Decision Frameworks — Inventory Demand Planning

Decision Frameworks — Inventory Demand Planning

This reference provides the detailed decision logic, optimization models, method selection trees, and segmentation methodologies for inventory demand planning at multi-location retailers. It is loaded on demand when the agent needs to make or recommend nuanced planning decisions.

All thresholds, formulas, and cost assumptions reflect US multi-location retail operations managing hundreds of SKUs across grocery, general merchandise, seasonal, and promotional categories.

1. Forecast Method Selection Trees

1.1 Primary Selection Algorithm

The goal is to match each SKU to the forecasting method that minimizes WMAPE on out-of-time holdout data. In practice, most organizations cannot afford per-SKU model optimization across hundreds of items. Instead, classify items into demand pattern archetypes and assign methods by archetype.

Step 1 — Classify the Demand Pattern

Compute the following statistics on the most recent 52 weeks of de-promoted demand data (remove promotional lift periods before computing):

Step 2 — Map to Demand Archetype

Step 3 — Assign Forecasting Method

1.2 Model Switching Rules

Do not switch methods based on a single bad week. Models need time to prove or disprove themselves.

1.3 Parameter Optimization Protocol

For exponential smoothing methods, optimize parameters using grid search on time-series cross-validation (rolling origin, 1-step ahead forecast, 26+ origins).

Grid search ranges:

Optimization metric: Minimize WMAPE on the holdout origins. If two parameter sets produce WMAPE within 1 percentage point, prefer the set with lower α (more smoothing) for stability.

Overfitting guard: If the optimized model produces WMAPE on the holdout that is

5 percentage points better than on the fitting data, the model is likely overfit. Increase smoothing (lower α) until the gap narrows to <3 points.

2. Safety Stock Optimization Models

2.1 Standard Safety Stock (Normal Demand, Fixed Lead Time)

When demand follows a roughly normal distribution and lead time is consistent:

SS = Z × σ_d × √(LT)

Where:

• Z = z-score for the target service level (see lookup table below)
• σ_d = standard deviation of demand per period (use same period as LT)
• LT = lead time in periods

Z-Score Lookup:

2.2 Safety Stock with Lead Time Variability

When vendor lead times are uncertain (CV of lead time > 0.15):

SS = Z × √(LT_avg × σ_d² + d_avg² × σ_LT²)

Where:

• LT_avg = average lead time in periods
• σ_LT = standard deviation of lead time in periods
• d_avg = average demand per period

Practical note: Many planners underestimate lead time variability because they measure "vendor ship date to DC receipt" without accounting for receiving delays, quality holds, or weekend/holiday dead time. Measure lead time from PO release to "available to sell" — this is the operationally relevant metric.

2.3 Safety Stock with Review Period

For periodic review systems (review every R periods):

SS = Z × σ_d × √(LT + R)

The review period adds exposure time — between reviews, you cannot react to demand changes. Weekly review (R=1) on a 2-week lead time item needs safety stock for √3 weeks. Monthly review (R=4) on the same item needs safety stock for √6 weeks — 41% more.

2.4 Safety Stock for Intermittent Demand

Normal-distribution formulas fail when demand has many zero periods. Use empirical (bootstrapped) safety stock instead:

1. Collect the last 52 periods of demand data (include zeros).
2. Generate 10,000 bootstrap samples of length (LT + R) by random sampling with replacement from the historical demand.
3. Compute the sum of each bootstrap sample (= simulated demand during lead time + review).
4. The safety stock is the (service level)th percentile of the simulated demand totals minus the mean simulated demand total.

Example: For 95% service level, safety stock = P95 of bootstrap demand — mean of bootstrap demand. This captures the skewed, zero-inflated distribution that parametric formulas miss.

2.5 Safety Stock for New Products (No History)

When an item has < 8 weeks of own demand history:

1. Identify 3–5 analogous items matching on: category, price point (±20%), brand tier (national/private label), pack size, and target demographic.
2. Compute the average σ_d and CV across the analogs.
3. Apply a "new product uncertainty premium" of 1.25× to the analog σ_d.
4. Use the standard formula with the inflated σ_d: SS = Z × (1.25 × σ_d_analog) × √(LT + R).
5. Every 2 weeks, blend own-history σ_d with the analog σ_d. By week 8, use 70% own history and 30% analog. By week 12, use 100% own history.

2.6 Safety Stock Cost-Optimization

The naive approach is to set a service level target and compute SS. The sophisticated approach is to optimize the tradeoff between holding cost and stockout cost:

Optimal SL = 1 − (H / (H + S × D/Q))

Where:

• H = holding cost per unit per period
• S = stockout cost per unit (lost margin + customer goodwill + substitution cost)
• D = demand per period
• Q = order quantity

For most retailers, stockout cost on A-items is 3–5× the unit margin (including lost customer visits and substitution effects), which pushes optimal SL to 96–98%. For C-items, stockout cost is approximately equal to the unit margin, yielding optimal SL of 88–92%.

3. Promotional Planning Frameworks

3.1 Promotional Lift Estimation Methodology

Promotional lift is always computed relative to the baseline forecast (the forecast that would have been generated without the promotion). Contaminating the baseline with promotional history is the #1 source of systematic forecast error in retail.

Step 1 — Establish Clean Baseline

Strip promotional periods from the demand history before fitting the baseline model. Flag weeks as "promotional" if any of the following were active:

• Temporary price reduction (TPR) > 5% off regular price
• Feature in circular, digital ad, or endcap display
• BOGO or multi-buy offer
• Cross-promotion in another category

After stripping, interpolate the gaps using the forecast model fitted to non-promotional periods. This creates a "what would have sold at regular price" baseline.

Step 2 — Compute Historical Lifts

For each historical promotional event on this SKU:

Lift Ratio = Actual Promo-Period Sales / Baseline Forecast for Promo Period

A lift ratio of 2.5 means the promotion drove 2.5× baseline volume (150% incremental).

Organize lift ratios by promotional mechanism:

Step 3 — Apply Lift to Current Forecast

Promo Forecast = Baseline Forecast × Lift Ratio

When multiple promotional mechanisms are combined, do NOT multiply individual lifts. Use the combined-mechanism lift from historical data or the table above. The interaction effects are sub-additive (display alone = 1.5× and circular alone = 1.8× does not mean display + circular = 2.7×; it's typically 2.0–2.5×).

Step 4 — Model the Post-Promotion Dip

Post-Promo Demand = Baseline × (1 − Dip Factor × Decay)

Default dip factors by product type:

3.2 Cannibalization Estimation

When SKU A is promoted, substitutable SKU B loses sales. The cannibalization rate is:

Cannibalization Rate = ΔB_down / ΔA_up

Where ΔA_up is the incremental lift on A and ΔB_down is the volume loss on B during the same period.

Default estimates when no cross-elasticity data exists:

Important: Cannibalization is bidirectional across the category. When building the category-level promotional plan, sum the cannibalization effects across all substitutes to compute the true category-level lift (which is always less than the item-level lift).

3.3 Forward-Buy and Pantry Loading

Deep promotions cause customers to purchase ahead of their consumption schedule. Forward-buy volume is demand pulled from future periods, not incremental category demand.

Forward-buy estimation:

Forward-Buy Volume = Incremental Lift × Forward-Buy Factor

The forward-buy factor tells you what fraction of the incremental lift came from pantry loading rather than true consumption increase. This directly feeds the post-promo dip calculation — the dip is approximately equal to the forward-buy volume spread over its consumption period.

3.4 Promotional Calendar Planning

When planning the annual promotional calendar, apply these rules:

1. Minimum inter-promotion gap: 4 weeks between promotions on the same SKU. Shorter gaps train customers to only buy on deal, eroding brand equity and baseline velocity.
2. Maximum promotional frequency: 13 weeks per year (25%) for any single SKU. Beyond this, the "promotional price" becomes the reference price in consumers' minds.
3. Seasonal alignment: Promote seasonal items during the build phase (first 40% of the season), not during peak or decline. Promoting at peak wastes money on demand that would have occurred anyway. Promoting during decline is a markdown, not a promotion.
4. Cross-category coordination: Avoid promoting close substitutes simultaneously. Stagger promotions across substitutes by at least 2 weeks to avoid self-cannibalization.
5. Vendor funding alignment: Match promotional timing to vendor trade fund availability. Many CPG manufacturers operate on calendar quarters — funds not committed by quarter-end expire. Plan key promos in weeks 8–12 of each quarter when vendors are motivated to spend remaining funds.

4. ABC/XYZ Segmentation Methodology

4.1 ABC Classification (Value)

ABC classification segments SKUs by their financial contribution. The classification drives differentiated investment in forecasting effort, safety stock, review frequency, and management attention.

Classification Procedure

1. Select the value metric. Options in order of preference:

   • Gross margin contribution (best — focuses investment on profit, not revenue)
   • Revenue (acceptable when margin data is unavailable)
   • Unit volume (use only for warehouse space planning, not financial investment)

2. Compute trailing 52-week value for each active SKU.

3. Sort descending by the value metric.

4. Compute cumulative % of total value and classify:

5. Exception overrides:
   • New items (< 13 weeks) are auto-classified one tier higher than their data suggests until they have sufficient history. A new item computing as C is treated as B.
   • Items with contractual obligations (planogram commitment, vendor agreement) are classified minimum B regardless of current sales velocity.
   • Items flagged as strategic by merchandising (e.g., traffic drivers, competitive price match items) are classified minimum A.

Reclassification Schedule

Run ABC reclassification quarterly. Between quarters, items are reclassified only if they cross a threshold by >50% (e.g., an item must contribute >120% of the A/B boundary to move from B to A mid-quarter). This prevents oscillation at class boundaries.

4.2 XYZ Classification (Predictability)

XYZ classification segments SKUs by demand forecast difficulty. It drives differentiated forecasting method selection and safety stock strategies.

Classification Procedure

1. Compute de-seasonalized, de-promoted demand for each SKU over the trailing 52 weeks. Remove seasonal indices and promotional lift periods so that the variability metric reflects genuine demand uncertainty, not planned variation.

2. Compute the coefficient of variation (CV):

   CV = σ_demand / μ_demand
   

   Use the de-seasonalized, de-promoted demand series.

3. Classify:

4. Supplement with ADI (Average Demand Interval): Items with ADI > 2.0 (meaning demand occurs less than every other period) should be classified Z regardless of CV, because the intermittency itself creates forecast difficulty that CV alone doesn't capture.

4.3 Combined ABC/XYZ Policy Matrix

4.4 Migration Tracking

Track SKU movement between segments quarterly. Key migration patterns to monitor:

5. Vendor Management Decision Logic

5.1 Vendor Tier Classification

Classify vendors into tiers based on annual purchase volume, strategic importance, and supply risk profile:

5.2 Vendor Scorecard Metrics

Score each vendor quarterly on a 0–100 scale across these dimensions:

Composite score thresholds:

5.3 Vendor Lead Time Management

Lead time management is the demand planner's most underleveraged tool for reducing inventory investment. A 1-day reduction in lead time across all vendors can reduce aggregate safety stock by 5–8%.

Lead time decomposition:

Actions to reduce lead time:

1. For strategic vendors: negotiate VMI (vendor-managed inventory) where the vendor monitors your inventory and ships proactively. Eliminates order processing delay.
2. For all vendors: provide rolling 8-week forecasts to allow pre-positioning. Reduces production/picking time on non-stock items.
3. Internally: invest in receiving automation (ASN-enabled receiving, barcode scanning) to cut receiving from 2–3 days to same-day.
4. Negotiate consolidated weekly shipments vs. per-PO shipments to reduce transit frequency while maintaining fill rate.

5.4 MOQ (Minimum Order Quantity) Negotiation Framework

When a vendor's MOQ creates excess inventory, evaluate these options in order:

5.5 Vendor Negotiation for Lead Time Reduction

Preparation checklist before negotiating:

1. Document your current order volume and growth trajectory with this vendor.
2. Compute the cost of their current lead time to your business: excess safety stock carrying cost + stockout cost from lead time variability.
3. Identify what you can offer in return: longer-term commitments, higher volumes, fewer order frequency changes, rolling forecasts.
4. Know your BATNA (best alternative): have a qualified secondary supplier identified.

Negotiation structure:

1. Present the data: "Over the past 6 months, your average lead time has been X days with a standard deviation of Y. This variability costs us $Z annually in excess safety stock."
2. Propose the target: "We're requesting a committed lead time of X−2 days with a guarantee of CV < 0.15."
3. Offer the exchange: "In return, we can commit to rolling 8-week forecasts updated weekly, and we'll consolidate to 2 orders per week instead of daily."
4. Set the timeline: "Let's implement this for Q2 and review the scorecard at the end of Q2 QBR."

6. Seasonal Buy and Markdown Timing Models

6.1 Seasonal Buy Planning

Seasonal categories require forward commitments because lead times exceed the selling season. The buy decision has two components: the initial buy (pre-season) and the in-season reorder (if the vendor supports it).

Initial Buy Calculation

Initial Buy = Season Forecast × Initial Commitment % − Carry-Over Inventory

In-Season Reorder Triggers

Monitor sell-through rate weekly starting from week 2 of the season:

Sell-Through Rate = Units Sold / (Units Sold + Units On-Hand + Units On-Order)

6.2 Markdown Timing and Depth Model

The markdown decision balances margin recovery against sell-through velocity. Every week of delay costs margin because holding costs accrue and the remaining selling window shrinks.

Markdown Decision Matrix

Markdown Velocity Curve

After applying a markdown, monitor the velocity response:

6.3 Season-End Liquidation Decision

When the selling season is ending and inventory remains:

Liquidation Net Recovery = (Liquidation Price × Remaining Units) − Logistics Cost
Hold-to-Next-Season Net = (Expected Sell Price × Sell-Through Estimate) 
                          − Holding Cost − Obsolescence Risk

Liquidation is preferred when:

• Hold-to-next-season sell-through estimate < 60% (style risk, trend change)
• Holding cost for 9–12 months > 15% of original cost (typical for most retailers)
• Warehouse space is constrained and the space has higher-value alternative use
• The product is trend/fashion and will be visually dated next season

Holding is preferred when:

• Product is a classic/carryover style with minimal fashion risk
• Hold-to-next-season sell-through estimate > 80%
• Warehouse space is available at low marginal cost
• Liquidation offers are below variable cost (you'd lose money selling)

7. New Product Introduction Forecasting

7.1 Analogous Item Selection

The quality of a new product forecast depends almost entirely on the quality of the analogous items selected. Bad analogs produce bad forecasts regardless of the method.

Selection Criteria (rank by importance)

Analog Scoring

Score each candidate analog on the criteria above (1–5 scale per criterion, weighted). Select the top 3–5 analogs with composite scores > 3.5. If no analogs score > 3.0, the new product is truly novel — use category average with a 40% confidence band.

7.2 New Product Lifecycle Curve

Most new products follow a lifecycle curve with four phases:

Forecast by phase:

7.3 New Product Safety Stock Protocol

7.4 New Product Kill Decision

Not every new product succeeds. The kill decision should be structured, not emotional:

When a kill decision is made:

1. Cancel all open POs immediately.
2. Halt any planned promotions.
3. Mark down remaining inventory at 30% off for 3 weeks, then 50% for 2 weeks.
4. Liquidate any remainder after 5 weeks.
5. Document the post-mortem: why did the analog-based forecast fail? Was it the analogs, the product, the pricing, or the competitive context?

8. Demand Sensing and Exception Management

8.1 Real-Time Demand Signal Monitoring

In addition to periodic forecast reviews, monitor for demand signals that require immediate attention between forecast cycles:

8.2 Forecast Override Governance

Manual overrides are necessary but dangerous. Ungoverned overrides introduce bias and degrade forecast accuracy over time.

Override rules:

1. All overrides must be documented with a reason code and quantitative justification.
2. Override authority by magnitude:
   • ±10%: Planner can override without approval
   • ±10–25%: Requires planning manager approval
   • ±25–50%: Requires director approval

   • ±50%: Requires VP approval (or planning committee)

3. Override accuracy tracking: Every override is tracked against actuals. If a planner's overrides have a WMAPE > 40% over a quarter, their override authority is reviewed.
4. Sunset rule: Overrides expire after 4 weeks. If the condition persists, a new override (with fresh justification) must be created. This prevents stale overrides from contaminating forecasts months later.
5. No "consensus" overrides: Overrides from demand review meetings where forecasts are adjusted to match sales team wishful thinking are the #1 source of positive bias. Require every meeting override to cite a specific, verifiable external signal.

9. Inventory Health Diagnostics

9.1 Weeks of Supply Analysis

Weeks of supply (WOS) is the primary pulse-check metric for inventory health. Compute at the SKU level, aggregate to category, and review weekly.

WOS = On-Hand Inventory (units) / Average Weekly Demand (units)

Use the forward-looking forecast for the denominator, not trailing sales. Trailing sales understates demand when items have been out of stock (you can't sell what you don't have).

WOS Health Bands:

9.2 Inventory Turns and GMROI

Inventory Turns:

Annual Turns = Annual COGS / Average Inventory at Cost

GMROI (Gross Margin Return on Inventory Investment):

GMROI = Gross Margin $ / Average Inventory at Cost

A GMROI of 2.0 means you earn $2 in gross margin for every $1 invested in inventory. Minimum acceptable GMROI varies by category but should generally exceed the company's cost of capital divided by the gross margin percentage. For a retailer with 8% cost of capital and 35% gross margin, minimum GMROI = 0.08 / 0.35 = 0.23. In practice, most retailers target GMROI > 1.5 for healthy categories.

9.3 Dead Stock and Obsolescence Identification

Dead stock is inventory with zero sales for a defined period. It is the most expensive form of excess inventory because it generates zero return while consuming warehouse space and working capital.

Dead stock tiers:

Root cause analysis for dead stock:

Run quarterly. Categorize dead stock by root cause to prevent recurrence:

9.4 Allocation Logic for Multi-Location Retailers

When DC inventory is insufficient to fill all store-level demand, allocate using a priority framework rather than pro-rata distribution:

Priority 1: Prevent store stockout on A-items. Allocate first to stores with < 3 days of supply on A-items. Quantity = minimum of (days-to-next-DC-shipment × daily demand) to bridge until the next allocation cycle.

**Priority 2

…(truncated)