation . Stage 2 (Conditional size model): A LightGBM regressor with a Tweedie objective predicts expected demand magnitude E(yt ∣ yt > 0) using the same feature set augmented with the SHOS conditional size representation (). Stage 2 (Conditional size model): A LightGBM regressor with a Tweedie objective predicts expected demand magnitude E(yt ∣ yt > 0) using the same feature set augmented with the SHOS conditional size representation (). The final forecast is obtained by multiplying the outputs of the occurrence and conditional-size stages, thereby generating an estimate of expected demand accounting for both event likelihood and demand magnitude. Gradient-boosted architectures (LightGBM and XGBoost) employed a Tweedie objective function to better model the zero-inflation and high skew of intermittent demand distributions. Classical intermittent-demand methods were incorporated into the study under three complementary roles. First, the TSB framework served as the primary conceptual basis for SHOS because both approaches decompose intermittent demand into components of occurrence probability and conditional demand size. Second, standalone Croston and TSB implementations were evaluated directly as statistical forecasting baselines within the same rolling-window evaluation framework used for all supervised learning models. Third, hybrid configurations such as SBA-LightGBM, Croston-LightGBM, and related variants were evaluated as feature-level integration strategies within common machine-learning pipelines. These comparisons were intended to assess the relative effectiveness of statistical feature representations rather than to compare fundamentally different forecasting paradigms. All preprocessing steps, latent embedding construction procedures, feature-generation pipelines, train-test splits, and evaluation folds were kept identical across all forecasting architectures unless explicitly stated otherwise. This controlled design ensures that observed performance differences arise from representation quality and model structure rather than inconsistencies in preprocessing, optimization, or evaluation methodology. Hyperparameter configuration All forecasting models use fixed hyperparameters determined through a preliminary hyperparameter optimization stage conducted prior to final evaluation. Hyperparameter selection was performed using rolling time-series cross-validation on a dedicated validation subset to preserve temporal ordering and avoid look-ahead bias. For each forecasting architecture, a grid-based search was conducted over the principal learning parameters, including tree depth, learning rate, regularization strength, number of estimators, and sampling ratios where applicable. Mean Absolute Error (MAE) on the validation folds was used as the primary optimization criterion because of its robustness under highly intermittent and skewed demand distributions. After identifying the best-performing configuration for each architecture, the selected hyperparameters were fixed and consistently applied across all rolling-window evaluation folds. This design ensures fair comparison between forecasting models and isolates the effect of feature representation from optimization variability. For the SHOS framework, the sparsity-scaling parameter was fixed at (k = 0.1) based on preliminary sensitivity analysis across representative demand categories. The recency-aware update mechanism used a recency window of three consecutive inactive periods prior to adaptive boosting, while smoothing coefficients were capped at 0.9 to maintain numerical stability under extreme sparsity26. Deep-learning architectures were additionally stabilized using fixed training configurations including: Adam optimizer, early stopping, dropout regularization, and fixed sequence-window construction across all folds. The final hyperparameter settings used throughout the experimental study are summarized in Table 2. Table 2Hyperparameter settings for evaluated forecasting models.ModelHyperparameterValueRidge regressionAlpha0.01XGBoostN estimators500Max depth9Learning rate0.05subsample0.8LightGBMN estimators500Max depth9Learning rate0.05Num leaves31Hurdle modelBase regressorLightGBMSHOSAlpha base0.2Beta base0.2Sparsity scaling (k)0.1Recency window3 periodsSmoothing cap0.9Initialization windowFirst 3 periods Hyperparameter settings for evaluated forecasting models. Deep-learning architectures were trained using a fixed sequence-window construction, with the same rolling-window validation framework applied to all machine-learning models. Model optimization employed the Adam optimizer with learning-rate stabilization, mini-batch training, dropout regularization, and early stopping to reduce overfitting under highly sparse demand conditions. Hyperparameters including hidden dimensions, dropout rates, learning rates, batch size, and epoch limits were selected during preliminary rolling-validation tuning and subsequently fixed across all evaluation folds to ensure fair comparison between architectures. All forecasting models were trained on identical preprocessing pipelines, rolling-window folds, and evaluation windows unless explicitly stated otherwise. This controlled experimental design ensures that reported performance differences arise primarily from feature representation and model structure rather than inconsistencies in preprocessing or optimization procedures. Deep-learning configuration summary The evaluated deep-learning architectures were implemented under a standardized training and evaluation framework to ensure controlled comparison across all forecasting configurations. All models were trained using identical preprocessing pipelines, rolling-window validation folds, feature-generation procedures, and train-test splits in order to isolate the effect of representation quality from differences in optimization or data preparation. The LSTM architecture employed a single recurrent hidden layer with dropout regularization and fixed sequence-window input construction to model temporal demand dependencies under sparse-demand conditions. The TCN configuration utilized causal temporal convolution blocks with residual connections and fixed dilation scheduling to capture localized temporal structure while preserving sequential ordering. The Transformer architecture employed a multi-head self-attention encoder with positional encoding and feed-forward projection layers to evaluate attention-based sequence learning behaviour in highly intermittent-demand environments. All deep-learning models were optimized using the Adam optimizer together with mini-batch gradient updates, learning-rate stabilization, dropout regularization, and early stopping to improve optimization stability and reduce overfitting. Hyperparameters including hidden dimensions, dropout rates, learning rates, batch size, sequence-window length, and epoch limits were selected during preliminary rolling-validation tuning and subsequently fixed across all evaluation folds to ensure reproducibility and fair comparison between architectures27. Despite these stabilization strategies, the evaluated deep-learning configurations remained sensitive to the extreme sparsity and prolonged inactive demand periods characteristic of the present automotive aftermarket forecasting environment. Consequently, the deep-learning results reported in this study should be interpreted within the context of the evaluated dataset characteristics, sparsity regime, feature representations, and rolling-window forecasting framework. Evaluation protocol and metrics Model performance is evaluated using a rolling-window cross-validation protocol that preserves temporal ordering and prevents look-ahead bias. In each fold, models are trained on a contiguous historical window and evaluated on the immediately following test window, thereby closely simulating real-world deployment conditions. All data-dependent transformations, including feature scaling and the construction of low-rank latent embeddings using truncated singular value decomposition (SVD), are fitted exclusively on the training data of each fold and then applied to the corresponding test data. This ensures strict temporal integrity and avoids information leakage. Model performance is reported as the mean and standard deviation of evaluation metrics computed across all eleven rolling validation folds. Let denote the observed demand and the corresponding forecast for observation i, with N representing the total number of observations in the evaluation set and denoting the mean of the observed demand. Forecast accuracy is assessed using multiple complementary metrics that capture both absolute and relative error characteristics relevant to intermittent demand forecasting. The Root Mean Squared Error (RMSE) is used to quantify sensitivity to large forecast errors and is particularly important for avoiding severe underprediction that can lead to stockouts28. It is defined as The Mean Absolute Error (MAE) measures the average magnitude of forecast errors and provides an intuitive measure of typical deviation between predicted and observed demand. It is computed as The coefficient of determination, , evaluates the proportion of variance in the observed demand explained by the model and serves as a normalized measure of explanatory power. It is given by Although R2 provides a normalized measure of variance explanation, its interpretation under highly intermittent and zero-inflated demand distributions requires caution. In sparse-demand environments dominated by extended inactive periods, variance-based metrics may become sensitive to isolated nonzero observations and may not fully reflect event-detection behaviour or demand-reactivation responsiveness. Accordingly, R2 is interpreted in this study only as a complementary diagnostic metric alongside absolute and weighted forecasting error measures. To assess relative forecast accuracy in a volume-weighted manner, the Weighted Mean Absolute Percentage Error (WMAPE) is employed. Unlike conventional percentage errors, WMAPE remains well-defined for intermittent demand series with frequent zero values. It is defined as In addition to aggregate accuracy metrics, overfitting is diagnosed by comparing training and test errors across validation folds to assess generalization behavior. Model interpretability and feature-level contributions are further examined using SHAP (SHapley Additive exPlanations) analysis, providing insight into the drivers of predictive performance. In addition to reporting aggregate error metrics, statistical significance testing is performed to assess whether observed performance differences between competing models are meaningful. Given the paired structure of rolling-window cross-validation and the non-normal error distributions typical of intermittent demand, the Wilcoxon signed-rank test is employed to compare per-fold MAE and WMAPE values between models. All tests are conducted at a 5% significance level. Results and discussion The results presented in this section follow directly from the experimental protocol described in section “SHOS framework and theoretical formulation”. All models are trained and evaluated on the same zero-padded, monthly-aggregated automotive aftermarket dataset using an identical rolling-window cross-validation framework. The dataset comprises approximately 1.4 million monthly observations derived from 56,000 dealer-part-time series. By holding data selection, preprocessing, feature construction, and evaluation protocols constant across all experiments, observed performance differences can be attributed solely to differences in feature representation and learning models. Performance of machine learning models This subsection presents a global comparison of all evaluated forecasting models under a unified experimental protocol. All models were trained and evaluated on the same zero-padded, monthly-aggregated automotive aftermarket dataset using identical rolling-window cross-validation splits of 11 folds, preprocessing steps, and evaluation metrics. As a result, differences in forecasting performance can be attributed solely to differences in modeling approach rather than experimental artifacts. Table 3 reports the mean performance across all validation folds for statistical baselines, standard machine learning models, SHOS augmented machine learning models, deep learning architectures, and hybrid hurdle formulations. Performance is evaluated using absolute error metrics (MAE, RMSE), relative error metrics (WMAPE, MAPE), and goodness-of-fit (R2)29. Accordingly, aggregate forecasting metrics such as MAE, RMSE, WMAPE, and R2 should be interpreted jointly under intermittent-demand conditions, because sparse-demand forecasting effectiveness depends on both magnitude accuracy and stability under zero inflation, as well as responsiveness to irregular demand-occurrence behaviour. Table 3Comparative performance of machine learning models (Mean ± Std).ModelgroupModelarchitectureMAERMSEMAPE (%)WMAPE (%)R²IntegratedSHOS (ML)SHOS + LightGBM0.0712 ± 0.00281.0310 ± 0.141617.73 ± 0.4720.39 ± 0.570.8430 ± 0.0460SHOS + Random Forest0.0702 ± 0.00320.8748 ± 0.090317.63 ± 0.80420.11 ± 0.9050.8879 ± 0.0245SHOS + XGBoost0.0793 ± 0.02401.2222 ± 0.137818.54 ± 0.66022.74 ± 0.9920.7799 ± 0.0545Standard MLLightGBM0.1326 ± 0.00461.2900 ± 0.107938.25 ± 1.2038.00 ± 0.680.7578 ± 0.0418RandomForest0.1364 ± 0.00461.3411 ± 0.105740.28 ± 1.5439.08 ± 1.030.7377 ± 0.0484XGBoost0.1413 ± 0.00451.3960 ± 0.107440.41 ± 1.3740.49 ± 1.260.7156 ± 0.0494RegressionRidge0.3439 ± 0.01341.5101 ± 0.116169.99 ± 0.6198.57 ± 4.280.6685 ± 0.0535ElasticNet0.3439 ± 0.01341.5101 ± 0.116070.00 ± 0.6198.59 ± 4.260.6685 ± 0.0535Hybrid(Two-stage)Hurdle_Baseline0.1341 ± 0.00501.2911 ± 0.093437.91 ± 1.2938.41 ± 0.710.7577 ± 0.0379Hurdle_SHOS0.0701 ± 0.00231.1136 ± 0.110216.84 ± 0.4420.10 ± 0.470.8186 ± 0.0413StatisticalTSB0.28311.290559.9482.20.7619Croston1.15583.96882.1335.63−1.2511DeeplearningLSTM0.6539 ± 0.0553.3033 ± 0.089103.06 ± 3.86187.14 ± 11.77-0.5775 ± 0.100TCN0.6534 ± 0.0253.4656 ± 0.207101.86 ± 2.92187.24 ± 6.49-0.7405 ± 0.211Transformer0.6801 ± 0.0503.3057 ± 0.064101.20 ± 7.17195.03 ± 15.81-0.5806 ± 0.105IntegratedSHOS (DL)SHOS + LSTM0.6370 ± 0.0253.4152 ± 0.151102.79 ± 1.47182.53 ± 6.15-0.6851 ± 0.121SHOS + TCN0.6681 ± 0.0443.5537 ± 0.142100.54 ± 5.12191.46 ± 12.61-0.8265 ± 0.146SHOS + Transformer0.6660 ± 0.0503.5936 ± 0.133106.34 ± 4.19190.93 ± 15.07-0.8679 ± 0.153NaiveNaivezero0.34922.6557-100−0.0176Significant values are in [italics].Note: MAPE is undefined for the Naive Zero model due to zero actual demand values, rendering percentage error metrics meaningless. Comparative performance of machine learning models (Mean ± Std). Integrated SHOS (ML) SHOS + LightGBM SHOS + Random Forest SHOS + XGBoost Random Forest Hybrid (Two-stage) Deep learning Integrated SHOS (DL) SHOS + LSTM SHOS + TCN SHOS + Transformer Naive zero Significant values are in [italics]. Note: MAPE is undefined for the Naive Zero model due to zero actual demand values, rendering percentage error metrics meaningless. Across all model groups, SHOS-augmented machine learning models consistently achieve the strongest performance. In particular, the SHOS + LightGBM model attains a mean MAE of 0.0712 ± 0.0028, an RMSE of 1.0310 ± 0.1416, and a WMAPE of 20.39% ± 0.57. The proposed model achieves a substantial reduction in MAE compared to the baseline (LightGBM without SHOS), computed as , which corresponds to an approximately 50% improvement. This improvement is achieved without increasing model complexity, indicating that the gains arise from the statistical structure embedded in the SHOS feature representation rather than from architectural sophistication. The two-stage Hurdle_SHOS formulation achieves comparable absolute accuracy (MAE = 0.0701 ± 0.0023), but does not consistently outperform the single-stage SHOS augmented models. Given its increased modelling complexity and training overhead, this result suggests that explicit architectural decomposition into occurrence and size is unnecessary when high-quality occurrence-size priors are already encoded at the feature level. Relative to the baseline LightGBM configuration (MAE = 0.1326), the SHOS-augmented LightGBM model (MAE = 0.0712) achieved an approximate 46.3% reduction in mean absolute error under the evaluated rolling-window validation framework. Similarly, the weighted mean absolute percentage error decreased from 38.00% to 20.39%, corresponding to an improvement of approximately 46.3% in weighted forecasting error. These results confirm that SHOS-derived occurrence-size representations substantially improve sparse-demand forecasting performance under the evaluated experimental conditions. Deep learning architectures, including LSTM, TCN, and Transformer models, perform markedly worse than both SHOS-augmented and standard machine learning models, yielding negative R² values and extremely high WMAPE. Even when augmented with SHOS features, deep learning architectures remained less effective under the evaluated configurations, highlighting the difficulty of training high-capacity sequence models under extreme sparsity and low signal-to-noise conditions. Finally, to verify that performance gains are not artefacts of zero inflation, a Naive Zero baseline is included. This model fails to capture variance (R2 = −0.0176) and exhibits the highest error across all metrics, confirming that SHOS-based models capture genuine demand signals rather than trivially exploiting sparsity. To assess the statistical significance of the performance difference between the single-stage SHOS model and the two-stage hurdle structure, a Wilcoxon signed rank test was applied to per-fold MAE and WMAPE values across the 11 rolling validation folds. The test results indicate statistically significant differences for both metrics (MAE: W=0.0, p=0.00098; WMAPE: W=0.0, p=0.00098) are presented in Table 4. This finding confirms that the single-stage SHOS model consistently outperforms the hurdle-based formulation under the evaluated experimental setting. Table 4Statistical significance testing between SHOS and Hurdle models.MetricTestTest statistic (W)p-valueSignificance(α = 0.05)MAEWilcoxon signed-rank0.00.00098SignificantWMAPEWilcoxon signed-rank0.00.00098Significant Statistical significance testing between SHOS and Hurdle models. Effect of SHOS feature representation This subsection isolates the impact of SHOS-derived feature representations on forecasting performance by comparing models trained on identical architectures with and without SHOS features. By holding the model class, hyperparameters, training protocol, and evaluation windows constant, this analysis serves as a controlled ablation study that attributes performance differences solely to the feature representation. Across all evaluated machine learning architectures, the introduction of SHOS features leads to substantial and consistent reductions in forecasting error. Tree-based ensemble models exhibit the largest absolute gains. For example, augmenting LightGBM with SHOS features reduces MAE by approximately 46% relative to its raw-feature counterpart, alongside comparable reductions in RMSE and WMAPE. Similar improvements are observed for XGBoost and Random Forest models, indicating that the benefits of SHOS are not model specific but arise from the statistical structure embedded in the feature representation. Importantly, these improvements are achieved without altering the prediction target or introducing architectural decomposition. SHOS outputs are used strictly as additional input features, enabling standard supervised learners to internalize demand occurrence and conditional size dynamics without modifying the learning objective30. This design ensures that observed performance gains reflect enhanced representation of the underlying demand process rather than changes in model formulation. The observed forecasting improvements associated with SHOS-augmented models cannot be attributed solely to indirect interaction effects with Croston-type statistical signals because SHOS-based configurations consistently outperform both: (i)raw-feature forecasting baselines, and raw-feature forecasting baselines, and (ii)Croston/SBA-integrated hybrid feature configurations evaluated under the same rolling-window validation framework. Croston/SBA-integrated hybrid feature configurations evaluated under the same rolling-window validation framework. The relative magnitude of performance improvement is particularly pronounced for metrics sensitive to sparsity and scale, such as WMAPE. This suggests that SHOS features improve not only point accuracy but also the stability of forecasts across heterogeneous demand volumes, a critical consideration for large-scale supply chain applications. Figure 6 illustrates the percentage reduction in MAE obtained by introducing SHOS features across representative machine learning models. The consistent downward shift across all architectures confirms that SHOS acts as a general-purpose representation enhancement rather than a model-specific tuning mechanism. Percentage reduction in MAE obtained by augmenting standard machine learning models with SHOS-derived features. Taken together, these results demonstrate that statistically grounded, event-preserving feature engineering can yield larger and more reliable gains than architectural modification alone. By embedding adaptive estimates of demand occurrence probability and conditional size directly into the input space, SHOS enables standard machine learning models to learn from highly intermittent data more effectively, without requiring complex multi-stage structures or specialized loss functions. Feature versus architecture decomposition While section “Preprocessing and feature engineering” established that SHOS-derived features substantially improve forecasting accuracy, this subsection examines whether these gains are attributable to feature representation or to model architectural complexity. To this end, we explicitly compare models across different architectural classes, linear models, tree-based ensembles, and deep learning architectures under both raw feature and SHOS augmented settings. The results reveal a clear and consistent pattern; improvements in forecasting accuracy align primarily with the quality of feature representation rather than with increasing architectural sophistication. Models trained on raw features show limited or negative returns from added complexity31. In particular, deep learning architectures such as LSTM, TCN, and Transformer exhibit significantly worse performance than simpler tree-based models when trained directly on sparse intermittent demand data, despite their higher expressive capacity. In contrast, introducing SHOS features leads to substantial performance gains across all architectural classes. Notably, SHOS-augmented tree-based models outperform both raw-feature deep learning models and SHOS-augmented deep learning models. For example, SHOS + LightGBM and SHOS + Random Forest achieve the lowest error across all evaluated metrics, while deep learning models, even when provided with SHOS features, remain substantially less accurate. This indicates that although SHOS improves learning across architectures, complex sequence models struggle to fully exploit sparse, event-driven signals relative to simpler learners. Table 5 summarizes this decomposition by reporting representative performance metrics grouped by architectural class and feature representation. The table highlights that switching from raw features to SHOS features yields a larger reduction in error than switching from a simple architecture to a more complex one. In several cases, a simple SHOS-augmented model outperforms a significantly more complex raw-feature model by a wide margin32. Table 5Decomposition of forecasting performance by feature representation and model architecture class (raw features vs. SHOS features).ModelgroupModelarchitectureMAERMSEMAPE (%)WMAPE (%) R 2 IntegratedSHOS (ML)SHOS + LightGBM0.0712 ± 0.00281.0310 ± 0.141617.73 ± 0.4720.39 ± 0.570.8430 ± 0.0460SHOS + Random Forest0.0702 ± 0.00320.8748 ± 0.090317.63 ± 0.80420.11 ± 0.9050.8879 ± 0.0245SHOS + XGBoost0.0793 ± 0.02401.2222 ± 0.137818.54 ± 0.66022.74 ± 0.9920.7799 ± 0.0545StandardMLLightGBM0.1326 ± 0.00461.2900 ± 0.107938.25 ± 1.2038.00 ± 0.680.7578 ± 0.0418RandomForest0.1364 ± 0.00461.3411 ± 0.105740.28 ± 1.5439.08 ± 1.030.7377 ± 0.0484XGBoost0.1413 ± 0.00451.3960 ± 0.107440.41 ± 1.3740.49 ± 1.260.7156 ± 0.0494Significant values are in [italics]. Decomposition of forecasting performance by feature representation and model architecture class (raw features vs. SHOS features). Integrated SHOS (ML) SHOS + LightGBM SHOS + Random Forest SHOS + XGBoost Standard ML Random Forest Significant values are in [italics]. These findings demonstrate that, under extreme demand intermittency, architectural complexity alone is insufficient to compensate for weak or noisy representations. Instead, performance gains arise from embedding statistically grounded structure into the input space. By explicitly encoding demand occurrence and conditional size dynamics, SHOS enables even relatively simple models to outperform more complex architectures trained on raw demand data. This decomposition provides strong empirical support for the central thesis of this study in intermittent demand forecasting: feature representation exerts a substantial influence in the evaluated sparse-demand setting. Hybrid dominance analysis To avoid ambiguity between standalone forecasting baselines and feature-level hybridization strategies, the hybrid comparisons presented in this subsection were evaluated within a unified supervised-learning framework. Specifically, classical intermittent-demand estimators such as Croston and SBA were not treated as complete forecasting systems in these experiments, but rather as statistical signal generators embedded within identical downstream machine-learning pipelines. This design enabled controlled comparison of representation quality while holding the learning architecture constant. Hybrid forecasting methods are frequently proposed for intermittent demand under the assumption that combining statistical estimators with machine learning or deep learning architectures improves performance relative to purely data-driven models. To examine whether the observed gains of SHOS arise from hybridization in general or from the specific design of SHOS features, we benchmark SHOS-based models against established statistical machine learning hybrid formulations reported in the intermittent demand literature. These include Croston-based and Syntetos-Boylan Approximation (SBA) based hybrids combined with tree-based learners and deep learning models. Table 6 demonstrates that incorporating classical intermittent-demand estimators into machine learning models yields heterogeneous and often limited performance gains. SBA-based hybrids consistently outperform Croston-based hybrids, reflecting SBA’s debiasing advantage under intermittency. However, even the best SBA-based hybrids remain substantially inferior to SHOS-based models reported in section “Dataset description” across all error metrics. Croston-based hybrids exhibit particularly poor performance, with elevated absolute and percentage errors and reduced explanatory power. This degradation reflects the sensitivity of fixed-parameter smoothing approaches to extreme sparsity, where isolated demand events exert disproportionate influence on the estimator. Table 6Performance comparison of classical statistical feature-signal integrations within common machine-learning forecasting pipelines for intermittent-demand forecasting.ModelarchitectureMAERMSEMAPE (%)WMAPE (%) R 2 SBA-LightGBM0.0848 ± 0.0061.3685 ± 0.09129.09 ± 2.1125.92 ± 1.150.7686 ± 0.038SBA-XGBoost0.0891 ± 0.0091.3902 ± 0.10232.24 ± 2.3327.16 ± 1.250.7582 ± 0.048Croston-LightGBM0.1504 ± 0.0121.4725 ± 0.11544.46 ± 2.4443.66 ± 2.150.6900 ± 0.042Croston-XGBoost0.1581 ± 0.0141.4974 ± 0.12247.60 ± 2.6145.90 ± 2.310.6794 ± 0.048Croston-LSTM0.6319 ± 0.0453.1514 ± 0.285105.46 ± 5.50183.47 ± 15.4-0.4200 ± 0.105SBA-LSTM0.6979 ± 0.0513.9382 ± 0.312106.96 ± 6.12202.65 ± 18.2-1.2175 ± 0.146 Performance comparison of classical statistical feature-signal integrations within common machine-learning forecasting pipelines for intermittent-demand forecasting. Croston- LightGBM Croston- XGBoost Notably, extending these classical hybrids to deep learning architectures does not recover performance. Both Croston-LSTM and SBA-LSTM exhibit severe degradation, including negative R2 values and inflated error measures. These results indicate that simply injecting classical statistical estimates into complex architectures does not resolve the fundamental challenges of intermittent demand and may amplify instability when demand signals are sparse and irregular33. In contrast, SHOS-based hybrids consistently dominate all alternative statistical-machine learning combinations across MAE, RMSE, WMAPE, and R2. This dominance confirms that hybrid performance is not driven by the inclusion of any statistical signal, but by the quality, adaptivity, and event-preserving nature of the statistical representation. By embedding series-specific occurrence and size priors at the feature-representation level rather than stacking or averaging model outputs, SHOS provides a statistically coherent foundation for learning under extreme intermittency. Deep learning performance analysis Table 7 summarizes the forecasting performance of recurrent (LSTM), convolutional (TCN), and attention-based (Transformer) architectures evaluated with both raw input features and SHOS-derived feature augmentation. Across all evaluated configurations, deep learning models exhibited substantially higher forecasting error than SHOS-augmented tree-based machine learning models. All deep learning variants produced negative R2 values and elevated WMAPE, indicating limited ability to model highly sparse intermittent demand under the present experimental setting. Table 7Forecasting performance of deep learning models with and without SHOS feature augmentation.ModelMAERMSEMAPE (%)WMAPE (%) R 2 LSTM0.6539 ± 0.0553.3033 ± 0.089103.06 ± 3.86187.14 ± 11.77-0.5775 ± 0.100TCN0.6534 ± 0.0253.4656 ± 0.207101.86 ± 2.92187.24 ± 6.49-0.7405 ± 0.211Transformer0.6801 ± 0.0503.3057 ± 0.064101.20 ± 7.17195.03 ± 15.81-0.5806 ± 0.105SHOS + LSTM0.6370 ± 0.0253.4152 ± 0.151102.79 ± 1.47182.53 ± 6.15-0.6851 ± 0.121SHOS + TCN0.6681 ± 0.0443.5537 ± 0.142100.54 ± 5.12191.46 ± 12.61-0.8265 ± 0.146SHOS + Transformer0.6660 ± 0.0503.5936 ± 0.133106.34 ± 4.19190.93 ± 15.07-0.8679 ± 0.153 Forecasting performance of deep learning models with and without SHOS feature augmentation. SHOS + Transformer Negative R2 values observed for several deep-learning configurations primarily reflect the difficulty of stable variance explanation under highly sparse and zero-inflated demand environments rather than implying complete predictive failure. Under such conditions, variance-based metrics become highly sensitive to isolated nonzero demand events and should therefore be interpreted jointly with MAE, RMSE, and WMAPE. Introducing SHOS-derived features produced modest improvements in selected absolute error metrics, particularly for the LSTM configuration. However, these improvements remained substantially smaller than those observed for SHOS-augmented tree-based models and did not alter the overall model ranking. The results, therefore suggest that informative statistical representations alone are insufficient to fully overcome the optimization and representation challenges associated with extreme demand sparsity in deep sequential architectures. This behaviour is consistent with the structural characteristics of intermittent demand, in which long inactive periods and irregular event occurrences reduce the availability of stable temporal learning signals. In contrast, tree-based learners operate directly on event-level feature representations and are therefore better suited to exploit sparse occurrence and conditional-size information encoded in SHOS features34. Figure 7 further illustrates that forecasting performance clusters more strongly by feature representation quality than by architectural depth. SHOS-augmented tree-based models consistently occupy the lowest-error region, while deep learning models remain comparatively unstable despite feature augmentation. Comparison of forecasting error across deep learning and SHOS-augmented models. Overall, the results indicate that the evaluated deep-learning architectures, under the present dataset characteristics, sparsity regime, feature representations, and rolling-window forecasting framework, remained less effective than representation-oriented machine-learning approaches for highly intermittent demand forecasting. These findings should therefore be interpreted as experimental observations rather than universal conclusions regarding the suitability of deep learning for intermittent-demand forecasting. Volatility retention and event-preserving behaviour To assess whether SHOS provides an event-preserving representation rather than conventional low-pass smoothing, we evaluate its behaviour using time-domain, frequency-domain, and structural metrics that explicitly capture phase distortion, peak attenuation, and temporal dependence. These analyses are designed to determine whether SHOS suppresses informative volatility or preserves the timing and structure of demand events under intermittency. In intermittent-demand forecasting, preserving the temporal structure of demand events is operationally important because replenishment and inventory decisions depend not only on aggregate forecast accuracy, but also on timely recognition of demand reactivation and event intensity. Excessive smoothing or temporal phase distortion can delay detection of genuine demand resurgence following prolonged inactivity, potentially increasing service-level risk and stockout exposure. Accordingly, the analyses presented in this subsection were intended to evaluate whether SHOS preserves informative event structure while reducing stochastic sparsity-induced noise. Phase lag measures temporal displacement between smoothed and raw demand signals, peak preservation quantifies the fraction of original peak magnitude retained, and FFT similarity measures spectral alignment between signals in the frequency domain. Table 8 evaluates temporal and structural preservation properties across smoothing approaches. SMA exhibits a non-zero phase lag and substantial peak attenuation, consistent with low-pass filtering behaviour. In contrast, SHOS introduces zero phase lag for both cyclical and intermittent demand, indicating that the timing of demand events is fully preserved. Moreover, SHOS achieves higher peak preservation and greater similarity to the raw signal than SMA, demonstrating that noise reduction is achieved without suppressing informative demand bursts. Because SHOS determines the temporal representation, the SHOS+LGBM pipeline inherits these event-preserving properties without introducing additional temporal distortion35. While temporal metrics assess event alignment and peak fidelity, frequency domain analysis provides complementary insight into how much of the original signal’s spectral content is retained after smoothing. Table 8Temporal and structural signal preservation metrics for simple moving average (SMA), SHOS, and SHOS+LGBM across cyclical and intermittent demand.Demand categoryMetricSMA (Low-Pass)SHOS (Adaptive)SHOS+LGBM (Hybrid)CyclicalPhase Lag (Months)0.00500Peak Preservation0.33330.14080.1395FFT Similarity0.40290.11160.1093Amplitude0.18590.05590.0544IntermittentPhase Lag (Months)-0.20500Peak Preservation0.33330.24080.2408FFT Similarity0.40150.13040.1304Amplitude0.12710.08690.0869 Temporal and structural signal preservation metrics for simple moving average (SMA), SHOS, and SHOS+LGBM across cyclical and intermittent demand. FFT energy reflects the total spectral content of the signal and provides a coarse measure of volatility retention. Excessive energy loss indicates over smoothing, while uncontrolled energy retention reflects noise amplification. Frequency-domain analysis provides complementary insight into how smoothing methods affect the structural variability of intermittent demand signals across different temporal scales. In the present context, retaining meaningful spectral structure indicates that the representation preserves recurring event behaviour and burst-driven dynamics rather than excessively flattening the signal through averaging. This property is particularly relevant in intermittent-demand environments where isolated demand events may carry disproportionate operational importance. Table 9 reports aggregate frequency-domain energy and average amplitude for each smoothing method. SMA removes a substantial portion of spectral energy, particularly for cyclical demand, reflecting aggressive attenuation of higher-frequency components. In contrast, SHOS retains significantly more frequency domain energy while still reducing noise relative to raw demand. For intermittent demand, SHOS maintains a balanced reduction in amplitude without eliminating structurally meaningful variability. The SHOS+LGBM results closely match those of SHOS, indicating that the hybrid pipeline preserves spectral characteristics rather than reintroducing artificial smoothing. Table 9Frequency-domain energy and average amplitude of raw demand, SMA, SHOS, and SHOS+LGBM signals for cyclical and intermittent demand categories.DemandcategoryMetricRaw demandSMASHOSSHOS+LGBMCyclicalFFT energy(Sum)120.62540.15288.00728.1189Avg. amplitude0.35490.18590.05590.0544IntermittentFFT energy(Sum)66.2518.2521.537521.5375Avg. amplitude0.26060.12710.08690.0869 Frequency-domain energy and average amplitude of raw demand, SMA, SHOS, and SHOS+LGBM signals for cyclical and intermittent demand categories. FFT energy (Sum) FFT energy (Sum) Figure 8 illustrates the time domain behaviour for representative SKUs. SMA attenuates and delays demand peaks, introducing phase distortion and flattening event magnitude. In contrast, SHOS preserves both peak timing and magnitude while reducing stochastic noise. These properties remain unchanged when SHOS-derived features are used within the LightGBM model. For cyclical demand, SMA spreads and delays demand peaks, resulting in phase distortion and amplitude loss. SHOS preserves peak timing while reducing high-frequency noise, maintaining the underlying cyclical structure. For intermittent demand, SHOS remains responsive to isolated demand events while avoiding artificial oscillations during extended zero demand periods. Importantly, these properties are preserved when SHOS-derived features are used within a machine learning model. Time-domain comparison for representative cyclical (top row) and intermittent (bottom row) SKUs. Autocorrelation analysis further confirms that SHOS does not behave as a low-pass filter. As shown in Fig. 9, SHOS closely preserves the temporal dependence structure of the raw demand signal, particularly for cyclical SKUs where seasonal correlations remain intact. In contrast, conventional smoothing disrupts temporal dependence by averaging across event boundaries. The SHOS+LightGBM pipeline maintains the same autocorrelation profile, indicating that integrating SHOS features does not introduce artificial temporal structure36. Autocorrelation functions for representative cyclical and intermittent SKUs. The signal-preservation metrics presented in this subsection should therefore be interpreted as diagnostic and mechanistic evidence explaining the behaviour of the SHOS representation under intermittency, rather than as direct proof of downstream inventory or replenishment optimization performance. While preservation of event timing and structural demand characteristics may support improved operational responsiveness, the primary validation of SHOS effectiveness in this study remains the forecasting accuracy, robustness, and sparsity-response analyses presented in the preceding sections. Taken together, these results demonstrate that SHOS fundamentally differs from conventional low-pass smoothing. Rather than averaging demand values over time, SHOS performs event-preserving smoothing by operating on latent occurrence and conditional size components. This design enables effective noise reduction under sparsity while preserving cyclical structure, peak timing, and temporal dependence. These properties provide complementary evidence supporting why SHOS-derived features may improve learning stability and event responsiveness under intermittent demand without suppressing informative temporal structure. Although improved responsiveness and event preservation may contribute to better replenishment awareness under intermittent demand, the analyses presented in this study were not intended to directly quantify inventory-policy performance or operational cost reduction. Rather, these metrics were evaluated as forecasting-oriented indicators of representation quality and sparse-demand learning stability. Extreme sparsity stress test: responsiveness and recovery under demand reactivation Intermittent demand forecasting models are particularly vulnerable under extreme sparsity, where long sequences of zero demand are followed by isolated demand reactivation events. In such regimes, learning systems trained on raw demand inputs often exhibit delayed response and unstable adjustment when demand resumes. To evaluate whether SHOS-derived feature representations mitigate these effects, we conducted an extreme sparsity stress test focusing on two operationally relevant criteria responsiveness, defined as the delay in reacting to demand reappearance, and recovery, defined as the time required for forecasts to stabilize following a demand event37. Responsiveness is measured as the number of periods required for the forecast to exhibit a statistically meaningful increase following the first non-zero demand observation, while recovery measures the number of periods required for forecast error to return within a stable tolerance band after the reactivation event. Table 10 reports stress test results across three sparsity bins. In the most extreme sparsity regime (sparsity ≥ 0.9), when augmented with SHOS-derived features, LightGBM exhibits a dramatic reduction in both responsiveness and recovery relative to LightGBM trained on raw demand features. Responsiveness is reduced by nearly an order of magnitude, indicating substantially faster reaction to demand reactivation, while recovery time is similarly compressed, reflecting more stable post-event adjustment. These improvements highlight SHOS’s ability to preserve actionable demand signals during prolonged periods of inactivity. Table 10Extreme sparsity stress test.Sparsity binModelResponsivenessRecovery0.9+LightGBM0.36960.4693LightGBM + SHOS0.0590.03230.8–0.9LightGBM0.56141.3321LightGBM + SHOS0.30960.75660.7–0.8LightGBM1.47752.6577LightGBM + SHOS1.15972.5779 Extreme sparsity stress test. In moderately sparse regimes (sparsity 0.8–0.9 and 0.7–0.8), LightGBM + SHOS continues to outperform the raw-feature baseline, though the magnitude of improvement diminishes as demand density increases. This behaviour is expected, as raw demand signals become more informative as sparsity decreases. Importantly, SHOS does not introduce lag or excessive smoothing in these regimes, indicating that its adaptive mechanism remains responsive across sparsity levels. Figure 10 provides a compact visual summary of the extreme sparsity stress test. Across all sparsity bins, LightGBM + SHOS consistently achieves lower responsiveness and recovery values than the raw-feature baseline, with the largest relative gains observed under extreme sparsity. The monotonic reduction in improvement as sparsity decreases further confirms that SHOS primarily enhances learning behaviour in regimes where raw demand signals are least informative, rather than artificially inflating performance in denser demand settings. Bar chart comparison of responsiveness and recovery metrics. Taken together, the extreme sparsity stress test demonstrates that SHOS-derived features substantially improve model behaviour precisely under conditions where intermittent demand forecasting is most challenging. By explicitly encoding demand occurrence probability and conditional demand size, SHOS stabilizes learning during prolonged zero-demand periods while enabling rapid adaptation when demand resumes38. These findings reinforce the central claim of this study: the primary benefits of SHOS arise at the feature-representation level, improving responsiveness and recovery without increasing architectural complexity. These findings complement the volatility analysis in section “ Evaluation protocol and metrics” by demonstrating that SHOS not only preserves events, but also enables faster recovery when events reoccur after prolonged inactivity. From an operational perspective, faster responsiveness directly reduces the risk of stockouts following demand reactivation, while faster recovery prevents prolonged overreaction after isolated demand events, both of which are critical in spare-parts inventory planning. SHAP interpretability analysis To further understand the mechanisms underlying the observed performance improvements, we analyze model interpretability using SHAP (SHapley Additive exPlanations). SHAP provides a consistent, game-theoretic framework for quantifying the marginal contribution of each input feature to the model’s predictions, enabling both global and local interpretability. This analysis focuses on the best-performing SHOS augmented machine learning model and examines which features drive forecast accuracy under extreme demand intermittency. Figure 11 presents the global SHAP feature importance summary, ranked by mean absolute SHAP value across all validation folds. The results reveal that SHOS-derived features dominate the importance ranking. In particular, the smoothed occurrence probability and the smoothed conditional demand size consistently emerge as the most influential predictors, surpassing traditional lag-based features, rolling statistics, and static product attributes. Global SHAP feature importance for the SHOS-augmented model, showing mean absolute contribution of each feature to forecast predictions across all validation folds. The prominence of indicates that the model places substantial weight on the estimated probability of demand occurrence, reflecting the critical role of event timing in intermittent demand forecasting. Similarly, the importance of confirms that conditional magnitude estimation provides essential information once demand occurrence is anticipated. Together, these features closely align with the theoretical decomposition of intermittent demand into occurrence and size components, thereby validating the conceptual foundations of the SHOS framework. In contrast, conventional temporal features such as raw lagged demand values and rolling averages exhibit lower and more dispersed SHAP contributions. This suggests that, in highly sparse settings, such features provide weaker, noisier signals than statistically smoothed occurrence-size representations. Static attributes and latent embeddings contribute additional explanatory power but play a secondary role relative to SHOS-derived signals. Importantly, the SHAP analysis also highlights the interpretability advantage of the proposed approach. Unlike deep learning architectures, which make it difficult to trace feature contributions, the SHOS-augmented model provides transparent, domain-consistent explanations. Decision-makers can directly relate forecast outputs to intuitive quantities such as recent demand occurrence likelihood and expected conditional demand size, facilitating trust and adoption in operational settings39. Overall, the SHAP results confirm that the performance gains achieved by SHOS-enhanced models are not driven by spurious correlations or overfitting, but by meaningful statistical representations that capture the fundamental structure of intermittent demand. This interpretability evidence complements the quantitative results presented in sections “Dataset description”–“Evaluation protocol and metrics” and reinforces the conclusion that SHOS provides both accuracy and explanatory value for large-scale intermittent demand forecasting. Managerial Insights The findings of this study have several direct implications for practitioners managing large-scale intermittent demand environments, such as automotive aftermarket, spare parts, and service logistics networks. Beyond forecasting accuracy, the proposed SHOS framework offers operational advantages in terms of robustness, interpretability, and deployability that are particularly relevant for real-world decision making. First, SHOS enables materially improved forecast reliability for highly intermittent items without increasing system complexity. The empirical results show that SHOS-enhanced single-stage machine learning models outperform both standard ML baselines and more complex multi-stage or deep learning architectures. For practitioners, this implies that meaningful accuracy gains can be achieved without introducing additional modelling layers, custom architectures, or specialized training pipelines40. This is especially important in enterprise forecasting systems, where model simplicity directly translates to lower maintenance cost, faster retraining cycles, and easier integration with existing planning workflows. Second, SHOS improves responsiveness to demand reactivation, reducing the risk of stockouts after prolonged inactivity. The extreme sparsity stress tests demonstrate that SHOS-based models recover more rapidly when demand reappears after long periods of zero demand. From an inventory management perspective, this behaviour directly addresses a common operational failure mode in intermittent-demand forecasting: delayed response to genuine demand signals. By stabilizing occurrence estimation while preserving event timing, SHOS supports earlier replenishment decisions when demand resumes, reducing service level failures without inflating inventory during inactive periods. Third, SHOS preserves demand volatility and event structure, supporting downstream planning decisions. Unlike conventional smoothing or low-pass filtering approaches that distort peaks and introduce phase lag, SHOS retains the timing and magnitude of demand events while suppressing noise induced by sparsity. This property is particularly valuable for planners who rely on forecasts not only for point estimates, but also for understanding demand variability, planning safety stock, and evaluating risk. Preserving event-driven dynamics ensures that forecasts remain operationally meaningful rather than overly smoothed artefacts. Fourth, the SHOS framework provides transparent and interpretable signals aligned with domain intuition. The SHAP analysis confirms that SHOS-derived features, with occurrence probability and conditional demand size as the dominant drivers, are the dominant drivers of model predictions. These quantities correspond to intuitive managerial concepts of the likelihood of demand and the size it is expected to be if it occurs. This interpretability facilitates trust, auditability, and stakeholder acceptance, particularly in settings where planners must justify decisions to procurement, finance, or service teams. Finally, SHOS is operationally scalable and parameter-light, supporting large-scale deployment. The k-sensitivity analysis demonstrates that SHOS performance varies smoothly across a wide range of sparsity conditions and does not require fine-grained parameter tuning across product categories or lifecycle stages. As a result, a single global parameter setting can be applied across tens of thousands of SKUs, enabling consistent and scalable deployment in large assortments without manual intervention or per-item calibration14. Taken together, these insights suggest that the primary managerial value of SHOS lies not only in improved forecast accuracy, but also in shifting the focus of intermittent-demand forecasting from architectural complexity to representation quality. By embedding a statistically grounded, event-preserving demand structure directly into the features, SHOS enables simpler, more robust, and more interpretable forecasting systems that closely align with the operational realities of intermittent demand planning. Practical implications for inventory management From an operational perspective, the performance characteristics of the SHOS-augmented model have direct implications for inventory planning and control in intermittent-demand environments. Forecast errors in such settings are inherently asymmetric: underprediction increases the risk of stockouts and service level failures, while overprediction leads to excess inventory, obsolescence, and unnecessary holding costs. The consistently low MAE and RMSE achieved by SHOS enhanced models indicate not only improved point accuracy, but also tighter error distributions, which are critical for managing this trade-off effectively. For inventory managers, this reliability translates into greater confidence in replenishment decisions, particularly for slow-moving items where even small forecast deviations can have disproportionate cost impacts. Equally important is the operational simplicity of the proposed framework. The use of a single-stage machine learning model augmented with SHOS-derived features reduces implementation complexity relative to multi-stage or hurdle-based pipelines. Such simplicity lowers barriers to deployment, facilitates routine retraining and monitoring, and improves transparency when forecasts must be communicated to planners and stakeholders. These considerations are especially relevant in large-scale supply chains that manage tens of thousands of SKUs amid frequent demand volatility. The robustness of the SHOS-enhanced model under extreme sparsity further supports its practical applicability. Overfitting diagnostics and stress test evaluations confirm stable generalization across validation windows, indicating reliable out-of-sample performance. This stability is essential for dynamic inventory systems that must adapt continuously to evolving demand patterns without frequent model recalibration. Taken together, these findings suggest that effective intermittent demand forecasting for inventory management does not require increasingly complex modelling architectures. Instead, embedding statistically grounded, event-aware representations of demand occurrence and magnitude into the forecasting pipeline enables simpler models to deliver more reliable and operationally useful predictions. This shift from architectural complexity to representation quality offers a practical pathway for improving inventory performance while maintaining system scalability and interpretability. Conclusions and managerial implications Intermittent demand forecasting remains one of the most operationally critical challenges in large-scale supply chains, particularly in aftermarket and spare-parts environments where thousands of slow-moving SKUs exhibit prolonged inactivity punctuated by irregular demand events. While recent research has largely focused on increasing architectural sophistication, the present study demonstrates that substantial and more robust gains can be achieved at the feature-representation level. This work introduced the SHOS framework as an event-preserving feature engineering approach that explicitly decomposes demand into an occurrence probability and a conditional size component, and embeds these statistically grounded priors directly into supervised learning pipelines. Extensive large-scale experiments on more than 56,000 intermittent time series show that SHOS-derived features consistently reduce forecasting error by approximately 46.3% relative to conventional feature sets, while also improving generalization stability, responsiveness, and recovery under extreme sparsity. Beyond predictive accuracy, the analysis demonstrates that SHOS preserves event timing, peak structure, and temporal dependence characteristics that are frequently distorted by conventional smoothing or deep sequence models. These properties explain why SHOS-augmented models outperform both raw-feature machine learning models and more complex two-stage or deep learning architectures. The findings indicate that, in highly intermittent regimes, representation quality dominates architectural complexity in determining forecasting performance. The findings of this study should be interpreted within the context of the evaluated automotive aftermarket forecasting environment, the selected preprocessing strategy, and the forecasting configurations considered in the present investigation. The dataset originates from a single national automotive aftermarket network, and therefore, the reported behaviour may reflect market-specific sparsity characteristics, dealer behaviour, and inventory dynamics. The present work focused primarily on point forecasting and event-preserving representation learning under highly intermittent demand conditions. Although the reported forecasting improvements, robustness behaviour, and signal-preservation characteristics are operationally relevant for intermittent-demand planning, the study did not explicitly evaluate probabilistic forecasting, uncertainty quantification, downstream inventory policies, service-level optimization, replenishment cost behaviour, or multi-echelon inventory performance. Consequently, the forecasting improvements reported here should not be interpreted as direct proof of inventory-level optimization performance. In addition, the reported forecasting behaviour depends on the evaluated feature-generation framework, temporal aggregation strategy, and preprocessing configuration adopted in the present study. While SHOS demonstrated strong performance under the evaluated sparse-demand conditions, further validation is required across alternative spare-parts ecosystems, retail forecasting environments, geographical markets, and heterogeneous supply-chain structures. Future work should therefore focus on extending SHOS toward probabilistic forecasting environments, integrating downstream inventory simulation and replenishment optimization, and evaluating representation-preserving sparse-demand forecasting across broader industrial and retail settings. Managerial and operational implications. From a practical perspective, the SHOS framework provides several actionable advantages for organizations managing large intermittent-demand portfolios: Plug-and-play deployment: SHOS operates as a feature-generation layer rather than a forecasting engine. It can be integrated directly into existing ERP, MRP, or demand-planning systems without restructuring current forecasting models or pipelines. Plug-and-play deployment: SHOS operates as a feature-generation layer rather than a forecasting engine. It can be integrated directly into existing ERP, MRP, or demand-planning systems without restructuring current forecasting models or pipelines. Cold-start and slow-moving SKU support: By explicitly modelling demand occurrence probability, SHOS provides meaningful signals even when historical demand is extremely sparse, enabling more informed planning for new, rare, and long-tail spare parts. Cold-start and slow-moving SKU support: By explicitly modelling demand occurrence probability, SHOS provides meaningful signals even when historical demand is extremely sparse, enabling more informed planning for new, rare, and long-tail spare parts. Improved stock-out prevention: The extreme-sparsity stress tests show that SHOS-augmented models react significantly faster when demand reappears after long inactivity periods. This responsiveness directly supports earlier replenishment decisions and reduces the risk of unexpected stock-outs. Improved stock-out prevention: The extreme-sparsity stress tests show that SHOS-augmented models react significantly faster when demand reappears after long inactivity periods. This responsiveness directly supports earlier replenishment decisions and reduces the risk of unexpected stock-outs. Stabilized safety-stock planning: By reducing noise amplification and preserving genuine volatility structure, SHOS improves forecast consistency across heterogeneous SKUs, supporting more reliable service-level and safety-stock calculations. Stabilized safety-stock planning: By reducing noise amplification and preserving genuine volatility structure, SHOS improves forecast consistency across heterogeneous SKUs, supporting more reliable service-level and safety-stock calculations. Scalable large-portfolio forecasting: Because SHOS embeds statistical structure into features rather than architectures, it remains computationally efficient and suitable for enterprise-scale deployment across tens of thousands of parts. Scalable large-portfolio forecasting: Because SHOS embeds statistical structure into features rather than architectures, it remains computationally efficient and suitable for enterprise-scale deployment across tens of thousands of parts. Scientific and strategic implications. More broadly, this study contributes to a growing body of evidence that learning performance in sparse and intermittent systems is governed primarily by the quality of representation rather than by model complexity. The success of SHOS suggests that statistically grounded, event-aware feature engineering can reduce reliance on complex hybrid architectures, enabling simpler, more interpretable, and more scalable forecasting systems. Limitations and future work. Future work will extend SHOS toward adaptive parameter learning, automated lifecycle-aware representations, and integration with prescriptive inventory-optimization frameworks. Further evaluation across multi-echelon supply chains, real-time deployment environments, and additional industry sectors will help establish broader generalizability and operational impact. The present study evaluates SHOS representations under a fixed monthly aggregation framework. Although monthly discretization improves statistical stability under highly sparse demand conditions, alternative temporal resolutions may preserve additional renewal dynamics and short-term demand structure. Future work should therefore investigate weekly aggregation, adaptive temporal discretization, and multi-resolution intermittent-demand representation frameworks to evaluate the sensitivity of SHOS-derived features to aggregation bucket size across different industrial forecasting environments. The present evaluation framework focuses primarily on aggregate forecasting accuracy and representation-oriented comparative analysis under sparse-demand conditions. Future work should extend the evaluation protocol to include event-oriented intermittent-demand metrics, such as the sensitivity and specificity of nonzero-demand detection, conditional evaluation restricted to positive-demand observations, and Average Detection Delay (ADD) to measure responsiveness to demand reactivation after prolonged inactivity periods. Future work should also investigate alternative representation-learning approaches for sparse dealer-part interaction matrices, including zero-inflated probabilistic matrix factorization, Poisson factorization, and nonnegative matrix factorization methods specifically designed for highly sparse count-based environments. Such approaches may further improve latent behavioural representation under extreme intermittent-demand conditions. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 2 Acknowledgements The authors gratefully acknowledge the support of the funding agencies and institutions listed above. Author contributions **S. N. B.** : Conceptualization (equal); Data curation (lead); Formal analysis (lead); Investigation (equal); Methodology (equal); Writing-original draft (lead); Writing-review & editing (equal). **V. S. R. B.** : Conceptualization (equal); Data curation (lead); Formal analysis (lead); Investigation (equal); Methodology (equal); Writing-original draft (lead); Writing & editing (equal). **C. C. S.** : Conceptualization (equal); Methodology (equal); Formal analysis (lead); Investigation (equal); Writing & editing (lead). **S. S.** : Formal analysis (lead); Investigation (equal); Methodology (equal); Writing-original draft (lead); Writing & editing (equal). **R. C.** : Formal analysis (lead); Investigation (equal); Methodology (equal); Writing-original draft (lead); Writing & editing (equal). Funding This work was co-funded by the European Union under the REFRESH - Research Excellence for Region Sustainability and High-tech Industries project (Project No. CZ.10.03.01/00/22_003/0000048) via the Operational Programme Just Transition. This article was also supported by the Students Grant Competition SP2024/087, Specific Research of Sustainable Manufacturing Technologies, financed by the Ministry of Education, Youth and Sports (MEYS), Czech Republic, and the Faculty of Mechanical Engineering, VŠB-Technical University of Ostrava. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. Data availability All the data and material used in this study is available in the manuscript, and further details if required, the corresponding author will provide the same, through proper requisition.Code AvailabilityThe custom code and computational scripts used for SHOS-based event-preserving feature engineering and intermittent demand forecasting analysis are publicly available at https://doi.org/10.5281/zenodo.20525971. Code availability The custom code and computational scripts used for SHOS-based event-preserving feature engineering and intermittent demand forecasting analysis are publicly available at https://doi.org/10.5281/zenodo.20525971. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Contributor Information Veera Siva Reddy B, Email: sivareddy.bobbili@gmail.com. C. Chandrasekhara Sastry, Email: chandrasekhar@iiitk.ac.in, Email: chandrasekhar041991@gmail.com. References Associated Data Data Availability Statement All the data and material used in this study is available in the manuscript, and further details if required, the corresponding author will provide the same, through proper requisition.Code AvailabilityThe custom code and computational scripts used for SHOS-based event-preserving feature engineering and intermittent demand forecasting analysis are publicly available at https://doi.org/10.5281/zenodo.20525971. The custom code and computational scripts used for SHOS-based event-preserving feature engineering and intermittent demand forecasting analysis are publicly available at https://doi.org/10.5281/zenodo.20525971.