Detecting Seasonality in Your Data

Granularity and history requirements

• Granularity: Use the level you plan at (daily or weekly for ecommerce; weekly often balances noise and signal).

• History: Aim for 2–3 full seasonal cycles. For yearly seasonality, 2–3 years. For weekly seasonality, 8–12 weeks can be enough.

• Coverage: Confirm continuous data without long gaps or stockouts masquerading as low demand.

Decomposition (STL) and seasonal strength

• Use STL (seasonal-trend decomposition using LOESS) to separate trend, seasonality, and remainder.

• Seasonal strength (Fs) ≈ 1 − Var(remainder)/Var(remainder + seasonal). Rule of thumb:

  • Fs ≥ 0.6: strong seasonality; use seasonal models.

  • 0.3 ≤ Fs < 0.6: moderate; consider seasonality with shrinkage.

  • Fs < 0.3: weak; default to non-seasonal methods.

Inspect stability: if seasonal patterns shift year to year, use adaptive methods.

ACF/PACF and period detection

• Autocorrelation (ACF) spikes at lag s suggest a seasonal period s (e.g., 7 for daily, 12 for monthly).

• Partial autocorrelation (PACF) helps set AR/MA orders in SARIMA.

• Use periodograms or Fourier terms if multiple seasonalities exist (e.g., weekly and yearly).

Managing movable holidays and events

• Create features for movable holidays (Easter, Ramadan, Lunar New Year) and region-specific events.

• Spread effects across lead/lag windows (build-up and post-event returns).

• Maintain a calendar layer you can update each year rather than hard-coding dates.

Data and Feature Engineering

Internal signals (sales, price, stockouts, returns)

• Demand signal: Use orders or unconstrained demand, not shipments, when possible.

• Price and promo flags: Discount depth, sitewide vs SKU-level, ad spend pulses.

• Stockouts and substitutions: Mark periods where you were out-of-stock; avoid training models on suppressed demand.

• Returns and cancellations: Net out returns or model them separately to plan reverse logistics.

External drivers (weather, promos, search, macro)

• Weather: Temperature, precipitation, humidity, snow days; sensitivity varies by category.

• Marketing: Campaign launches, influencer drops, email sends; use spend or reach as features.

• Search and social trends: Relative interest can provide early lift signals.

• Macro and local events: Income shocks, school openings, sports playoffs, festivals near stores.

Add external drivers only when they add lift in backtests; avoid overfitting to one-off events.

Calendar and event catalogs

• Static: Black Friday/Cyber Monday, Mother’s Day, Singles’ Day, back-to-school start windows.

• Movable: Easter, Lunar New Year, Ramadan, Diwali; regionally varied.

• Retail cadence: Paydays, payday weekends, month-end closeouts.

• Operational: Fulfillment cutoffs, shipping holidays, regional delivery constraints.

Keep a canonical catalog with effective-date ranges and geographies.

Data quality checks and outlier treatment

• Detect outliers via robust z-scores or median absolute deviation.

• Tag anomalies (promo, one-time bulk order, data error) and treat accordingly.

• Impute missing days/weeks with local medians or model-based fills.

• Audit duplicate orders, time-zone shifts, and SKU merges/splits.

Methods That Work for Seasonality

Seasonal naive and seasonal indices

• Seasonal naive: Forecast t equals observed value from the same season last cycle (e.g., last year same week). Strong baseline.

• Seasonal indices: Compute multiplicative or additive indices by period, then multiply a baseline forecast by the index.

  • Multiplicative suits proportional effects (e.g., December 2.0x).

  • Additive suits constant lifts (e.g., +30 units in December).

Useful when patterns are stable and interpretable.

Holt-Winters (additive vs multiplicative)

• Additive: Best when seasonal amplitude is constant across the year.

• Multiplicative: Best when amplitude scales with trend (common in retail growth).

• Strengths: Adaptive smoothing; handles level, trend, and seasonality with few parameters.

• Watchouts: Sensitive to initial states on short history; tune seasonal period carefully.

SARIMA and Prophet

• SARIMA: Flexible ARIMA with seasonal orders; good for stable, single-seasonality series. Supports holiday regressors.

• Prophet: Decomposable model with piecewise trend, multiple seasonalities via Fourier terms, and holiday effects; robust defaults, quick to iterate.

Both benefit from clean calendars and outlier flags.

Intermittent demand (Croston, SBA)

• Croston’s method: Separately models demand size and interval; reduces bias for sparse SKUs.

• SBA (Syntetos–Boylan Approximation): Bias-corrected Croston variant; often better in practice.

• Bootstrapping: Simulate demand occurrences to generate probabilistic forecasts.

• Consider zero-inflated or hurdle models if using ML for intermittent series.

Machine learning with time features

• Gradient boosting or random forests with:

  • Time features: day-of-week, week-of-year, month, Fourier terms.

  • Lagged demand and rolling stats (7/28/56-day means).

  • Exogenous features: price, promos, weather, search interest.

• Use grouped cross-validation by time; enforce causal lags to avoid leakage.

ML can capture nonlinear effects but needs careful feature discipline and backtesting.

Hierarchical reconciliation

• Forecast at SKU, category, and total, then reconcile to ensure additivity.

• Approaches: Bottom-up, top-down, middle-out, and MinT (optimal reconciliation).

• Multi-location: Reconcile across stores/warehouses; share information for thin series while preserving local patterns.

Implementation Step-by-Step

Define scope, aggregation, and hierarchies

• Forecast objectives: replenishment, capacity, staffing, or pricing windows.

• Granularity: Weekly for buying; daily for labor and fulfillment; aggregate where noise overwhelms signal.

• Hierarchies: SKU → subcategory → category → total; and location → region → network.

Decide what will be forecasted vs derived (e.g., channel split shares).

Train/validation splits and rolling backtests

• Use rolling-origin evaluation (e.g., 6–12 backtest windows).

• Keep feature availability realistic (no future promos, use lagged variables).

• Evaluate across seasonal periods (include Q4) to avoid cherry-picking.

Track errors by level (SKU, category, total) and by time (peak vs off-peak).

Model selection using a decision tree

• Is demand intermittent (≥30–50% zeros)?

  • Yes: Croston/SBA; or aggregate to family/location and use hierarchical reconciliation.

  • No: Continue.

• Do you have ≥2 full seasonal cycles?

  • Yes: Holt-Winters, SARIMA, or Prophet; optionally seasonal indices.

  • No: Use seasonal naive and indices; borrow strength from analog SKUs or hierarchy.

• Are multiple seasonalities present?

  • Yes: Prophet or ML with Fourier terms.

  • No: SARIMA/Holt-Winters likely sufficient.

• Strong exogenous drivers?

  • Add regressors to SARIMAX/Prophet or ML models; verify lift in backtests.

Tuning, diagnostics, and ensembling

• Tune smoothing/ARIMA orders via grid search guided by AIC/BIC and cross-validation.

• Check residuals: no autocorrelation, roughly white noise, stable variance.

• Ensemble complementary models (e.g., seasonal naive + Holt-Winters + ML) with weights learned on validation windows.

• Guardrails: shrink seasonal effects when Fs drops; cap promo lift to historical maxima.

Deployment and monitoring cadence

• Refit cadence: weekly for retail; daily during peak periods.

• Update calendars and promo plans continuously.

• Monitor bias and error drift; alert on changes in seasonal strength or lead-time variance.

If you run on Shopify, the Verve AI Shopify inventory forecasting app can centralize demand signals and seasonal calibrations in one workflow:

From Forecast to Decisions

Safety stock and service levels

• Translate forecast error into safety stock using target service levels and lead-time variability.

• Typical approach: safety stock = service factor × demand variability over lead time.

• Use distinct policies for peak vs off-peak (raise service levels before Q4, then normalize).

Replenishment and capacity planning

• Convert forecasts into planned orders using reorder points and lot sizes.

• Account for vendor calendars, MOQs, and port/holiday constraints.

• Capacity: align receiving, putaway, and carrier bookings with weekly peaked demand.

Pricing and promotion alignment

• Plan promo timing where you have headroom. Avoid discounting into constrained supply.

• Use uplift models to estimate incremental demand on top of seasonal baseline.

• Synchronize price changes with inventory and fulfillment capacity to protect margins.

Staffing and scheduling

• Staff to volume by day-of-week and season. Cross-train to flex during peaks.

• Use short-horizon forecasts for labor (intra-week) and long-horizon for hiring (8–12 weeks out).

• Coordinate with CX to anticipate post-peak returns and inquiries.

Measuring Accuracy and Value

Metrics: MAE, MAPE, WAPE, sMAPE

• MAE: average absolute error; robust and interpretable.

• WAPE: weighted absolute percentage error; good for portfolio rollups.

• MAPE/sMAPE: percentage errors; beware division by small actuals.

• For intermittent demand, use MAE/WAPE or scaled errors rather than MAPE.

Evaluate separately for peak periods and for high-value SKUs.

Bias and forecast value add (FVA)

• Bias = average signed error. Track positive vs negative bias by category and season.

• FVA: measure incremental improvement from each process step (stat model → adjusted → consensus).

• Stop steps that do not add value; double down on those that do.

Scenario planning and stress tests

• Build what-if scenarios: stronger/weaker season, promo-on/off, longer lead times.

• Use uplift caps and elasticity ranges derived from backtests.

• Stress-test fulfillment: simulate pick/pack and carrier constraints on peak days.

For deeper guides on KPIs and processes, browse our inventory management blog.

Drift detection and alerts

• Track rolling MAPE/WAPE vs last season; alert on >25–30% relative change.

• Monitor seasonal strength; if Fs drops >0.15, review calendar drivers.

• Change-point detection on level/trend; re-tune models after major shifts.

Common Pitfalls and How to Avoid Them

Overfitting and chasing noise

• Too many features without cross-validation leads to brittle forecasts.

• Keep parsimonious models; prefer interpretable seasonal indices early on.

• Regularize promo and weather effects; require repeated evidence across years.

Ignoring exogenous shocks and stockouts

• Exclude stockout periods from model training or impute unconstrained demand.

• Use event overrides for shocks (policy changes, supplier outages).

• Rebuild baselines post-shock and shrink seasonal effects until stability returns.

SKU sparsity and long-tail issues

• Aggregate thin SKUs to families or use hierarchical reconciliation.

• Use analogs and attribute-based clustering for cold starts.

• Adjust service levels by ABC class; don’t over-engineer the long tail.

Misaligned time buckets and calendars

• Align weekly calendars to your operational week (e.g., Sun–Sat).

• Standardize time zones across order and fulfillment systems.

• Maintain consistent holiday calendars by region and channel.

Quick-Start Toolkit

Method selection flowchart

• Step 1: Is demand intermittent (≥30–50% zeros)? Use SBA/Croston or aggregate.

• Step 2: Do you have ≥2 cycles? If yes, Holt-Winters/SARIMA/Prophet; if not, seasonal indices + naive.

• Step 3: Multiple seasonalities? Use Prophet or ML with Fourier terms.

• Step 4: Strong exogenous drivers? Add regressors; verify lift via backtests.

• Step 5: Need additivity across levels? Apply hierarchical reconciliation.

Seasonal index calculator template

• Compute baseline: moving average excluding seasonal window.

• For each period (e.g., week 1–52), index = actual / baseline (multiplicative) or actual − baseline (additive).

• Normalize indices so they average to 1.0 (multiplicative) or 0 (additive).

• Apply indices to baseline forecasts; shrink toward 1.0/0 if data is sparse.

Holiday and event calendar checklist

• Fixed-date retail holidays and shipping cutoffs.

• Movable holidays with build-up and cooldown windows.

• Region-specific school terms and local events.

• Promo schedule with discount depth and channels.

• Supplier shut-downs and logistics constraints.

Data readiness checklist

• 2–3 cycles of clean history at the planning granularity.

• Stockout flags, returns handling, and backorders identified.

• Price/promo histories aligned to order timestamps.

• Weather and search trend features (optional) with proper lags.

• Rolling backtest harness configured with realistic splits.

Recommended tools and resources

• Consider centralizing drivers, calendars, and forecasts in one workspace; explore inventory planning tools to standardize inputs and outputs:

• WooCommerce stores can embed forecasting in WP-Admin with the Verve AI WooCommerce inventory forecasting plugin.

Soft CTA: If you’re tackling seasonal planning for the first time, start with seasonal indices and a rolling backtest. A simple worksheet, a clean calendar, and a few what-if scenarios will surface most gaps before peak season hits.

FAQs

How much history do I need to model seasonality well?

Aim for 2–3 full seasonal cycles at the planning granularity. For yearly patterns, 2–3 years; for weekly patterns on daily data, 8–12 weeks may suffice. If history is short, start with seasonal naive and indices, borrow strength from category-level patterns, and use analog products.

What is the difference between seasonality and cyclicality?

Seasonality repeats at a fixed period (weekly, monthly, yearly). Cyclicality spans multi-year business cycles without a fixed period. Models handle seasonality with periodic terms; cycles are often captured in trend or via external macro indicators.

Which model should I start with for seasonal demand?

Use seasonal naive and seasonal indices as baselines. If you have 2+ cycles, try Holt-Winters (multiplicative for proportional effects). For stable series with one seasonality, SARIMA works well; for multiple seasonalities or holiday effects, Prophet is a good choice. Always benchmark against the naive model.

How do I forecast new or low-history products?

Use analog SKUs (similar attributes, price tier, channel) and inherit category-level seasonal indices. Apply proportional scaling from comparable products. Cluster by attributes to share patterns. As data accrues, transition to SKU-level models.

How do promotions and stockouts impact seasonal forecasts?

Promos create uplift on top of seasonality; model them as regressors and cap effects to historical ranges. Stockouts suppress observed demand; flag and adjust training data or impute unconstrained demand. Never let stockout periods teach the model that true demand was zero.

What accuracy is realistic and how should I measure it?

Target portfolio WAPE aligned to your category volatility (often 10–25% for retail categories, higher for sparse SKUs). Measure by value and by unit. Track bias, peak-period accuracy, and forecast value add (FVA). For intermittent items, prefer MAE/WAPE over MAPE.

Should I forecast daily, weekly, or monthly for seasonality?

Forecast at the cadence of decisions. Weekly is common for replenishment and buying. Daily is useful for labor and fulfillment. Aggregate when noise dominates signal; disaggregate with profiles if you need intra-week staffing.

How do I handle moving holidays like Easter or Lunar New Year?

Maintain a calendar table with movable holiday dates per region. Include lead and lag windows to capture build-up and post-event effects. Use holiday regressors in SARIMAX/Prophet or add engineered features to ML models. Re-validate effects each year as dates shift.