A lightweight POS+ERP demand-sensing system halved the pain of simple heuristics: on a public 10-store, 50-item benchmark, gradient-boosted demand sensing cut simulated stockout incidence by roughly 30% compared with a week-lag rule while keeping order volume essentially unchanged. The paper shows small grocers can gain meaningful availability and working-capital benefits from modest analytics and disciplined feature engineering rather than enterprise AI rollouts.

Main Finding

A lightweight demand-sensing route-to-market (RTM) framework built from ordinary POS and ERP event tables can meaningfully improve short-term store-item demand forecasts for independent grocers. Using simple feature engineering and a tree‑boosting model, the study reduced forecast error relative to a lag-7 heuristic and—when forecasts were converted to replenishment rules—reduced simulated stockout incidence by 30.5% without materially increasing average order size (-0.8% change in the order index).

Key Points

• Practical objective: show what small grocers/distributors can implement with POS/ERP exports and modest analytics (Python, Excel, Power BI).
• Lightweight four-layer RTM architecture: data capture → feature engineering → forecasting → RTM decisioning.
• Feature set restricted to easily derived signals: store/item IDs, day-of-week, month/quarter, weekend flag, promotion proxy, lag-1, lag-7, 7-day and 28-day rolling means.
• Models compared: seasonal naive (lag-7), ridge regression, random forest, histogram gradient boosting (HistGradientBoosting).
• Best forecasting performance (holdout 90 days):
  • HistGradientBoosting: MAE = 4.43, RMSE = 5.54, sMAPE = 12.54%
  • Ridge: MAE = 4.59, RMSE = 5.80, sMAPE = 12.81%
  • Random forest: MAE = 4.72, RMSE = 5.91, sMAPE = 13.37%
  • Seasonal naive (lag-7): MAE = 5.56, RMSE = 7.02, sMAPE = 15.81%
• Replenishment simulation: stockout incidence fell from 14.20% (lag-7) to 9.86% (demand-sensing) → 30.5% reduction.
• Forecast error is concentrated in particular store-item cells; heat maps and exception flags are recommended managerial tools.
• Implementation roadmap: five phases (data audit → baseline dashboard → forecast model → replenishment rules → governance) with suggested KPIs (fill rate, on‑shelf availability, forecast MAE/bias, inventory days, OTIF, emergency replenishment frequency).
• Limitations: demonstration uses a clean public Kaggle benchmark; real-world issues (missing POS days, substitutions, stock-on-hand, price changes, inconsistent item coding, supplier constraints) require pre-modeling data work and extensions.

Data & Methods

• Dataset: Kaggle Store Item Demand Forecasting Challenge
  • Period: 2013-01-01 to 2017-12-31 (5 years)
  • Units: store-item-day
  • Stores: 10; Items: 50; Observations: ~913,000 (training rows 868,000; holdout rows 45,000)
  • Mean daily sales ≈ 44.02; range 0–132
• Feature engineering: calendar features, store/item identifiers, lag-1 and lag-7 sales, 7- and 28-day rolling means, binary promotion proxy.
• Models: seasonal naive (lag-7), ridge regression (linear), random forest, histogram gradient boosting (tree-based boosting).
• Evaluation: last 90 days held out; metrics MAE, RMSE, sMAPE; also heat‑map visualization of mean demand and forecast error.
• Replenishment simulation: converted forecasts to reorder-point rules and simulated stockout incidence and an “order-index” metric to gauge order volume change.
• Interpretability: feature importance / correlation analyses to show leading predictors (recent velocity, weekly seasonality, store/item identity).

Implications for AI Economics

• Adoption economics and productivity:
  • Demonstrates high value-to-cost ratio for modest ML investments in SMEs: non‑trivial service-level and working-capital improvements achievable without full-scale AI stacks.
  • Lowers the barrier for diffusion of forecasting ML into small retailers, potentially increasing operational efficiency and reducing waste at scale.
• Returns to data investments:
  • Highlights that “event discipline” (cleaning POS/ERP tables and basic feature engineering) delivers large marginal returns before needing complex models—important for prioritizing scarce SME resources.
  • Reinforces the economic case for public or subsidized technical assistance programs that improve data quality in small firms.
• Market structure and competition:
  • Improved availability and reduced emergency replenishments can strengthen small grocers’ competitiveness versus larger chains, potentially affecting local market shares and retail structure.
  • At scale, aggregate improvements in small-retailer inventory efficiency could alter wholesale demand smoothing and supplier production planning.
• Working capital, cash flow and supplier relations:
  • Better forecasts enable more efficient inventory allocation rather than inventory growth, improving cash conversion cycles and reducing forced emergency orders—important for margin‑constrained SMEs.
  • Forecast-driven prioritization can change supplier service dynamics (e.g., which SKUs/stores get preferential OTIF).
• Labor and logistics effects:
  • Route optimization informed by demand-sensing can improve drop density and reduce delivery frequency where appropriate, affecting fleet utilization and driver workload.
  • May reduce ad-hoc labor/urgent-order costs but could shift work toward analytics and exception management roles.
• Policy and public-good considerations:
  • Suggests policy levers (grants, training, standardized POS/ERP export formats, data‑governance templates) to accelerate SME adoption of lightweight ML.
  • Raises the case for publicly available benchmark data and toolkits so small retailers can adopt proven implementations with lower risk.
• Research and measurement implications:
  • Encourages economic research linking SME micro-level forecasting adoption to macro retail statistics (e.g., impacts on measured retail inventories, sales volatility).
  • Points to extensions that are economically important: integrating prices/margins, supplier lead times, stock-on-hand, perishability, weather and causal promotion effects, and combining forecasts with vehicle routing and credit analytics.
• Distributional and equity notes:
  • Benefits accrue only to firms that can get to “event discipline”; support may be needed for smaller or resource-constrained operators, especially in underserved communities where these grocers are anchor institutions.
• Potential negative externalities / caution:
  • Improved forecasting could concentrate supplier attention on higher-performing stores/products, potentially disadvantaging low-velocity locations unless governance ensures fairness.
  • Data quality and coding inconsistencies mean real-world gains may be lower than benchmark results without investments in cleaning and governance.

Overall, the paper provides an applied blueprint showing that modest, interpretable ML for tabular retail data can generate economically meaningful improvements for SMEs, with clear policy and research pathways to scale benefits while managing distributional and data-quality risks.