Qualitative methods

Expert judgment

• What: domain experts produce estimates.

• When: new products, limited data, structural breaks.

• Pros: fast, context-aware; Cons: bias, low reproducibility.

Delphi method

• What: iterative anonymous expert rounds to reach consensus.

• When: strategic horizons, disruptive markets.

• Pros: mitigates groupthink; Cons: time-consuming.

Market research and surveys

• What: intent, preference, and awareness surveys.

• When: pre-launch, pricing, elasticity sensing.

• Pros: captures latent demand; Cons: stated vs revealed behavior gap.

Scenario planning

• What: coherent narratives of alternative futures.

• When: long-term strategy, high uncertainty.

• Pros: stress-tests plans; Cons: qualitative, not precise.

Analogous forecasting

• What: use similar past products/markets as proxies.

• When: NPI and assortment extensions.

• Pros: actionable starting point; Cons: matching analogs is subjective.

Quantitative time-series methods

Naive and seasonal naive

• What: next value equals last (naive) or last season’s same period (seasonal naive).

• When: baselines, intermittent checks.

• Pros: zero setup, strong benchmark; Cons: ignores structure.

Moving average

• What: mean of last k observations.

• When: noise reduction, short-term smoothing.

• Pros: simple; Cons: lags trend, weak on seasonality.

Exponential smoothing (SES, Holt, Holt-Winters)

• What: weighted averages with decay, optionally modeling trend and seasonality.

• When: most operational horizons with clear patterns.

• Pros: robust, fast, handles level/trend/seasonality; Cons: less interpretable than regression with drivers.

Classical decomposition

• What: separate series into trend, seasonality, residuals (additive or multiplicative).

• When: diagnostics, stable seasonal patterns.

• Pros: interpretability; Cons: static seasonality, sensitive to outliers.

ARIMA and SARIMA

• What: autoregressive and moving average terms with differencing; seasonal variants capture periodicity.

• When: stationary or differenced series, medium history length.

• Pros: strong for autocorrelation structures; Cons: parameter tuning, assumes linearity.

Structural time series and Prophet-style approaches

• What: state-space models with components (trend, seasonality, holidays); Prophet popularized an accessible interface.

• When: multiple seasonalities, holiday effects, missing data.

• Pros: flexible components, interpretable; Cons: requires care to avoid overfitting.

Tip for Shopify merchants: if you prefer a packaged approach to automate baselines and seasonality handling, a Shopify inventory forecasting app can streamline setup and monitoring. Explore options such as Shopify inventory forecasting to streamline this process.

Causal and machine-learning methods

Regression with external drivers

• What: linear or regularized models using price, promotions, marketing, weather, macro signals.

• When: demand is driver-sensitive; need “what-if” analysis.

• Pros: interpretable coefficients, scenario modeling; Cons: needs feature engineering, risk of leakage.

Econometric models (VAR, VECM)

• What: model multiple interdependent time series and cointegration.

• When: macro-finance, categories that move together.

• Pros: captures cross-effects; Cons: data-hungry, complex diagnostics.

Gradient boosting and random forests

• What: tree ensembles (e.g., gradient boosting) with lag features, calendars, and drivers.

• When: nonlinear effects and interactions.

• Pros: strong accuracy, handles mixed data; Cons: opaque, careful validation needed.

Neural networks (RNN/LSTM)

• What: sequence models capturing long-range dependencies.

• When: high-frequency data with complex patterns.

• Pros: expressive; Cons: requires substantial data, tuning, and compute.

Dynamic regression and transfer functions

• What: regression with time-series error structure and lagged driver effects.

• When: delayed promo/price impacts or advertising carryover.

• Pros: aligns to business causality; Cons: setup complexity.

Price elasticity and promotion modeling

• What: estimate demand response to price and promotions.

• When: pricing strategy, markdown optimization.

• Pros: actionable sensitivity; Cons: confounded by seasonality and assortment changes without careful controls.

By horizon, granularity, and hierarchy

Short vs medium vs long-term approaches

• Short-term: exponential smoothing, SARIMA, state-space, nowcasting with high-frequency signals.

• Medium-term: combine time-series with drivers (dynamic regression, Prophet with holidays).

• Long-term: causal econometrics, scenarios, and blended ensembles; emphasize intervals.

Intermittent demand (Croston and variants)

• What: sporadic, low-volume items (service parts, long-tail SKUs).

• Methods: Croston, SBA (bias-corrected), TSB (Teunter–Syntetos–Babai) for decaying demand probability.

• Tips: forecast demand size and interval separately; evaluate with intermittent metrics (e.g., MAE, MASE over nonzero periods).

If your catalog includes long-tail items on WooCommerce, consider using a WooCommerce inventory forecasting plugin to handle sporadic demand patterns and automate safety stock logic.

Hierarchical forecasting and reconciliation

• Structures: product category → SKU → size/color; region → store → shelf.

• Strategies:

  • Bottom-up for high-variance leaf nodes with rich history.

  • Top-down when leaf data is noisy or scarce.

  • Reconciliation (e.g., MinT) to ensure totals equal the sum of parts.

• Practice: forecast at multiple levels, then reconcile to improve coherence.

New product forecasting and analogs

• Map to closest analog SKU/category and adjust for price, channel, and expected lift.

• Use expert ranges and early signal assimilation (preorders, waitlists, click-through).

Nowcasting with real-time signals

• Incorporate search trends, site traffic, add-to-cart rates, and weather.

• Use as complementary short-horizon updates on top of baseline models.

Spatial and portfolio aggregation

• Aggregate across regions or channels to improve stability.

• Use portfolio variance to set safety stock at higher levels while allowing local flexibility.

How to choose the right method

Quick method-chooser matrix

• Few data points, new item, or regime change:

  • Start: expert judgment, analogs, scenario bands.

  • Add: simple moving average or SES once minimal data accrues.

• Clear seasonality and stable trend:

  • Start: Holt‑Winters or SARIMA.

  • Upgrade: structural time series with holidays.

• Strong driver effects (price, promo, ads):

  • Start: regression with drivers, dynamic regression.

  • Upgrade: gradient boosting with lagged features.

• Intermittent demand:

  • Start: Croston/SBA/TSB.

  • Add: bootstrapped intervals and service-level safety stock.

• Large hierarchies:

  • Start: bottom-up or middle-out.

  • Upgrade: reconciled forecasts (e.g., MinT).

Text flowchart (visual guide):

• Is history < 6–12 periods? → Use qualitative + naive/SES → Collect more data.

• Otherwise, is there clear seasonality? → Yes → Holt‑Winters/SARIMA; No → SES/Holt/ARIMA.

• Do external drivers meaningfully explain variance? → Yes → Dynamic regression/GBM; No → stick to time-series.

• Is demand intermittent? → Yes → Croston variants.

• Are there hierarchical constraints? → Apply reconciliation post-forecast.

Diagnostic questions: trend, seasonality, intermittency

• Trend: is there a consistent upward/downward slope?

• Seasonality: repeating patterns weekly, monthly, yearly?

• Intermittency: many zeros or long gaps between sales?

• Events: are holidays/promos/stockouts visible in history?

Data length and quality thresholds

• Trend detection: 12+ points; seasonality: 2–3 full seasonal cycles (e.g., 24–36 months for annual seasonality).

• ARIMA/SARIMA: 50–100 observations recommended.

• ML models: more is better; ensure at least hundreds to thousands of rows with reliable features.

• Quality checklist:

  • Remove or flag stockouts and backorders.

  • Record price and promotion flags.

  • Normalize calendar (time zones, DST).

  • Impute or explain data gaps.

Compute, tooling, and skills

• Spreadsheets handle baselines (naive, moving average, SES) and simple Holt‑Winters.

• BI tools visualize diagnostics; Python/R scale to ARIMA, Prophet, ML.

• If you prefer guided options, curated inventory planning tools can accelerate setup for common ecommerce workflows.

Cost of error and business constraints

• Define asymmetry: Is over-forecasting (overstock) worse than under-forecasting (stockout)?

• Translate into service levels and loss functions.

• Set horizons by decision cadence (replenishment vs S&OP).

Pilot, compare, and iterate workflow

• Frame: objective, horizon, aggregation level, and cost-of-error.

• Split: rolling-origin backtests to mimic real decisions.

• Compare: use a simple baseline; only keep models that beat it.

• Calibrate: tune parameters and intervals.

• Operationalize: monitor error, bias, and drift; retrain on schedule.

Measuring accuracy and uncertainty

Metrics: MAPE, sMAPE, MASE, RMSE

• MAPE: intuitive percent error; misleading near zeros.

• sMAPE: symmetric alternative; still sensitive to low volumes.

• MASE: scale-free, comparable across series; robust choice.

• RMSE/MAE: absolute errors; choose based on penalty for large misses.

Backtesting and time-series cross-validation

• Use rolling-origin evaluation:

  • Train on initial window → forecast next h → roll forward → repeat.

• Keep folds aligned with your decision cadence (e.g., weekly).

• Log all splits, metrics, and winners to avoid hindsight bias.

Prediction intervals and scenario bands

• Produce 50/80/95% intervals to communicate uncertainty.

• For promotions and price tests, add scenario bands (best/base/worst) using driver assumptions.

• Use bootstrapping or model-based variance to construct intervals.

Bias and calibration checks

• Track bias (mean forecast error) by item, category, and region.

• Calibration: a 90% interval should contain actuals ~90% of the time.

• Address persistent bias with model re‑specification or post‑adjustments.

Monitoring drift and model governance

• Watch for drift in level, variance, and seasonality.

• Set retrain triggers (e.g., MASE > threshold for 3 consecutive periods).

• Version models, document assumptions, and control access.

Combining and ensembling forecasts

• Simple average of diverse models is often robust.

• Weighted ensembles by past performance across items.

• Use stacking with caution; avoid leakage and overfitting.

Common pitfalls and best practices

Data leakage and overfitting

• Never include future-known info in training (post-period features).

• Keep validation strictly out-of-sample and time-ordered.

Holiday, promotion, and event effects

• Add explicit holiday calendars and promo flags.

• Model lag effects for ads and markdowns.

Outliers and regime shifts

• Detect with robust stats; cap or explain spikes.

• Segment pre/post regime changes; don’t force a single model across breaks.

Feature engineering and external signals

• Lag features (t−1, t−7, t−52), rolling means, and holiday proximity.

• External: price, stockouts, marketing spend, weather; ensure alignment and quality.

Documentation and stakeholder communication

• Keep model cards: data used, assumptions, metrics, and limitations.

• Share forecast ranges, not just points; clarify confidence levels.

Change management and adoption

• Co-design with planners; align to existing planning cadences.

• Start with a small portfolio, prove accuracy, then scale.

Templates, tools, and next steps

Spreadsheet template and checklist

• Columns: date, actuals, promo_flag, price, stockout_flag, moving_avg_k, SES, Holt_Winters, baseline_forecast, error, abs_error, MASE_baseline.

• Steps:

  • Compute naive and seasonal naive baselines.

  • Add SES/Holt‑Winters using built-in functions or simple formulas.

  • Create rolling-origin evaluation: lock training window, project h steps, capture errors, roll forward.

  • Summarize metrics (MAE, MASE, sMAPE) by item and horizon.

• Checklist:

  • At least 24 months of data for annual seasonality.

  • Cleaned anomalies and documented events.

  • Defined service levels and cost-of-error.

Example datasets and benchmarks

• Use a subset of your catalog across demand profiles (fast, medium, slow, intermittent).

• Include at least one promotional period, one stockout case, and one price change.

Tooling options: spreadsheets, BI, Python/R

• Spreadsheets: fast pilots and baseline methods.

• BI: diagnostics and stakeholder visibility.

• Python/R: ARIMA, Prophet-style models, gradient boosting, hierarchical reconciliation.

If you run a Shopify storefront and want automated baselines without heavy coding, consider exploring Shopify inventory forecasting solutions that integrate with your catalog and order history.

Implementation timeline

• Week 1: data audit, baseline setup, diagnostic plots.

• Weeks 2–3: method trials (time-series vs causal), backtests, interval calibration.

• Week 4: pilot go-live for a subset; set monitoring thresholds.

• Ongoing: monthly retrain or after major assortment/price changes.

Roles and responsibilities

• Business owner: define decisions, horizons, and cost-of-error.

• Analyst: data prep, modeling, evaluation.

• Ops/planner: feedback loop, overrides under governance.

• Engineering/IT: data pipelines, deployments, access control.

Further reading and standards

• Explore practical guides and case studies on our inventory management blog for deeper dives into demand planning and replenishment topics.

FAQs

What are the main types of forecasting methods?

• Qualitative: expert judgment, Delphi, scenarios, analogs.

• Quantitative time-series: naive, moving average, exponential smoothing, decomposition, ARIMA/SARIMA, structural models.

• Causal/ML: regression with drivers, VAR/VECM, gradient boosting, neural nets, dynamic regression.

How do I choose between time-series and causal approaches?

• Use time-series when past patterns (trend/seasonality) dominate and drivers are weak or unavailable.

• Use causal when external factors (price, promotions, marketing, weather) explain variance or you need what‑if scenarios.

• Many teams blend both: time-series baseline plus driver adjustments.

Which method works best when I have very little historical data?

• Start with qualitative judgment, analogs, and scenario ranges.

• Use naive or SES once minimal data exists.

• Incorporate early real-time signals (preorders, traffic) to refine.

How much history do I need to capture seasonality reliably?

• Aim for 2–3 full seasonal cycles (e.g., 24–36 months for annual seasonality).

• For weekly seasonality, at least 52–104 weeks improves stability.

What accuracy metric should I use and why can MAPE be misleading?

• Prefer MASE or MAE when volumes are low or intermittent.

• MAPE can explode near zero actuals, overstating error.

• sMAPE mitigates some issues but still needs caution at low volumes.

How do I forecast intermittent or low-volume demand?

• Use Croston, SBA, or TSB to model demand size and interval separately.

• Evaluate with MAE/MASE and use probabilistic safety stock rather than point estimates.

Should I combine multiple methods to improve accuracy?

• Yes. Simple or performance-weighted ensembles often improve robustness, especially across diverse items and horizons.

What is the difference between a forecast, a projection, and a scenario?

• Forecast: model-based estimate of the most likely outcome (often with intervals).

• Projection: arithmetic extension of assumptions (e.g., +5% growth) without fitting.

• Scenario: coherent alternative path based on narrative assumptions (best/base/worst).

• For curated utilities that support the selection workflow and data checks, see inventory planning tools.

• If your stack is WooCommerce, a WooCommerce inventory forecasting plugin can help operationalize these methods for products with irregular sales patterns.