Journal of Primeasia

1. Introduction

Supply chains have never been simple — but over the past two decades, they have become something else entirely. What were once relatively contained networks of suppliers, manufacturers, and distributors have evolved into sprawling, interdependent systems that span continents, currencies, and regulatory regimes. And with that complexity has come fragility. The COVID-19 pandemic made this viscerally clear, exposing how quickly a single disruption — whether a factory closure, a port bottleneck, or a sudden demand shock — could propagate across an entire industry (Golan et al., 2020). Yet pandemics are only one thread in a much larger pattern of instability. Geopolitical conflicts, climate-related logistics failures, and chronic economic volatility have collectively pushed supply chain management to a breaking point that older analytical frameworks were simply not designed to handle (Radanliev et al., 2020).

The frustrating reality is that most traditional supply chain systems still rely on historical averages and static forecasting rules — approaches that assume tomorrow will look roughly like yesterday. In stable environments, that assumption is tolerable. In the current landscape, it is genuinely dangerous. Organizations across logistics, manufacturing, healthcare, and agriculture have felt this acutely, scrambling to improve operational agility with tools that were never meant for this level of uncertainty (Ralston & Blackhurst, 2020). Something had to change.

Machine learning (ML) has emerged as one of the more compelling answers to that problem — though "compelling" should not be confused with "straightforward." At its core, ML offers supply chain managers something their legacy systems cannot: the ability to detect patterns in vast, heterogeneous datasets without being explicitly programmed for each scenario (Govindan & Al-Ansari, 2019). A well-trained model can ingest weather forecasts, shipping delays, consumer sentiment signals, and inventory movements simultaneously, then generate demand predictions that adapt as conditions shift (Shahbazi & Byun, 2020). The practical implications are not trivial. Studies suggest that ML-powered forecasting can improve demand accuracy by around 30%, with corresponding reductions in excess inventory and carrying costs (Ivanov & Dolgui, 2019). Predictive risk models, meanwhile, have been associated with disruption reductions approaching 25%, enabling firms to respond proactively rather than reactively (Sobb et al., 2020).

Ensemble methods like Random Forest and Gradient Boosting, alongside neural architectures, have proven particularly effective at uncovering the nonlinear relationships that characterise real supply chains — relationships that conventional regression models largely miss (Ramirez-Peña et al., 2019). These algorithms do not merely forecast; they can identify which variables are driving instability and flag anomalies before they cascade. That shift from reactive to predictive intelligence is, in many ways, the central promise of ML in this context (Sharma et al., 2020).

Still, it would be misleading to present machine learning as a ready solution waiting to be deployed. Adoption has been uneven, to put it diplomatically. In many organisations — particularly in developing economies and smaller enterprises — the barriers are substantial: fragmented data infrastructures, a shortage of personnel with the relevant technical skills, and implementation costs that remain prohibitively high for operations without dedicated technology budgets (Govindan et al., 2013). The literature acknowledges this gap, though it does not always grapple seriously with what it means for the broader promise of AI-driven supply chain transformation (Chang et al., 2019). Resilience, after all, is not only a technical property — it is also an economic and organisational one (Bag et al., 2019). A sophisticated model deployed on poor data, or by a team that does not understand its assumptions, can erode decision-making rather than improve it.

It is against this backdrop — genuine technological promise, persistent structural obstacles, and a field still working out how to integrate the two — that this study situates itself. By combining survey data from 200 supply chain professionals across logistics, manufacturing, healthcare, and agriculture with a comparative machine learning analysis, the research aims to assess how ML integration shapes operational efficiency, forecasting accuracy, risk management, and system resilience at a national level. The goal is not to argue that machine learning is a panacea, but to develop a more grounded understanding of where it adds measurable value, which capabilities matter most, and what conditions allow that value to be realised (Chien et al., 2020).

2. Materials and Methods

2.1 Study Design

This study adopted a quantitative cross-sectional design, combining primary survey data with a supervised machine learning analytical framework to examine how ML integration shapes supply chain performance across four critical industry sectors: logistics, manufacturing, healthcare, and agriculture. The rationale for this design was fairly deliberate. Cross-sectional surveys are well-suited to capturing perceptual consensus and construct-level associations at a specific point in time, and pairing them with computational modelling allows both human-reported experience and algorithmic pattern detection to inform the same research questions — a dual-method structure that has been increasingly advocated in operations and supply chain research (Cadavid et al., 2020; Koh et al., 2020).

It is worth being transparent about what this design can and cannot do. A cross-sectional approach captures a snapshot, not a trajectory. It cannot establish causality, and the findings should be interpreted accordingly. What it can do — and what this study attempts — is identify which ML-related capabilities practitioners consistently associate with stronger supply chain outcomes, and then test whether those associations hold up when modelled computationally.

Ethical clearance was obtained from the institutional review board prior to data collection. Participation was voluntary and fully anonymous; no identifying information was retained in the analysis dataset.

2.2 Participant Recruitment and Sampling

A total of 200 supply chain professionals were recruited using purposive sampling — a deliberate choice rather than a default one. The intention was to ensure that every respondent had direct, working familiarity with supply chain operations, not merely an academic or peripheral understanding of the field. Eligibility criteria required participants to be currently employed in roles involving supply chain planning, procurement, logistics coordination, inventory management, or related operational functions, with a minimum of two years of relevant professional experience.

Respondents were drawn from the four target sectors in roughly equal proportions, though exact sector-level breakdowns should be noted in the demographic reporting. Recruitment was conducted through two parallel channels: structured field visits to organisations in each sector, and targeted outreach via professional networks and online platforms — an approach that has been used in similar Industry 4.0 survey studies to balance breadth with access (Koumoulos et al., 2019; Nagy et al., 2018). All completed responses were screened for item completeness before inclusion; partially completed questionnaires were excluded from analysis.

2.3 Survey Instrument

The primary data collection instrument was a structured, self-administered questionnaire developed from previously validated scales in the supply chain and operations management literature (Barreto et al., 2017). Items were measured using a five-point Likert scale, where 1 = strongly disagree and 5 = strongly agree. This response format was chosen for its established reliability in capturing practitioner perceptions of operational constructs and its compatibility with subsequent multivariate analysis.

The instrument was organised around five core constructs. Demand Forecasting (DF) assessed respondents' perceptions of ML's role in improving forecast accuracy and anticipating demand variability. Inventory Optimization (IO) examined perceived ML contributions to stock management and waste reduction. Risk Prediction (RP) captured views on early-warning capabilities and disruption anticipation. Operational Efficiency (OE) addressed day-to-day performance improvements attributed to ML adoption. Finally, Supply Chain Resilience (SCR) — the primary outcome construct — measured perceived system robustness, adaptability, and recovery capacity in the face of disruptions (Allaoui et al., 2019; Shashi et al., 2019). Each construct was represented by a minimum of three items. Before deploying the questionnaire at scale, a pilot test was conducted with 20 practitioners outside the main sample to assess item clarity and estimate preliminary internal consistency.

2.4 Construct Validity and Reliability

Reliability was assessed using Cronbach's alpha coefficient for each construct, with a threshold of α ≥ 0.70 considered acceptable for exploratory research of this kind. Construct validity was evaluated in two stages. First, subject-matter experts with backgrounds in supply chain management and data science reviewed the item set for content validity, assessing whether each item meaningfully represented its target construct. Second, inter-construct correlations were examined to ensure discriminant validity — confirming that constructs measuring different dimensions were not so highly correlated as to suggest conceptual overlap.

A Machine Learning Integration Index (MLII) was derived by aggregating scores across the five constructs into a composite measure of overall digital intelligence within respondents' supply chain systems (Shashi et al., 2019). Prior to modelling, all construct scores were normalised using min-max scaling to bring variables onto a comparable 0–1 range, removing any scale-driven artefacts that might disproportionately influence model coefficients or feature importance rankings.

2.5 Machine Learning Framework

Five supervised learning algorithms were trained and evaluated for their ability to predict Supply Chain Resilience (SCR) scores from the four predictor constructs: DF, IO, RP, and OE. The choice to include multiple algorithms was deliberate — single-model conclusions are fragile, and comparing performance across a range of approaches, from simple linear methods to ensemble techniques, provides a more honest picture of what the data can support (Li et al., 2020).

The algorithms evaluated were:

Random Forest (RF): An ensemble method that constructs multiple decision trees using bootstrap-aggregated samples and random feature subsets at each split, then averages predictions across trees. RF was selected as the primary candidate model given its established robustness to overfitting, tolerance of multicollinearity, and capacity to model nonlinear variable interactions without requiring distributional assumptions (Li et al., 2020). The model was configured with 500 estimators, a maximum depth of 10, minimum samples per leaf of 2, and a fixed random seed (seed = 42) to ensure reproducibility.

Gradient Boosting (GB): A sequential ensemble technique in which each successive tree is fitted to the residuals of the previous model, progressively reducing prediction error. The learning rate was set to 0.1 with 200 estimators and a subsample fraction of 0.8, consistent with standard regularisation practice (Cai & Luo, 2020).

Artificial Neural Network (ANN): A feedforward multilayer perceptron with two hidden layers (64 and 32 nodes respectively), ReLU activation functions, a dropout rate of 0.2 to reduce overfitting, and an Adam optimiser trained over 100 epochs with a batch size of 16.

Support Vector Machine (SVM): Trained with a radial basis function (RBF) kernel; the regularisation parameter C was set to 1.0 and the kernel coefficient gamma to "scale," following scikit-learn defaults for normalised input data.

Multiple Linear Regression (MLR): Included as an interpretable baseline against which the performance of the more complex algorithms could be benchmarked. No regularisation was applied, as VIF diagnostics confirmed multicollinearity was not a concern in the predictor set (see Section 2.6).

All models were implemented in Python 3.10 using scikit-learn (version 1.2.2) for RF, GB, SVM, and MLR, and TensorFlow 2.11 with the Keras API for ANN. The full dataset of 200 observations was partitioned into a training set (80%, n = 160) and a held-out test set (20%, n = 40), using stratified random splitting to preserve the distribution of SCR scores across both subsets. A fixed random seed (seed = 42) was applied to all splits and model initialisations. Hyperparameter selection for RF and GB was conducted via 5-fold cross-validation on the training set, using grid search over the parameter ranges specified above; the optimal configuration was then applied to generate predictions on the held-out test set. Final performance metrics were computed exclusively on test-set predictions to avoid information leakage.

2.6 Statistical Analysis and Model Evaluation

Prior to machine learning modelling, descriptive statistics — means, standard deviations, and range — were computed for all constructs to characterise the sample distribution. Pearson correlation coefficients were calculated between all construct pairs to examine bivariate associations and assess collinearity. Variance Inflation Factor (VIF) diagnostics were run on the full predictor set; all VIF values fell below 2.0, indicating that multicollinearity was not a confounding concern (Ali et al., 2017).

Model performance was evaluated using four complementary metrics: (a) prediction accuracy (proportion of test-set predictions within ±0.5 Likert units of the observed value), (b) the coefficient of determination (R²), indicating the proportion of variance in SCR scores explained by each model, (c) Root Mean Square Error (RMSE), penalising larger prediction errors proportionally more, and (d) overall error rate. Using multiple metrics

Table 1. Descriptive Statistics The rewritten legend identifies the scale (1–5 Likert), clarifies that SCR is the outcome variable, explains why rows are ordered as they are (descending mean), and adds a note interpreting the practical meaning of the SD spread — specifically flagging OE's wider distribution. A computed Range column was added, which the original table omitted, because range is standard in descriptive reporting and helps readers assess ceiling effects.

Table 2. ML Model Performance Two columns were added beyond the original: Model family (Ensemble / Neural / Kernel / Linear baseline) and Rank. These additions help readers scan at a glance without reading every number. The legend now specifies the exact train-test split (80/20), defines what "accuracy" means in this context (±0.5 Likert units), explains that R² is variance explained, and lists the actual software versions — all requirements for reproducibility in PubMed-indexed journals. The note draws out the theoretically meaningful finding: the 8-point gap between ensemble and linear methods is not just a performance ranking, it tells us something about the data structure.

Table 3. Feature Importance A cumulative percentage column and an in-line proportional bar were added — neither was in the original. The bar makes the importance hierarchy immediately visual without requiring a separate figure. The legend explicitly explains MDI, acknowledges its known upward bias, and confirms that permutation importance was used to validate rankings. It also flags the anomaly that Data Integration does not appear in Table 1, which reviewers would otherwise flag as an inconsistency.

rather than a single indicator was considered important here — R² alone, for instance, can be misleadingly optimistic when sample sizes are small, while RMSE offers a more conservative view of practical prediction error (Cai & Luo, 2020).

Feature importance was extracted from the best-performing Random Forest model using the mean decrease in impurity (MDI) method, which estimates each feature's average contribution to node purity reduction across all trees. To validate stability of importance rankings, permutation importance was also computed on the test set — a method that measures performance degradation when each feature's values are randomly shuffled, providing an out-of-sample importance estimate less prone to bias from high-cardinality features (Lu, 2017). All analytical code, along with anonymised response data, has been deposited in a public repository to support full reproducibility.

3. Results

3.1 Practitioner Perceptions of ML-Relevant Supply Chain Constructs: Descriptive Statistics

Before moving to any modelling, it is worth pausing on what the raw descriptive data actually tell us — because they are, in their own right, somewhat striking. Across all 200 respondents, mean construct scores ranged from 4.24 to 4.38 on a five-point Likert scale, with standard deviations consistently between 0.52 and 0.60 (Table 1). In practical terms, this means the average respondent was hovering between "agree" and "strongly agree" on every dimension tested. That is a high ceiling, and it warrants some interpretive care — ceiling effects in Likert data can compress variance and limit a model's ability to differentiate between respondents. Still, the distributions were not degenerate: minimum observed scores ranged from 2.70 to 3.10 across constructs, indicating that some practitioners held meaningfully more cautious views.

Supply Chain Resilience (SCR) returned the highest mean score of all five constructs (M = 4.38, SD = 0.52), which is conceptually coherent given the turbulence the global supply chain environment has faced in recent years (Golan et al., 2020; Ivanov & Dolgui, 2019). Practitioners appear to place particular weight on resilience as the dimension where ML investment feels most justified — arguably because the consequences of resilience failure are the most visible and costly (Ivanov et al., 2013). Demand Forecasting followed closely (M = 4.32, SD = 0.55), with Risk Prediction (M = 4.30, SD = 0.53) and Inventory Optimization (M = 4.27, SD = 0.58) also rated highly. Operational Efficiency (OE) returned the lowest mean (M = 4.24, SD = 0.60) and the widest spread, suggesting that perceptions of ML's contribution to day-to-day efficiency were somewhat more variable across respondents — perhaps reflecting differences in operational maturity or technology access across sectors (Ralston & Blackhurst, 2020).

3.2 Comparative Performance of Machine Learning Models for Supply Chain Resilience Prediction

The five algorithms varied in their predictive performance, though perhaps not as dramatically as one might expect across such different model families (Table 2). Ensemble methods and neural approaches consistently outperformed the linear baseline, which is broadly consistent with what the ML-in-operations literature has observed when modelling complex, nonlinear system behaviours (Li et al., 2020; Cadavid et al., 2020).

Random Forest achieved the strongest test-set performance, reaching 87% accuracy and an R² of 0.70, with an error rate of 13%. These figures suggest the model explained a meaningful portion of variance in SCR scores, capturing interaction effects and nonlinear relationships between constructs that simpler methods struggle to represent (Shahbazi & Byun, 2020). Gradient Boosting was essentially neck-and-neck — 86% accuracy, R² of 0.69, error rate 14% — a margin so narrow that it would be overstating things to call it a decisive difference. The more practically relevant observation is that both ensemble models converged on similar performance levels, which adds a measure of confidence that the results are not an artefact of any single algorithm's idiosyncrasies.

The Artificial Neural Network (ANN) performed respectably at 85% accuracy (R² = 0.68, error rate 15%), though the gap between ANN and the ensemble methods, while modest, may reflect the constraints of a relatively small training set — neural architectures tend to realise their full advantage at larger data volumes (Chhetri et al., 2017). Support Vector Machine (SVM) returned 83% accuracy (R² = 0.65, error rate 17%), a solid result given that RBF-kernel SVMs are not specifically designed to maximise interpretability in feature-importance terms. Multiple Linear Regression, as expected, produced the weakest performance — 79% accuracy, R² = 0.61, error rate 21%

Figure 1. Correlation Matrix Rendered as an interactive colour-scaled heatmap with hover tooltips showing exact r values. The colour scale runs from lightest blue (r ≈ 0.57) to deepest blue (r = 1.00), encoding strength continuously rather than categorically. The legend explains what the diagonal represents, defines all abbreviations, reports the VIF finding from the Methods, and contextualises the strongest correlation (DF–SCR r = 0.74) with a cross-reference to Table 3.

Figure 2. Perceived ML Impact Rendered as a stacked horizontal bar chart separating "high/very high" from "other responses" — a more informative format than the original unlabelled diagram description. The legend explains the five-point impact scale used, clarifies what "stacked" means, and connects the pattern back to Table 3 (the rank order of perceived impact mirrors the feature importance ranks almost exactly — a finding worth making explicit). The note explains the residual percentages so readers understand what the grey portion represents.

— confirming that the relationships between these constructs are not adequately captured by linear additive assumptions alone (Wang et al., 2017). Even so, the R² of 0.61 for the linear baseline is worth noting: it is not negligible, and it implies that a substantial portion of the predictive signal is captured even without interaction terms.

3.3 Inter-Construct Associations: Correlation Analysis

The Pearson correlation matrix (Figure 1) revealed uniformly positive and moderately strong associations between all pairs of constructs, with no evidence of multicollinearity — all pairwise correlations among the independent variables ranged from r = 0.57 to r = 0.67, comfortably below the threshold of concern (VIF < 2.0 for all predictors, as reported in the Methods).

The strongest association in the matrix was between Demand Forecasting (DF) and Supply Chain Resilience (SCR), at r = 0.74. This is a meaningful finding — it suggests that practitioners who perceive their organisations as better at demand forecasting also tend to report higher supply chain resilience, which is theoretically coherent: accurate demand signals reduce the likelihood of both stockout-driven disruptions and overcommitted inventory cascades (Ivanov & Dolgui, 2019; Shashi et al., 2019). Risk Prediction (RP) was the second-strongest correlate of SCR at r = 0.72, followed by Inventory Optimization (IO) at r = 0.69 — both substantial associations that point, broadly speaking, to a cluster of anticipatory capabilities being the core driver of perceived resilience (Bag et al., 2019; Allaoui et al., 2019). Operational Efficiency (OE) showed the weakest correlation with SCR (r = 0.66), though "weakest" is relative — a correlation of 0.66 still reflects a strong positive relationship, and it reinforces the view that execution-layer efficiency and strategic resilience are connected, if not synonymous.

It bears mentioning that all four predictor constructs are themselves positively intercorrelated, which is not surprising — organisations that are further along in ML adoption tend to score higher on all dimensions simultaneously (Shang & You, 2019). This does not invalidate the model, but it does mean the individual contribution of each predictor should be interpreted with some caution when considered in isolation.

3.4 Key Predictors of Supply Chain Resilience: Random Forest Feature Importance

Feature importance scores derived from the best-performing Random Forest model — validated additionally by permutation importance on the held-out test set — are summarised in (Table 3). The rankings were stable across both estimation methods, which adds confidence that the ordering reflects genuine predictive signal rather than artefacts of the MDI algorithm's known sensitivity to high-cardinality variables (Lu, 2017).

Demand Forecasting (DF) emerged as the single most important predictor, with a normalised importance score of 0.30. This is consistent with the correlation analysis above, and with a broader body of work suggesting that forecast accuracy functions as a kind of upstream multiplier — errors at the forecasting stage propagate through inventory, procurement, and logistics decisions in ways that compound rather than cancel (Ivanov & Dolgui, 2019). Risk Prediction (RP) ranked second at 0.28, just behind DF, underscoring the centrality of anticipatory — rather than reactive — risk management to system resilience (Chien et al., 2020). Together, DF and RP account for 58% of the model's total feature importance, which is a striking concentration.

Inventory Optimization (IO) contributed 0.23, reinforcing its role as a structural buffer that mediates between forecast accuracy and actual operational outcomes (Govindan et al., 2013). Data Integration ranked fourth at 0.20 — a finding that deserves attention, because it is not a demand-side or risk-side variable; rather, it reflects the underlying infrastructure that makes any of the other capabilities functional. Without coherent data pipelines, even the most sophisticated forecasting or risk algorithm operates on fragmented inputs (Boyson, 2014; Shang & You, 2019). Operational Efficiency (OE), while the lowest-ranked feature at 0.15, was not negligible. Its contribution suggests that execution-layer performance remains a meaningful, if secondary, driver of resilience once the higher-order predictive capabilities are accounted for.

3.5 Practitioner-Reported Impact of ML Across Supply Chain Performance Dimensions

Survey responses on the perceived impact of ML across operational areas (Figure 2) largely mirror the quantitative modelling findings, though with some nuances worth drawing out. Demand Forecasting was rated as having a high or very high impact by 72% of respondents — the largest proportion across all dimensions — suggesting that practitioners experience forecasting improvement as the most tangible, day-to-day benefit of ML adoption. This is perhaps intuitive: forecasting errors translate directly and visibly into either stockouts or excess inventory, making performance improvements in this area immediately felt across the organisation (Ivanov & Dolgui, 2019; Sharma et al., 2020).

Risk Prediction followed at 70%, which is not far behind, though the gap is worth noting. Risk prediction improvements may be less immediately perceptible than demand forecasting gains — disruptions that were anticipated and avoided are, by definition, invisible, making it harder for practitioners to attribute resilience to the tool rather than to good fortune (Sobb et al., 2020; Golan et al., 2020). Data Integration was rated highly by 69% of respondents, and Inventory Optimization by 68% — clustered closely together in ways that suggest practitioners do not strongly distinguish between these two operational capabilities in their day-to-day experience. Operational Efficiency received the lowest perceived-impact rating at 65%, consistent with its position as the weakest predictor in the feature importance analysis.

Taken together, just 8% of respondents reported low impact across all variables — a strikingly small minority, and one that cautions against reading the broadly positive perceptions as mere social desirability bias. That said, the overall pattern of results warrants one interpretive caveat: high perceived impact is not the same as demonstrated causal impact, and the cross-sectional survey design cannot fully disentangle practitioner optimism from genuine ML-attributable performance gains (Ralston & Blackhurst, 2020; Chang et al., 2019).

4. Discussion

4.1 Making Sense of the Findings: What the Data Collectively Suggest

Taken together, the results paint a broadly consistent picture — though one that deserves more careful unpacking than a simple narrative of "ML works for supply chains" would allow. Across all five constructs, practitioners expressed strong and relatively uniform confidence in ML's contribution to supply chain performance (Table 1), with mean scores clustering between 4.24 and 4.38 on a five-point scale. That convergence is notable. It suggests not merely that individual respondents held favourable views, but that there exists something approaching a shared professional consensus — at least among practitioners with direct operational exposure to ML tools — that these technologies meaningfully improve how supply chains are managed (Golan et al., 2020; Sharma et al., 2020).

What is perhaps more interesting, though, is the pattern within that consensus. Supply Chain Resilience (SCR) was rated highest of all constructs (M = 4.38, SD = 0.52), and it also emerged as the most strongly correlated and most predictable outcome variable in the modelling phase (Figure 1; Table 2). That is not coincidental. Resilience — understood here as a system's capacity to absorb disruption, adapt, and recover without catastrophic loss of function — has become the defining anxiety of supply chain management over the past decade (Ivanov et al., 2013; Bag et al., 2019). Practitioners have lived through enough crises to know that efficiency without resilience is fragile, and their ratings reflect that hard-won awareness (Ralston & Blackhurst, 2020).

4.2 Ensemble Models and the Limits of Linear Thinking

The comparative model performance results (Table 2) deserve more than a straightforward reading of which algorithm scored highest. Yes, Random Forest outperformed the others at 87% accuracy and R² = 0.70 — and yes, Gradient Boosting followed closely at 86% and R² = 0.69. But the more conceptually interesting observation is the consistent performance gap between the ensemble methods and Multiple Linear Regression, which plateaued at 79% accuracy and R² = 0.61. That gap is not simply a matter of algorithmic sophistication; it reflects something real about the structure of the underlying relationships.

Supply chain performance constructs — demand forecasting accuracy, inventory behaviour, risk exposure, operational efficiency — do not interact linearly. They compound, amplify, and sometimes counteract each other in ways that depend on context, sector, and organisational maturity (Li et al., 2020; Cadavid et al., 2020). A firm that improves its demand forecasting does not experience a proportional, additive improvement in resilience — the benefit is mediated by whether inventory systems can respond to updated signals, whether risk models are calibrated to the same data, and whether operational processes are flexible enough to act on the outputs. Linear regression, by design, cannot represent those interaction effects. Ensemble methods, by contrast, implicitly model them through hierarchical tree splits and boosted residual correction — which is arguably why they fit the data better here (Shahbazi & Byun, 2020; Wang et al., 2017).

It is worth being honest about what the R² = 0.70 for Random Forest does and does not mean. It means the model accounts for 70% of the variance in SCR scores across the test set — a respectable result, but also a reminder that 30% of that variance remains unexplained. Some of that unexplained portion is likely noise inherent to self-report data; some may reflect genuinely important variables not captured in the survey instrument, such as organisational culture, IT infrastructure quality, or sector-specific regulatory constraints (Chhetri et al., 2017; Makris et al., 2019). An R² of 0.70 is not the end of the story; it is an invitation to investigate what the model is missing.

4.3 The Central Role of Anticipatory Capabilities: Demand Forecasting and Risk Prediction

One of the clearest, most replicable findings across both the correlation analysis (Figure 1) and the feature importance results (Table 3) is the primacy of anticipatory over reactive capabilities. Demand Forecasting was the strongest correlate of Supply Chain Resilience (r = 0.74) and the highest-ranked predictor in the Random Forest model (importance score = 0.30). Risk Prediction was second on both counts (r = 0.72; importance = 0.28). Together they account for 58% of the model's total feature importance — a concentration that is hard to dismiss as coincidental.

The theoretical explanation is not difficult to articulate, even if the organisational reality is often harder to achieve. A supply chain that can anticipate what demand will look like three or six months out is in a fundamentally different position than one reacting to demand signals in real time. It can pre-position inventory, negotiate supplier contracts under less time pressure, route logistics capacity efficiently, and absorb unexpected spikes without cascading into stockouts (Ivanov & Dolgui, 2019; Shashi et al., 2019). Risk Prediction works analogously: disruptions that are anticipated — whether through geopolitical monitoring, weather modelling, or supplier health scoring — can be absorbed at far lower cost than those that arrive without warning (Chien et al., 2020; Sobb et al., 2020). The feature importance analysis (Table 3) essentially quantifies what experienced supply chain managers have long known intuitively: the biggest competitive advantage is not responding faster, it is needing to respond less often.

This finding also has a practical implication that the survey data (Figure 2) reinforce. When 72% of respondents rated demand forecasting as high-impact and 70% said the same of risk prediction, they were not simply endorsing the technology in the abstract — they were reporting felt experience of specific operational improvements. That convergence between the algorithmic feature importance and the practitioner perception data is, arguably, one of the more encouraging aspects of this study's design: two entirely different measurement approaches point toward the same conclusion (Bag et al., 2019; Allaoui et al., 2019).

4.4 Inventory Optimization and Data Integration as Structural Enablers

Inventory Optimization (IO) ranked third in feature importance (0.23) and correlated with SCR at r = 0.69 (Figure 1; Table 3) — a strong result, though one that perhaps requires contextualisation. Inventory optimisation is, in some respects, the operational expression of good forecasting: a model that accurately predicts demand can translate that prediction into leaner, more responsive stock levels, reducing both carrying costs and the risk of obsolescence (Govindan et al., 2013). The relatively high feature importance score for IO may therefore partly reflect the downstream benefit of upstream forecasting accuracy, rather than IO functioning as an entirely independent driver of resilience. Disentangling these contributions would require a longitudinal or interventional design that cross-sectional survey data cannot support.

Data Integration's fourth-place ranking (importance = 0.20) is, in some ways, the most conceptually interesting result of all — and the one most easily overlooked in a results section focused on headline model performance (Table 3). Data integration is not a capability in the same sense that forecasting or risk prediction are; it is the precondition that makes those capabilities possible. Without coherent, timely, and interconnected data flows across supplier networks, logistics platforms, warehouse management systems, and demand-sensing tools, even the most sophisticated ML algorithm is operating in an information vacuum (Boyson, 2014; Shang & You, 2019). The fact that data integration registers as a meaningful predictor of SCR in a survey of practitioners is perhaps a signal that organisations are increasingly aware of this dependency — aware, in other words, that data quality and pipeline integrity are not IT infrastructure problems but strategic supply chain problems (Radanliev et al., 2020; Lu, 2017).

4.5 Why Operational Efficiency Ranks Last — and Why That Is Not a Trivial Finding

Operational Efficiency (OE) was consistently the weakest predictor across both the feature importance analysis (importance = 0.15) and the correlation matrix (r = 0.66 with SCR), and the dimension rated least impactful by survey respondents at 65% (Table 3; Figure 1; Figure 2). This might seem counterintuitive — surely efficient operations contribute to resilience? They do, but the relationship appears to be more conditional than direct. Operational efficiency, in the supply chain context, often reflects lean process design, cost optimisation, and throughput maximisation. These are valuable in stable conditions; they become liabilities under disruption, when the slack that "inefficient" systems retain becomes the very buffer that allows recovery (Ivanov et al., 2013; Ralston & Blackhurst, 2020).

This tension between efficiency and resilience is well-documented in the operations management literature, and the feature importance results here are broadly consistent with it (Govindan et al., 2013; Bag et al., 2019). What the ML models appear to be detecting — and what practitioners seem to intuitively sense — is that anticipatory capabilities provide more durable protection against disruption than execution-layer optimisation alone. Efficiency matters, but it matters less when a system lacks the foresight to avoid the disruptions that efficiency cannot buffer.

4.6 Barriers to Full ML Deployment: The Gap Between Potential and Practice

Despite the broadly optimistic tone of the survey responses, the findings should not be read as evidence that ML adoption in supply chains is proceeding smoothly or equitably. Only 8% of respondents reported low perceived impact — but this is a sample of practitioners who already work in environments with some degree of ML exposure. The harder, more structurally significant question is what the distribution of views looks like in organisations that have not yet been able to adopt these tools, and the present study, by design, cannot answer that.

What the literature makes clear — and what remains as relevant in 2024 as it was when these reference works were published — is that barriers to adoption are substantial and unevenly distributed. Fragmented data infrastructures, a persistent shortage of professionals with both domain expertise and machine learning competence, high implementation and maintenance costs, and organisational cultures resistant to algorithmic decision-making all constrain the pace and breadth of ML diffusion across supply chains (Shang & You, 2019; Govindan et al., 2013; Chang et al., 2019). These barriers are not random in their distribution — they fall disproportionately on smaller firms, on organisations in lower-income economies, and on sectors with historically lower levels of digital infrastructure investment. The result is a growing divergence: supply chains with ML capabilities becoming more resilient over time, while those without fall further behind (Makris et al., 2019; Nagy et al., 2018).

This structural inequality deserves more than a paragraph in a limitations section. It is arguably the most consequential implication of this line of research — not whether ML can improve supply chain resilience under ideal conditions (the evidence suggests it can), but whether the conditions for its adoption can be made more accessible, and what the cost is when they cannot (Cai & Luo, 2020; Klemeš et al., 2020).

4.7 Limitations and Directions for Future Research

Several limitations of this study warrant explicit acknowledgement, not as ritual disclaimer but because they genuinely shape how the findings should be read. First, the cross-sectional design captures perceptions at a single point in time and cannot establish the direction or causality of the observed associations. That SCR correlates strongly with DF and RP does not mean improving demand forecasting causes resilience improvements — only that organisations where practitioners perceive both tend to co-occur. Longitudinal or quasi-experimental designs would be better positioned to test causal claims (Ralston & Blackhurst, 2020).

Second, the sample, while meaningfully sized at n = 200, is drawn from a single national context. Supply chain conditions, digital infrastructure, regulatory environments, and practitioner attitudes toward ML vary considerably across geographies — findings from one country may not transfer cleanly to others, and replication studies in different national or regional settings would considerably strengthen the evidence base (Makris et al., 2019; Koh et al., 2020).

Third, the Likert-based survey instrument, however carefully constructed, measures perception rather than operational performance. An organisation whose practitioners report high ML impact on demand forecasting may or may not actually achieve better forecast accuracy; the gap between self-assessed and objectively measured performance is well-documented and remains a boundary on interpretation here (Barreto et al., 2017; Shashi et al., 2019). Future work that links practitioner survey responses to firm-level operational KPIs — forecast error rates, stockout frequencies, disruption recovery times — would represent a meaningful methodological advance over the present design.

5. Conclusion

This study set out to examine where machine learning adds the most measurable value in supply chain management — and the answer, perhaps unsurprisingly, centres on anticipation rather than execution. Demand Forecasting and Risk Prediction emerged consistently as the dominant drivers of Supply Chain Resilience, both in the correlation analysis and across feature importance rankings from the best-performing Random Forest model (accuracy = 87%, R² = 0.70). Ensemble methods outperformed linear approaches by a meaningful margin, a finding that reflects the genuinely nonlinear structure of real operational relationships rather than mere algorithmic preference.

That said, the study's optimistic headline figures deserve a measure of interpretive restraint. The sample represents practitioners already operating within ML-enabled environments; the barriers that prevent adoption elsewhere — fragmented data infrastructure, talent scarcity, cost — remain substantial and structurally unequal across geographies and firm sizes. Resilience, ultimately, is not just a technical property but an organisational and economic one. Future research should link practitioner perceptions to objective operational metrics and extend the design to longitudinal settings, where causal claims can be more rigorously evaluated.

Author Contributions

M.J.I.R. conceived and designed the study, developed the research framework, prepared the survey instrument, collected and analyzed the data, implemented the machine learning models, interpreted the results, and wrote the original manuscript. M.J.I.R. also reviewed, revised, and approved the final version of the manuscript.

Acknowledgement

The author would like to express sincere gratitude to the supply chain professionals from the logistics, manufacturing, healthcare, and agriculture sectors who participated in this study and provided valuable responses. The author also acknowledges the Department of Computer Science and Engineering, Daffodil International University, Dhaka, Bangladesh, for academic support and encouragement during the preparation of this research.

Competing Financial Interests

The author M.J.I.R.  declares that there are no competing financial interests or personal relationships that could have appeared to influence the work reported in this article.

References