Start with hard sales telemetry: retailer POS, EDI order files, and direct e-commerce sales (Shopify, Amazon, Etsy) because they reflect actual sell-through rather than just orders. Supplement with marketplace impressions and conversion rates to estimate demand elasticities during promotions. Add high-frequency signals—Google Trends for search interest, Instagram and TikTok engagement metrics, and pre-order volumes—so you can detect emergent trends fast. Finally, record promotional calendars and influencer/activity spikes as event flags; these are causal inputs to regression-based models. Align all signals to the same time bucket (weekly recommended) and store raw and reconciled figures so you can trace forecast drivers during post-mortems.

Model seasonality explicitly rather than relying on ad hoc rules. For stable seasonality use Holt-Winters (triple exponential smoothing) to capture level, trend, and seasonal components; for intermittent seasonal patterns, aggregate to a higher level (product family or channel) to stabilize estimates. Overlay event-based adjustments for holidays, prom/wedding seasons, and major influencer campaigns—create binary event regressors and limit uplift to measured historical magnitudes to avoid overfitting. Always run scenario forecasts (baseline, +25% promotion, +50% viral uplift) and use them to size buffers; convert the uplift scenario into additional lead-time demand to compute temporary safety stock needed for the event window.

Choose models by demand behavior: for continuous, smooth SKUs use exponential smoothing families (Holt-Winters) or ARIMA for autocorrelated series; for intermittent demand (many zeros or sporadic orders) use Croston’s method or its variants (SBA adjustment) because standard smoothing produces biased forecasts. For new-product or low-history SKUs, use analog forecasting—map to similar existing SKUs by attributes (length, finish, packaging) and borrow the maturity curve. For campaigns and price changes, build causal regression (OLS or penalized regressions) with marketing spend, price, and search volume as regressors. Always benchmark models with appropriate accuracy metrics: for intermittent demand prefer MASE or scaled absolute errors, and validate with rolling backtests rather than single holdouts.

Use a service-level driven formula: Safety Stock = z * σ_demand × sqrt(lead time), where σ_demand is the standard deviation of demand per period and z is the normal z-score for your target service level (e.g., z≈1.65 for ~95% service). Reorder Point = average demand per period × lead time + Safety Stock. If lead time varies, replace sqrt(lead time) with sqrt(mean lead time^2 + variance_of_lead_time/mean_period_length) or incorporate lead-time variability into σ_total. For intermittent SKUs, estimate σ_demand over an appropriately aggregated period (e.g., 4-week buckets) to avoid underestimating volatility. Review service level targets by SKU class (A/B/C) using an ABC analysis so you don’t over-capitalize inventory on low-contribution items.

High-value external inputs include Google Trends (category and motif searches), marketplace category sales and bestseller lists, and social listening for content velocity on platforms where nail content is consumed. Supplier-side indicators—component lead times, lacquer shortages, and freight-rate indexes—inform supply-side constraints and should feed into constrained-supply scenarios. Use macro retail indicators (monthly apparel/beauty retail sales reports) for broad apparel-beauty demand shifts. Integrate these external signals as features in causal models or as early-warning triggers for scenario simulations rather than direct one-to-one demand conversions; calibrate each external input against historical impact to quantify uplift multipliers.

Adopt a hierarchical forecasting process: forecast at an aggregated level (family or collection), then disaggregate to SKU using historical mix ratios and attribute similarity for new SKUs. For custom designs, track design cohorts—launch week, influencer exposure, and repeat reorder behaviour—and use cohort decay curves to model drop-off. Apply a demand planning workflow: collect weekly forecasts from SKU owners, reconcile with channel P&L targets, and generate a single consensus forecast that feeds replenishment calculations. Automate re-forecasting on a rolling weekly basis (e.g., 12-week horizon) and enforce monthly S&OP reviews for lead-time alignment; this keeps factory schedules and packaging orders synchronized with real demand while minimizing rush shipments and markdown risk.