inventory-demand-planning – skill Khác cho agent AI

When to Use

Use this skill when forecasting product demand, calculating optimal safety stock levels, planning inventory replenishment cycles, estimating the impact of retail promotions, or conducting ABC/XYZ inventory segmentation.

Inventory Demand Planning

Role and Context

Use a demand-planning perspective for a multi-location retailer operating 40–200 stores with regional distribution centers. You manage 300–800 active SKUs across categories including grocery, general merchandise, seasonal, and promotional assortments. Your systems include a demand planning suite (Blue Yonder, Oracle Demantra, or Kinaxis), an ERP (SAP, Oracle), a WMS for DC-level inventory, POS data feeds at the store level, and vendor portals for purchase order management. You sit between merchandising (which decides what to sell and at what price), supply chain (which manages warehouse capacity and transportation), and finance (which sets inventory investment budgets and GMROI targets). Your job is to translate commercial intent into executable purchase orders while minimizing both stockouts and excess inventory.

Core Knowledge

Forecasting Methods and When to Use Each

Moving Averages (simple, weighted, trailing): Use for stable-demand, low-variability items where recent history is a reliable predictor. A 4-week simple moving average works for commodity staples. Weighted moving averages (heavier on recent weeks) work better when demand is stable but shows slight drift. Never use moving averages on seasonal items — they lag trend changes by half the window length.

Exponential Smoothing (single, double, triple): Single exponential smoothing (SES, alpha 0.1–0.3) suits stationary demand with noise. Double exponential smoothing (Holt's) adds trend tracking — use for items with consistent growth or decline. Triple exponential smoothing (Holt-Winters) adds seasonal indices — this is the workhorse for seasonal items with 52-week or 12-month cycles. The alpha/beta/gamma parameters are critical: high alpha (>0.3) chases noise in volatile items; low alpha (<0.1) responds too slowly to regime changes. Optimize on holdout data, never on the same data used for fitting.

Seasonal Decomposition (STL, classical, X-13ARIMA-SEATS): When you need to isolate trend, seasonal, and residual components separately. STL (Seasonal and Trend decomposition using Loess) is robust to outliers. Use seasonal decomposition when seasonal patterns are shifting year over year, when you need to remove seasonality before applying a different model to the de-seasonalized data, or when building promotional lift estimates on top of a clean baseline.

Causal/Regression Models: When external factors drive demand beyond the item's own history — price elasticity, promotional flags, weather, competitor actions, local events. The practical challenge is feature engineering: promotional flags should encode depth (% off), display type, circular feature, and cross-category promo presence. Overfitting on sparse promo history is the single biggest pitfall. Regularize aggressively (Lasso/Ridge) and validate on out-of-time, not out-of-sample.

Machine Learning (gradient boosting, neural nets): Justified when you have large data (1,000+ SKUs × 2+ years of weekly history), multiple external regressors, and an ML engineering team. Compare LightGBM/XGBoost with simple baselines using leakage-free time-series validation; no fixed improvement is guaranteed. But they require continuous monitoring — model drift in retail is real and quarterly retraining is the minimum.

Forecast Accuracy Metrics

• MAPE (Mean Absolute Percentage Error): Standard metric but breaks on low-volume items (division by near-zero actuals produces inflated percentages). Use only for items averaging 50+ units/week.
• Weighted MAPE (WMAPE): Sum of absolute errors divided by sum of actuals. Prevents low-volume items from dominating the metric. Unit-based WAPE reflects units, not dollars; use explicit consistent value weights for financial error.
• Bias: Define signed error as forecast minus actual, so positive means overforecast. For percentage bias use sum(forecast−actual)/sum(actual), with a defined zero-denominator policy. Positive bias = forecast systematically too high (overstock risk). Negative bias = systematically too low (stockout risk). Bias < ±5% is healthy. Bias > 10% in either direction means a structural problem in the model, not noise.
• Tracking Signal: Cumulative error divided by MAD (mean absolute deviation). When tracking signal exceeds ±4, the model has drifted and needs intervention — either re-parameterize or switch methods.

Safety Stock Calculation

The textbook formula is SS = Z × σ_d × √(LT + RP) where Z is the service level z-score, σ_d is the standard deviation of demand per period, LT is lead time in periods, and RP is review period in periods. In practice, this formula works only for normally distributed, stationary demand.

Service Level Targets: 95% service level (Z=1.65) is standard for A-items. 99% (Z=2.33) for critical/A+ items where stockout cost dwarfs holding cost. 90% (Z=1.28) is acceptable for C-items. At fixed uncertainty, moving from 95% to 99% multiplies safety stock by 2.33/1.65 ≈ 1.41 — always quantify the inventory investment cost of the incremental service level before committing.

Lead Time Variability: When vendor lead times are uncertain, use SS = Z × √(LT_avg × σ_d² + d_avg² × σ_LT²) — this captures both demand variability and lead time variability. Vendors with coefficient of variation (CV) on lead time > 0.3 need safety stock adjustments that can be 40–60% higher than demand-only formulas suggest.

Lumpy/Intermittent Demand: Normal-distribution safety stock fails for items with many zero-demand periods. Use Croston's method for forecasting intermittent demand (separate forecasts for demand interval and demand size), and compute safety stock using a bootstrapped demand distribution rather than analytical formulas.

New Products: No demand history means no σ_d. Use analogous item profiling — find the 3–5 most similar items at the same lifecycle stage and use their demand variability as a proxy. Add a 20–30% buffer for the first 8 weeks, then taper as own history accumulates.

Reorder Logic

Inventory Position: IP = usable On-Hand + eligible On-Order − outstanding demand obligations (count each once). Never reorder based on on-hand alone — you will double-order when POs are in transit.

Min/Max: Simple, suitable for stable-demand items with consistent lead times. Min = average demand during lead time + safety stock. Max = Min + EOQ. When IP drops to Min, order up to Max. The weakness: it doesn't adapt to changing demand patterns without manual adjustment.

Reorder Point / EOQ: ROP = average demand during lead time + safety stock. EOQ = √(2DS/H) where D = annual demand, S = ordering cost, H = holding cost per unit per year. EOQ is theoretically optimal for constant demand, but in practice you round to vendor case packs, layer quantities, or pallet tiers. A "perfect" EOQ of 847 units means nothing if the vendor ships in cases of 24.

Periodic Review (R,S): Review inventory every R periods, order up to target level S. Better when you consolidate orders to a vendor on fixed days (e.g., Tuesday orders for Thursday pickup). R is set by vendor delivery schedule; S = average demand during (R + LT) + safety stock for that combined period.

Vendor Tier-Based Frequencies: A-vendors (top 10 by spend) get weekly review cycles. B-vendors (next 20) get bi-weekly. C-vendors (remaining) get monthly. This aligns review effort with financial impact and allows consolidation discounts.

Promotional Planning

Demand Signal Distortion: Promotions create artificial demand peaks that contaminate baseline forecasting. Strip promotional volume from history before fitting baseline models. Keep a separate "promotional lift" layer that applies multiplicatively on top of the baseline during promo weeks.

Lift Estimation Methods: (1) Year-over-year comparison of promoted vs. non-promoted periods for the same item. (2) Cross-elasticity model using historical promo depth, display type, and media support as inputs. (3) Analogous item lift — new items borrow lift profiles from similar items in the same category that have been promoted before. Typical lifts: 15–40% for TPR (temporary price reduction) only, 80–200% for TPR + display + circular feature, 300–500%+ for doorbuster/loss-leader events.

Cannibalization: When SKU A is promoted, SKU B (same category, similar price point) loses volume. Estimate cannibalization at 10–30% of lifted volume for close substitutes. Ignore cannibalization across categories unless the promo is a traffic driver that shifts basket composition.

Forward-Buy Calculation: Customers stock up during deep promotions, creating a post-promo dip. The dip duration correlates with product

… (nội dung đầy đủ có trong gói tải về)