Sales forecasting is a quantitative estimate of future demand for an item or group of items over a specified period, expressed in volume (units, value) and with an uncertainty level. It provides concrete numbers for restocking, purchase planning and promotional campaigns.
How sales forecasting works
Forecasting reduces to three sequential steps: data collection, model selection and adapting the forecast to business realities. For a Kaspi.kz seller this practically means combining historical sales, promotional events, warehouse operations and external factors (seasonality, holidays, promotions).
- Data collection: orders by SKU, card page views, conversion rate, stock levels across warehouses (including FBS/FBO if you use Kaspi fulfilment), returns, cancellations, advertising budgets.
- Preprocessing: cleaning outliers (single large purchases), adjusting for promo spikes (sales like 11.11, Black Friday, Nauryz) and filling gaps.
- Modeling: simple methods (moving averages, exponential smoothing), regression models (accounting for price, traffic, promotions) and more advanced approaches — time series with seasonality and external regressors or ML models per SKU.
- Quality check: metrics like MAPE and bias to assess accuracy and systematic drift; typical targets are MAPE 10–30% per SKU depending on sales velocity.
Why a Kaspi.kz seller needs forecasting
Concrete benefits for a seller in Kazakhstan include reduced stockouts and overstock, lower warehousing costs and higher turnover of product cards on Kaspi:
- Reducing OOS (out-of-stock): even 1–2 days of downtime for a popular SKU on Kaspi can cost 5–15% of the monthly turnover for that item due to lost visibility and ranking.
- Optimizing purchases: forecasting helps form supplier orders considering MOQ and lead time; example: a supplier in China gives MOQ 100 pcs and delivery 21–35 days — a wrong forecast leads to stockouts or tied-up capital.
- Warehouse and logistics management: for sellers using Kaspi Fulfillment or FBS, forecasting allows planning shipments to Kaspi warehouses and reduces rush-delivery surcharges.
- Advertising planning: adjusting budgets for Kaspi CPA/CPM and other channels to expected demand improves ROI — e.g., increasing ad spend on a card with forecasted demand spike can boost sales without a significant CAC rise.
Examples and cases on Kaspi.kz
Below are realistic but simplified scenarios Kaspi sellers frequently encounter in Kazakhstan.
Case 1 — seasonal item: spring goods for Nauryz
- Context: the seller offers seasonal spring goods (decor, gifts, themed items) that see a strong, short-term demand spike around Nauryz.
- Approach: use 3–4 years of historical sales to measure typical Nauryz uplift, mark the Nauryz promo period as a separate regressor and build two forecasts: baseline and promo. Coordinate lead times with suppliers to ensure stock arrives 7–10 days before peak demand.
- Result: accurate stocking reduced OOS during the holiday and avoided large end-of-season markdowns, increasing seasonal margin.
Case 2 — launching a new SKU
- Context: a new product with no internal history needs initial stock for Kaspi marketplace testing.
- Approach: cluster similar items by category, price and attributes to form a proxy demand profile; start with a conservative test batch, run a short ad test on Kaspi, and update forecasts daily as real data comes in.
- Result: faster learning curve, minimized dead stock and the ability to scale replenishment once the SKU demonstrates steady conversion.
Case 3 — electronics with frequent promotions
- Context: high-value electronics with frequent flash sales and large promo-driven spikes.
- Approach: separate promo and non-promo demand in the model, include price and ad spend as regressors, and hold buffer stock for anticipated promo periods. Use shorter forecasting horizons around planned promotions.
- Result: fewer cancellations and backorders during promotions, and better margin control by avoiding overbuying for non-promo periods.
Practical tips for implementing forecasting
- Start simple: begin with moving averages or exponential smoothing to get a baseline; then layer in regressors and more complex models as data and capabilities grow.
- Granularity: forecast by SKU when possible, but for low-volume items group by cluster (category, price tier). Hybrid approaches (SKU for top items, cluster for the long tail) work well.
- Automate data pipelines: centralize sales, stock, returns and advertising data. Automate preprocessing (outlier handling, promo tagging) to ensure reproducible forecasts.
- Integrate lead times and MOQ: embed supplier lead times, transit delays and MOQ rules into the replenishment logic rather than treating forecasting and ordering as separate processes.
- Monitoring and feedback: track MAPE and bias regularly, analyze error patterns by category and update models. Implement simple alerting for large forecast deviations.
- Cross-functional collaboration: involve supply, procurement and marketing teams — marketing plans and planned promotions should feed into the forecast as explicit inputs.
- Plan for exceptions: create playbooks for flash sales, supplier failures and logistics delays so forecasts can be quickly adjusted.
Risks and typical mistakes
- Ignoring promotions: not marking promo periods biases the baseline down or up and leads to either stockouts during promos or excess inventory after them.
- Over-aggregation: forecasting only at category level can hide SKU-level trends and inflate OOS for fast movers.
- Incorrect lead times: underestimating supplier or transit lead times causes late replenishment; overestimating increases working capital tied in stock.
- Not accounting for returns and cancellations: especially relevant for electronics and fashion — include expected return rates in net demand calculations.
- Overfitting: overly complex models trained on limited SKU history can perform worse in production; prefer parsimonious models and regular retraining.
Conclusion
Forecasting is a practical tool for Kaspi.kz sellers to reduce stockouts, optimize purchases and improve ad ROI. Start with simple, transparent methods, ensure clean and complete data (including promo tags and multi-warehouse stocks), and iterate — combining business rules (MOQ, lead time) with statistical models yields the best operational outcomes.
Часто задаваемые вопросы
- What Kaspi.kz data should be collected for an accurate SKU-level forecast?
- Collect orders by SKU, card page views, conversion rate, stock levels across warehouses (including FBS/FBO), returns and cancellations, as well as advertising budgets and promotion dates. Add external factors — seasonality, holidays and weather anomalies. The broader the set of regressors, the easier it is to separate promo and traffic effects from baseline demand.
- How to adjust historical data for sales events (11.11, Black Friday, Nauryz)?
- Mark promo periods and treat them as a separate scenario: either exclude them from the baseline series or model them with a "promo" regressor. For SKUs heavily affected by sales, use separate forecasts for regular and promo periods. This reduces distortion for smoothing methods and improves inventory planning.
- What frequency and forecast horizon to choose for restocking on Kaspi.kz?
- For perishable or high-frequency sales use daily forecasts with horizons of 2–4 weeks; for supplier purchases use weekly or monthly forecasts for 4–12 weeks. Choose frequency based on operational needs: daily monitoring and weekly order reviews are optimal for most sellers. The horizon should cover total ordering time, delivery and safety buffer.
- Which forecast quality metrics to set and what realistic targets for SKUs on Kaspi.kz?
- Evaluate accuracy with MAPE and systematic bias — they show error magnitude and tendency toward stockouts or overstocks. Per-SKU target MAPE typically ranges from 10–30% depending on turnover. Monitor error distribution across categories and adjust methods for the most important items.
- How to forecast new or low-turnover SKUs without historical sales?
- Use analogy with similar products: cluster by category, price and attributes and take an averaged demand profile. For new SKUs plan small test shipments and rapidly update forecasts as data arrives. Alternatively, set Bayesian priors with a conservative demand norm and adapt based on observed sales.
- How to account for FBO and FBS stocks and Kaspi logistics delays when creating supplier orders?
- Sum available stock across all channels (FBO, FBS, own warehouse), include incoming shipments and expected returns when calculating available inventory. Account for different lead times and safety levels per channel and include MOQ in order formation rules. This helps avoid double ordering and reduces OOS and overstock risk.
