How to account for seasonal events (Nauryz, September back-to-school start, Kaspi sales) in demand forecasts?

Add indicator variables to the model for specific dates and for weeks before/after the event, and increase the weight of recent similar periods. Manually adjust the forecast before a promotion to reflect planned marketing activities and product availability, so you avoid under-ordering or excessive replenishment.

Which specific data from the Kaspi seller cabinet and warehouses should be collected for an accurate forecast?

Collect sales history by SKU, stock levels, returns, delivery dates and volumes, lead time from suppliers and FBS warehouses, plus advertising campaign and price data. These features let you account for promotion effects, price changes and logistics delays when building the model.

When are naive forecasts or moving averages enough, and when is ML required?

Naive methods and moving averages work for stable items with low volatility and minor seasonality. If an item has high variability, many external factors (promotions, price, traffic) or you manage a large assortment, move to multivariate models or machine learning to improve accuracy.

How to account for lead time and FBS supply constraints when planning orders?

Forecast demand for a period equal to lead time plus processing time, and recalculate recommended orders based on current stock and expected deliveries. For FBS add safety buffers according to supply variability and seasonality to compensate for delays.

What mistakes most often lead to excess stock or frequent stockouts and how to avoid them?

Common mistakes are relying only on historical sales without considering promotions and lead time, ignoring returns and poor SKU segmentation. Avoid these by including all relevant features in the model, segmenting SKUs by demand stability, and regularly applying manual corrections before major events.

Demand forecasting for a Kaspi.kz seller

Definition

Demand forecasting is the systematic calculation of expected product sales for a specific period based on historical data, seasonality, promotions and logistical constraints.

How demand forecasting works

Forecasting consists of three interconnected stages: data collection, model selection and adjustment of the forecast for external factors.

Data collection: sales history from the Kaspi.kz seller cabinet, stock levels, returns, advertising campaigns, prices, deliveries and lead time from suppliers or FBS warehouses FBS.
Model selection: from simple rules to statistical and algorithmic methods. In practice the following are used:

naive forecast (sales = last period sales);
moving average and exponential smoothing to reduce noise;
seasonality-aware models — seasonal exponential smoothing, SARIMA;
multivariate models: accounting for promotions, price, traffic and external factors;
machine learning for complex dependencies and a large number of features.

Adjustment: manual tuning of the forecast before major events — Nauryz, the September back-to-school start, sales and Kaspi marketing campaigns. This is critical: algorithms do not always capture simultaneous assortment changes or new marketing activities.

Why this matters for a Kaspi.kz seller

A forecast lets a seller reduce costs and increase revenue through managed inventory and timely logistics. Concrete benefits:

Reduced stockouts and lost sales: based on experience from Kazakhstan sellers, accurate planning cuts stockouts by 30–70% in electronics and home appliances—categories where lost sales are most visible.
Optimised working capital: reduced excess stock on average by 15–40% for retailers that implemented forecasting with seasonality and lead time accounted for.
Procurement and logistics planning: knowing demand 4–8 weeks ahead simplifies negotiations with suppliers in China and the CIS, where typical delivery times are 20–45 days. For deliveries from Almaty or Astana internal lead times are 1–5 days.
Promotion and campaign management: a forecast helps decide which items to include in Kaspi promotions and when, so demand is not pushed beyond supply capabilities.
Service metrics: controlling service level (fill rate). For example, a 95% fill-rate target requires extra safety stock but reduces order cancellations and negative reviews.

Examples and cases from sellers in Kazakhstan

Below are realistic scenarios and calculations commonly applied by Kaspi sellers.

Case 1 — Small electronics retailer (Almaty)

Scenario: weekly average sales of a phone accessory SKU = 120 units, with modest seasonality, lead time from the local supplier = 7 days (1 week). The seller uses a 4-week moving average and keeps a 1-week safety buffer.

Result: switching from manual ad-hoc orders to this simple forecast reduced stockout days by ~50% and lowered excess stock by ~20% within two months.

Case 2 — Large seller with imports from China

Scenario: a seller imports small appliances from China with lead time = 30 days and strong seasonality around Nauryz and autumn sales. They implemented SARIMA for seasonal trends and added promotion flags for Kaspi campaigns.

Result: forecasting demand 8 weeks ahead and coordinating promotions with suppliers reduced emergency air shipments and improved on-shelf availability during peak promotions.

Numerical example — reorder point calculation

Given: average weekly demand = 200 units, lead time = 3 weeks, demand standard deviation per week = 50 units, target service level ~95% (Z ≈ 1.65).

Lead time demand = 200 × 3 = 600
Safety stock ≈ Z × σ × sqrt(lead time) = 1.65 × 50 × √3 ≈ 143
Reorder point = lead time demand + safety stock = 600 + 143 = 743 units

When current stock falls to ~743 units the system should create a purchase recommendation considering supplier minimums and batch sizes.

Practical tips and formulas

Start simple: for many SKUs a naive forecast or 4–12 period moving average is sufficient. Move to exponential smoothing or SARIMA for seasonal SKUs and to ML for large assortments.
Key formulas:
- Naive forecast: Forecast_t = Sales_{t-1}
- Moving average (n periods): Forecast_t = (Sales_{t-1} + ... + Sales_{t-n}) / n
- Exponential smoothing: Forecast_t = α × Sales_{t-1} + (1−α) × Forecast_{t-1} (α between 0 and 1)
- Reorder point (ROP): ROP = Demand_during_lead_time + Safety_stock
- Safety stock (basic): SS = Z × σ_LT × sqrt(lead_time_periods)
Include promotion flags: add binary features for weeks with planned Kaspi promotions, your own campaigns or competitor promotions. Historical weeks around such events should be weighted higher when estimating uplift.
Segment SKUs: apply different models by segment: fast-moving, slow-moving, intermittent demand, promotional-only items. One model does not fit all.
Automate data pipelines: pull sales, returns and stock data from the Kaspi seller cabinet and your ERP; enrich with marketing calendars and supplier lead times.
Manual overrides: allow planners to adjust forecasts before big events (Nauryz, start of school season, major Kaspi campaigns).

Tools and integration

Common tools and integration points for Kaspi sellers:

Seller cabinet data: use exports or APIs from the Kaspi.kz seller interface as the primary source of truth for sales and returns.
ERP and inventory systems: integrate forecasts with procurement and warehouse systems to automate purchase orders and replenishment.
BI and spreadsheets: Excel/Google Sheets and BI tools are fine for prototypes and small assortments.
Advanced: dedicated forecasting platforms or in-house ML pipelines for large assortments and complex feature sets; integrate with FBS logistics and supplier portals.
APIs and webhooks: set up automated notifications when stock falls below ROP or when forecasted demand spikes due to campaigns.

Mistakes to avoid

Using only historical sales without accounting for promotions and lead time.
Ignoring returns, cancellations and seasonality.
Poor SKU segmentation — applying the same model to all items.
Lack of buffer for FBS variability — not adding safety stock when delivery variability is high.
Not allowing manual adjustments before big marketing events on Kaspi.

Conclusion

Demand forecasting for Kaspi.kz sellers is a mix of data, models and business rules. Start with simple methods, collect the right features from the seller cabinet and logistics partners (including FBS), and gradually move to more advanced models. Combine algorithmic forecasts with manual planning around major events to get the best business results.

Часто задаваемые вопросы

How to account for seasonal events (Nauryz, September back-to-school start, Kaspi sales) in demand forecasts?: Add indicator variables to the model for specific dates and for weeks before/after the event, and increase the weight of recent similar periods. Manually adjust the forecast before a promotion to reflect planned marketing activities and product availability, so you avoid under-ordering or excessive replenishment.
Which specific data from the Kaspi seller cabinet and warehouses should be collected for an accurate forecast?: Collect sales history by SKU, stock levels, returns, delivery dates and volumes, lead time from suppliers and FBS warehouses, plus advertising campaign and price data. These features let you account for promotion effects, price changes and logistics delays when building the model.
When are naive forecasts or moving averages enough, and when is ML required?: Naive methods and moving averages work for stable items with low volatility and minor seasonality. If an item has high variability, many external factors (promotions, price, traffic) or you manage a large assortment, move to multivariate models or machine learning to improve accuracy.
How to account for lead time and FBS supply constraints when planning orders?: Forecast demand for a period equal to lead time plus processing time, and recalculate recommended orders based on current stock and expected deliveries. For FBS add safety buffers according to supply variability and seasonality to compensate for delays.
What mistakes most often lead to excess stock or frequent stockouts and how to avoid them?: Common mistakes are relying only on historical sales without considering promotions and lead time, ignoring returns and poor SKU segmentation. Avoid these by including all relevant features in the model, segmenting SKUs by demand stability, and regularly applying manual corrections before major events.

Demand forecasting