Stable and time-aware explanations for supply-chain forecasts: A lean XAI framework

Machine-learning models now dominate retail demand forecasting, but their accuracy comes at a cost that managers feel every day: opacity. A gradient-boosted tree model can predict the next four weeks of grocery sales with remarkable precision, yet it offers no transparent account of which signals drove a forecast, when those signals mattered, or whether its reasoning is stable across time. A new study published in the open-access journal Heliyon by Ahmad N. Shawaqfeh and Juan F. De Paz tackles this trust gap head-on, proposing a lean explainable artificial intelligence (XAI) framework that produces explanations for supply-chain forecasts that are not only accurate but explicitly time-aware and statistically stable.

The motivation is rooted in one of forecasting’s most famous competitions. During the M5 competition, built on a massive Walmart sales dataset, LightGBM formed the backbone of the majority of the leading solutions, and machine-learning models dominated the leaderboard. Yet the competition’s own organizers documented a striking transparency deficit: despite targeted requests to the top 50 teams, detailed method descriptions were obtained for only 17 methods. Managers, the organizers noted, hesitate to deploy models whose internal reasoning they cannot follow. Subsequent research has reinforced the point, showing that higher forecast accuracy does not automatically translate into better inventory outcomes, and that firms treat demand forecasting as a high priority but struggle to operationalize AI unless outputs are explainable.

Standard explainability tools fall short for time series in a specific way. SHAP, the SHapley Additive exPlanations technique, assigns each feature a signed contribution to a model’s prediction, and it is widely preferred because it produces both local and global explanations. But a standard SHAP ranking only answers which features matter on average. It does not answer the critical business question of when they matter and how stable their importance is across time and forecast horizons. Demand is dynamic: the influence of a holiday or a price change spikes at specific moments and fades as the forecast extends. A single global score misses these horizon-dependent dynamics entirely, and prior work has shown that post-hoc explanations can fluctuate as a series progresses, making them unreliable for practitioners.

The new framework confronts this instability directly by combining SHAP with a time-based attribution layer called SHAPTime, which organizes explanations into a small set of horizon-aware temporal bins, and by validating explanations across multiple rolling-origin anchors rather than a single evaluation snapshot. The researchers trained LightGBM models on the FOODS category of the California (CA_1) subset of the M5 data, at the item-store level, to predict sales for the upcoming 28 days. To assess temporal robustness, they defined five rolling-origin anchors spaced 28 days apart, at day indices 1773, 1801, 1829, 1857, and 1885. Because each horizon faces different prediction challenges, they used a direct multi-horizon strategy: 28 separate LightGBM models per anchor, each specialized for one forecast day, yielding 140 horizon-specific fits in total.

Leakage prevention was engineered into the feature set with unusual rigor. The final specification contains 59 features organized into five families: categorical identifiers and events, demand recency via explicit lags from one to 28 days, shifted rolling demand statistics, price dynamics, and calendar and policy flags. All rolling statistics were computed on a temporally shifted version of the sales series, so that any feature value for a given day relies exclusively on sales data at least 28 days old, rendering the features horizon-agnostic and leakage-safe for all forecast horizons. Calendar variables, SNAP indicators, event flags, and prices were taken from the origin row rather than the target-day row, and lagged sales features were origin-aligned so that realized sales between the forecast origin and the target day were never used as predictors.

The model delivered anchor-stable accuracy. The Weighted Root Mean Squared Scaled Error, the official M5 metric, varied only modestly across the five anchors, ranging from 0.7041 to 0.7459 with an overall mean of 0.7284. This limited anchor-to-anchor variation indicates that the leakage-safe features generalize across time rather than overfitting to a particular origin. The authors emphasize that these scores are not intended to compete with top M5 leaderboard results, which rely on complex, uninterpretable hierarchy-wide ensembling; rather, the stable accuracy baseline provides a reliable predictive foundation for evaluating explanation consistency across forecast origins.

The global SHAP analysis revealed a clear hierarchy among the top 20 drivers. The ranking was led by the long-memory rolling feature mean28_180, followed by item identity, lag_1, and mean28_30, highlighting the combined importance of long-term demand history, item identity, and recent demand. Structural identity emerged as a major contributor, reflecting substantial cross-sectional heterogeneity that the model consistently uses. Calendar regularities such as day of week and weekend indicators sat mid-rank, while short-term persistence appeared as a decaying recency ladder of successive lags. Price dynamics and named events contributed relatively little in this specific FOODS/CA_1 slice. But, crucially, global importance answers what matters on average, not when it matters.

The time-aware decomposition answered the when question with three distinct patterns. First, origin-time context, the T0 bin capturing identity, calendar, SNAP, and current price, remained influential across horizons as a contextual backbone, with a mean attribution share of 0.265 and a strong positive association with horizon (Spearman’s rank correlation of +0.980). Second, recency decay: short lags were strong near the first forecast day and faded with horizon length, with every lag bin showing a strong negative trend, from lag_1 at -0.908 down to the 22-28 day bin at -0.789. Third, long-memory shift: the long-memory group carried the largest average share at 0.325, was among the most stable profiles with a coefficient of variation of 0.159, and increased in relative weight as near-term recency faded, with a Spearman correlation of +0.983. This redistribution empirically demonstrates the limitations of static global SHAP rankings.

Disaggregating the context bin clarified which components drive it. Item identity greatly exceeded all other origin-time components with a mean absolute SHAP of 0.301 while remaining remarkably stable across horizons, with a coefficient of variation of just 0.112, behaving like a persistent item-specific baseline that anchors forecasts from day 1 through day 28. The weekly rhythm came next: the day-of-week and weekend signal showed a mean of 0.038 with very strong weekly autocorrelation of 0.701 at lag 7 and a dominant period of roughly 3.5 days, precisely matching the end-of-week banding visible in the horizon profiles. Calendar features operating on longer spans made smaller, more even contributions, while episodic signals such as events and SNAP showed small average contributions and no significant horizon trend. Same-day price acted mainly as a near-term fine-tuner whose influence declined with horizon length.

The framework’s stability diagnostics formalize these patterns into compact, human-readable descriptors: mean strength, coefficient of variation, monotonic trend via Spearman correlation, periodicity via 7- and 14-day autocorrelations, and dominant periods extracted from Fast Fourier Transforms. Uncertainty was quantified with a block percentile bootstrap drawing 10,000 resamples of the five complete anchor-specific profiles, preserving within-profile temporal dependence. The authors also validated robustness with grouped permutation importance, a model-agnostic measure of forecast error increase when entire temporal groups are jointly permuted; the ranking agreed closely with SHAPTime’s mean-share ranking, with a Spearman correlation of 0.893 across seven identically defined groups, and both methods identified the long-memory group as the single most influential. Sensitivity analyses confirmed that the direction of attribution change persists across learned bin resolutions from 3 to 10 bins and across lead ranges of 14, 21, and 28 days.

The authors are careful about what generalizes. The numerical stability levels are specific to this retail context: grocery demand operates with comparatively stable regular prices, so the low variability of item identity and the strong weekly rhythm should not be expected at the same magnitudes in promotion-intensive sectors. Potential multicollinearity among lagged and rolling features also remains, which bin aggregation mitigates but does not fully resolve. What is intended to generalize is the procedure itself: train across multiple anchors, map SHAP attributions to temporal bins, and summarize each horizon profile with respect to strength, variability, trend, and periodicity. Because the pipeline relies on horizon-wise training, additive attributions, and memory-based temporal grouping, the same protocol can be applied to other forecasting datasets and horizons without redesigning the explanation logic.

For managers and analysts, the payoff is a compact view of persistent, horizon-dependent, and periodic model contributions that separates stable drivers from episodic ones. The horizon-specific profiles could support examination of order-up-to parameters, safety stocks, and the alignment of labor and replenishment with end-of-week peaks and calendar-driven demand. Firms that rely on more transparent models can react more quickly to sudden market changes than companies using black-box systems, and the framework’s diagnostics are designed for operational monitoring rather than one-off interpretive plots. In a field where explainability remains a critical gap between model accuracy and practical deployment, this work shows that trust in forecasting need not rest on faith: it can be measured, trended, and tracked across time, one horizon at a time.

Subject of Research: A multi-anchor, multi-horizon explainable AI framework that quantifies the temporal stability of SHAP-based feature attributions for LightGBM demand forecasting on retail supply-chain data.

Article Title: Stable and time-aware explanations for supply-chain forecasts: A lean XAI framework

Article References: Shawaqfeh, A. N., & De Paz, J. F. (2026). Stable and time-aware explanations for supply-chain forecasts: A lean XAI framework. Heliyon, 12(16), Article e45586. https://doi.org/10.1016/j.heliyon.2026.e45586

Image Credits: AI Generated

DOI: 10.1016/j.heliyon.2026.e45586

Keywords: explainable AI, XAI, supply chain forecasting, SHAP, SHAPTime, LightGBM, M5 competition, demand forecasting, temporal stability, rolling-origin validation, retail analytics, machine learning interpretability