Demand Forecasting

Quick answer / definition

Demand Forecasting predicts how many units of a product (or dollar sales) you’re likely to sell over a future period. It measures expected customer demand across channels and time windows and is used in inventory planning, purchasing, staffing, and marketing. It matters because better forecasts reduce stockouts, overstock, and missed revenue while improving cash flow and customer experience.

Why it matters

• Revenue: Accurate demand forecasts help you have the right SKUs available when customers buy, avoiding lost sales from stockouts.
• Profitability: Reduces excess inventory carrying costs and markdowns from overstock.
• Customer experience: Lower backorders and faster fulfillment increase lifetime value and repeat purchase rates.
• Marketing performance: Aligns ad spend and promotions with product availability to avoid driving demand you can’t fulfill.
• Operational efficiency: Improves purchasing, production scheduling, and labor planning.
• Decision-making: Provides quantitative inputs for pricing, launch cadence, and assortment decisions.

What is Demand Forecasting?

Demand Forecasting is the process of estimating expected product demand for a future period using historical sales data plus external inputs—seasonality, promotions, product launches, marketing plans, and supplier lead times. It includes short-term tactical forecasts (days–weeks) used for fulfillment and reorder decisions, and medium-to-long-term forecasts (months–quarters) used for budget, purchasing, and assortment planning.

What it typically includes:

• Time-series sales data by SKU, channel, and location
• Planned promotions, advertising spend, and launches
• Seasonality, calendar events, and anomalies
• Supplier lead times and minimum order quantities
• Inventory on-hand and pipeline (in-transit)

What it excludes (unless explicitly modeled):

• Micro-level customer intent not tied to observed behavior (e.g., unknown future viral events)
• Unplanned supply disruptions unless scenario-modeled

When businesses use it: daily/weekly for replenishment, monthly/quarterly for buying and financial planning, and continuously when running promotions or launching products. A high forecast relative to actual sales indicates overforecasting (risk of excess inventory); a low forecast indicates underforecasting (risk of stockouts).

Important terminology:

• SKU-level forecast: Forecast for a single product/variant.
• Aggregate forecast: Forecast for a group (category, brand, total demand).
• Bias: Systematic over- or under-forecasting direction.
• Forecast error: Difference between forecast and actual (measured with MAPE, MSE, MAE).
• Lead time: Time from order to receipt—critical for reorder calculations.

Formula / calculation

Demand Forecasting is a process, not a single numeric metric, but forecasts and forecast accuracy are measurable. Common formulas used in forecasting and evaluation:

1) Simple weighted moving average (forecast)

Forecast = (w1*Sales1 + w2*Sales2 + ... + wn*Salesn) / (w1 + w2 + ... + wn)

Where Sales1..n are recent periods and w1..n are weights (higher weight for recent periods).

2) Exponential smoothing (single)

Forecast_t = α * Actual_{t-1} + (1 - α) * Forecast_{t-1}

α (alpha) is the smoothing factor between 0 and 1.

3) Forecast error – Mean Absolute Percentage Error (MAPE)

MAPE = (1/n) * Σ |(Actual_i - Forecast_i) / Actual_i| * 100

Where n is number of periods. MAPE expresses average percentage error and is commonly used to benchmark accuracy.

Worked example: weighted moving average + MAPE

Scenario: Weekly sales for SKU-A for past 4 weeks: Week4=120, Week3=100, Week2=80, Week1=60. Weights (more recent higher): 4,3,2,1.

1. Compute weighted sum: (4*120) + (3*100) + (2*80) + (1*60) = 480 + 300 + 160 + 60 = 1,000
2. Sum of weights: 4+3+2+1 = 10
3. Forecast for next week = 1,000 / 10 = 100 units

If actual next-week sales = 110 units, forecast error = |110 - 100| = 10 units; MAPE for this one-period sample = (10 / 110) * 100 = 9.09%.

How it works (practical 6-step process)

1. Collect and clean data. Gather sales, inventory, marketing calendar, price changes, and supplier lead times. Measure: data completeness and anomalies. Why: garbage-in yields bad forecasts.
2. Segment SKUs. Group products by demand patterns (fast movers, slow movers, seasonal, intermittent). Measure: sales volume, variability, and intermittence. Why: different models work better for different segments.
3. Choose model(s). Apply simple statistical methods (moving averages, exponential smoothing) for stable SKUs; use causal models (regression, machine learning) for promotion-driven SKUs. Measure: out-of-sample accuracy (MAPE). Why: model choice affects accuracy and complexity.
4. Incorporate business inputs. Add planned marketing, promotions, and new-product introductions. Measure: expected uplift assumptions. Why: historical-only models miss planned demand shifts.
5. Generate and review forecasts. Produce SKU-level and aggregate forecasts; review by merchants and supply planners. Measure: bias and exceptions. Why: human review catches obvious errors (e.g., mistaken zeroes).
6. Monitor and iterate. Track forecast accuracy, log exceptions, and update models weekly/monthly. Measure: MAPE, bias, service level. Why: continuous improvement reduces inventory costs and stockouts.

Key components / factors

• Historical sales: Primary signal; accuracy depends on data quality and granularity.
• Promotions and marketing: Paid ads, email blasts, and discounts create demand spikes that must be modeled separately.
• Seasonality and calendar effects: Holidays, weather, and annual events change baseline demand predictably.
• Channel and device mix: Conversion differences by traffic source or device affect how marketing translates to sales.
• Product lifecycle: New launches, growth, maturity, and decline phases require different forecasting approaches.
• Pricing and assortment changes: Price increases, bundles, or removal of SKUs change demand elasticity.
• Lead times and supplier reliability: Longer or variable lead times require higher safety stock and different reorder points.
• Inventory and fulfillment constraints: Stock availability and shipping delays directly affect realized sales.
• Data/analytics/tracking: Missing or misattributed sales (e.g., offline, returns) reduce forecast accuracy.

Example (realistic ecommerce scenario)

Starting situation: DTC brand sells a seasonal jacket (SKU-J). Historical weekly sales for 6 weeks leading to autumn launch: [40, 45, 55, 70, 90, 120]. They expect a marketing push week after next. Lead time from supplier = 3 weeks. Current on-hand inventory = 300 units. Price = $120. Cost per unit = $48.

Diagnosis:

• Use weighted moving average (weights 1,2,3,4,5,6 for older→recent) to forecast week 7 baseline.

1. Weighted sum = (1*40)+(2*45)+(3*55)+(4*70)+(5*90)+(6*120) = 40+90+165+280+450+720 = 1,745
2. Sum weights = 1+2+3+4+5+6 = 21
3. Baseline forecast week 7 = 1,745 / 21 ≈ 83 units

Marketing team plans a 30% uplift the following week due to an email campaign starting week 8. Because lead time is 3 weeks, any reorder placed today arrives week 10. With week 7 baseline 83, week 8 expected = 83 * 1.3 ≈ 108 units. For safety, planner adds a 10% buffer for variability → reorder target for week 8 demand = 119 units.

Action taken: Place reorder for 400 units to cover expected demand for weeks 8–11 (108+100+90+85 approximate) plus safety stock because of supplier MOQs and ramping campaign. After shipping and fulfillment, actual week 8 sales = 110 (close to forecast).

Result & impact (conservative numbers):

• Revenue captured in week 8 = 110 * $120 = $13,200
• Had they under-forecasted and stocked only 50 units, they would have missed 60 sales = $7,200 lost revenue that week.
• Inventory cost for extra 289 units (400 ordered - 111 expected remaining after week 8) = 289 * $48 = $13,872 carrying cost basis (not including storage/financing).
• Trade-off: Preventing $7,200 immediate lost sales vs. added inventory carrying. With better accuracy and smaller safety buffer, reorder could be optimized to reduce carrying cost.

Benchmark / what is a good metric?

There is no universal "good" forecast accuracy: acceptable MAPE varies by SKU type and business. Typical guidance (industry-dependent):

• Low-variability, high-volume SKUs: MAPE under 10% is often achievable.
• Seasonal or new SKUs: MAPE of 20–50% is common due to limited history.
• Intermittent (sporadic) demand SKUs: MAPE can be very high and less meaningful; use service-level targets instead.

Benchmarks depend on product lifecycle, data quality, segmentation, and lead time. If you find published benchmarks, confirm the business context matches your model (B2C vs wholesale, SKU granularity, promotion activity).

How to improve / optimize demand forecasting (prioritized)

1. Segment SKUs by demand pattern (high impact).
   • What to change: Split SKUs into fast, slow, seasonal, and intermittent groups.
   • Why it works: Tailored models reduce error caused by averaging dissimilar behaviors.
   • How to implement: Use weekly sales variance and volume thresholds to assign segments, then pick models per segment.
   • Monitor: MAPE by segment and inventory turns.
2. Measure and report forecast error consistently (high).
   • What to change: Track MAPE, bias, and service level for SKU, category, and aggregate.
   • Why it works: Visibility identifies where models fail and whether bias exists.
   • How to implement: Automate weekly accuracy reports in your BI tool.
   • Monitor: MAPE trends and bias direction.
3. Include promotional and marketing plans (high).
   • What to change: Add planned campaign start/end, expected reach, and historical uplift multipliers.
   • Why it works: Prevents under-forecasting when marketing drives incremental demand.
   • How to implement: Integrate calendar inputs with model and test uplift percentages from past promotions.
   • Monitor: Promo lift accuracy and incremental ROAS.
4. Adjust for lead time variability (medium).
   • What to change: Use service-level based safety stock calculations incorporating lead-time variance.
   • Why it works: Reduces stockouts from unpredictable supplier delays.
   • How to implement: Calculate safety stock = z * σdemand * sqrt(lead_time), where z is z-score for desired service level.
   • Monitor: Stockout frequency and fill rate.
5. Automate data cleaning and anomaly detection (medium).
   • What to change: Remove returns, mis-logged orders, and one-off spikes before model training.
   • Why it works: Prevents outliers from skewing future predictions.
   • How to implement: Use rules (e.g., threshold on percent change) and manual review for flagged weeks.
   • Monitor: Number of flagged anomalies and accuracy improvements after cleaning.
6. Test machine learning models for complex SKUs (lower initially).
   • What to change: Try regression or tree-based models incorporating price, ads, and external signals.
   • Why it works: Can capture non-linear relationships and interactions.
   • How to implement: Pilot on 5–10 SKUs with rich history; backtest on holdout periods.
   • Monitor: Backtest improvement vs simpler models and compute maintenance cost.

Best practices

• Track forecast accuracy at SKU-level and aggregated roll-ups; use both MAPE and bias.
• Re-forecast frequently for short lead-time items (weekly) and monthly for long-term planning.
• Maintain a published promotion and launch calendar that forecasting and procurement teams sign off on.
• Use safety stock based on service-level targets and lead-time variability, not a flat percent buffer.
• Document assumptions for each forecast (promotion uplift, price change) so you can audit results.
• Hold regular S&OP (sales & operations planning) reviews to reconcile analytics forecasts with merchant insights.
• Invest in data hygiene: returns, cancellations, and offline sales should be reconciled into the dataset.
• Measure the cost of forecast errors: quantify lost sales vs carrying costs to inform risk tolerance.
• Use ensembled approaches: combine simple statistical models with causal adjustments for robust baseline forecasting.

Common mistakes to avoid

• Forecasting at the wrong level (why it happens): Teams use total-site forecasts and apply them to every SKU.
  • Why harmful: Hides SKU-level variability; leads to frequent stockouts or excesses.
  • Correct approach: Forecast at SKU×channel×location where feasible; use aggregate only for long-term planning.
• Ignoring promotions and marketing plans:
  • Why harmful: Under-forecasts during promo windows and overstocks afterward.
  • Correct approach: Tag historical promo events, measure uplift, and add planned promo inputs.
• Relying solely on last-year data for seasonal items:
  • Why harmful: Misses growth trends and changing customer behavior year-over-year.
  • Correct approach: Use multiple years with trend adjustments and recent-period weighting.
• No measurement of forecast bias:
  • Why harmful: Small systematic errors compound over time; planners won’t know if they consistently under- or over-forecast.
  • Correct approach: Track bias (signed error) and correct models or assumptions.
• Poor data integration:
  • Why harmful: Missing offline or 3P marketplace sales produces incorrect forecasts.
  • Correct approach: Consolidate all sales channels and returns into the forecasting dataset, and reconcile with accounting.

Demand Forecasting vs related concepts

Demand Forecasting vs Sales Forecasting

• Demand Forecasting: Focuses on expected customer demand (units/dollars) often used for inventory and supply planning.
• Sales Forecasting: Often more finance-focused, predicting revenue for budgeting and cash-flow with attention to price and channel mix.
• Key difference: Demand forecasts emphasize units and fulfillment; sales forecasts emphasize revenue and margin.

Demand Forecasting vs Demand Planning

• Demand Forecasting: Quantitative estimate of future demand (model-driven).
• Demand Planning: Process that uses forecasts plus business inputs to create a plan (purchasing, inventory, promotions).
• Key difference: Forecasting is the analytical output; planning is the cross-functional action that follows.

Demand Forecasting vs Inventory Replenishment

• Demand Forecasting: Predicts future demand.
• Inventory Replenishment: Operational calculation (reorder point, order quantity) that uses the forecast plus lead time and service-level targets.
• Key difference: Forecast informs replenishment; replenishment determines when/how much to order.

When should you track Demand Forecasting?

• Who: Ecommerce founders, merchants, inventory planners, finance, and growth teams should track forecasts.
• Stage of growth: Start basic forecasting as soon as you hold inventory and sell across multiple days/weeks; sophistication grows with SKUs and channels.
• Frequency: Review short-term forecasts weekly (or daily for high-turn SKUs); review monthly/quarterly for planning horizons.
• Segments to analyze: SKU×channel×location, promotions vs baseline, new products vs established, high-value vs low-value SKUs.
• Other metrics to view alongside: Inventory on-hand, days of inventory, lead time, fill rate, stockout rate, margin, and marketing spend.

Related ecommerce metrics

• Inventory turnover: How quickly inventory sells — connects to forecast accuracy and ordering frequency.
• Fill rate / service level: Percentage of demand met from stock — directly affected by forecast quality and safety stock.
• Stockout rate: Frequency of being unable to fulfill demand — indicates underforecasting or supply issues.
• Days of inventory (DOI): Inventory level relative to forecasted sales — used to size orders and cash needs.
• MAPE / MAE / MSE: Forecast error metrics used to measure model performance.
• Promotional uplift: Incremental sales from a campaign — must be modeled into forecasts for accuracy.

FAQs

1. How often should I update my demand forecast?

   Update short-term forecasts weekly (or daily for fast-moving SKUs) and update longer-term forecasts monthly or quarterly. Frequency depends on SKU lead times and variability.

2. What forecasting accuracy should I aim for?

   Aim for the lowest MAPE you can achieve given SKU type: under 10% for stable high-volume SKUs is realistic; seasonal or new SKUs will have higher error. Use segment-specific targets rather than a single site-wide number.

3. Can I forecast new products?

   Yes, but accuracy is limited. Use analogues (similar SKU historical patterns), pre-launch demand signals (email sign-ups), and conservative safety stock. Track early sales closely and adjust.

4. Should I use machine learning for forecasting?

   ML helps for complex, promotion-driven SKUs or when many explanatory variables exist. Start with simple statistical models; pilot ML on subsets and compare backtest performance before scaling.

5. How do promotions affect forecasts?

   Promotions produce incremental demand that historical baselines don’t capture. Model promo uplift from past events and include planned promo parameters in the forecast pipeline.

6. How do I measure if my forecasts are improving?

   Track MAPE, bias, fill rate, and stockout rate over time by SKU segment. Improvement should show reduced error and fewer stockouts without a disproportionate rise in carrying costs.

7. What data sources are essential?

   Sales by SKU and channel, returns, inventory on-hand, lead times, marketing calendar and spend, price changes, and any marketplace or offline sales that affect demand.