Dema’s order forecasting service transforms your historical order data into precise, forward-looking predictions using seasonal decomposition models. By analyzing recurring patterns, trends, and seasonal behaviors, the system delivers intelligent insights that empower you to optimize operations, plan capacity, and make data-driven business decisions.

​
How to access the forecasts

The forecasts are available in the Reports section of the platform as a selectable forecast metrics.

​
Understanding time series forecasting

Time series forecasting predicts future values based solely on historical patterns within your data over time. Unlike regression models that rely on external variables, our approach focuses exclusively on understanding inherent patterns within your order history. Additionally, the models are trained top-down, starting with the highest level and allocating the forecasted values down to the smallest level of granularity based on a smart reconciliation algorithm. The granularity hierarchy used for the service is Storefront × Country × Channel Group x Channel x Campaign.

Core training metrics:

  • Order count: The number of orders placed over time
  • Average order value: The mean value of orders for each time period

While models train only on these two metrics, the system generates comprehensive business metrics using rolling averages of stable relationships down to the smallest level of granularity.

Forecasted Value = Raw Predicted Metric(s) × Stable Relationship(s) × Distribution Share

An example of a derived metric is Gross Sales, which is calculated as below, for each granularity level.

Gross Sales = Forecasted Orders × Forecasted AOV × Gross Sales Share

Multi-dimensional forecasting:

  • Daily & hourly predictions: Daily forecasts distributed hourly using historical patterns
  • Business dimensions: Allocated across geography, channels, and categories using stable rolling shares

​
Available forecasting models

Seasonal decomposition model

Sophisticated model for complex seasonality using nonlinear regression y(t) = g(t) + s(t) + h(t) + ε(t) to decompose the time series into trend, seasonal, and residual components.

Naive model

Efficient model that takes last year’s performance and scales with recent trends. Fast and transparent for stable seasonal patterns.

Seasonal decomposition model works best with:

  • Strong seasonality and multiple seasons of historical data
  • Automatic changepoint detection and multiple seasonality handling
  • Holiday and special event incorporation
  • Requires at least 2 years of historical data

Naive model works well when:

  • Seasonal patterns are stable or slowly changing
  • Quick, interpretable forecasts are needed
  • Requires less historical data

The seasonal-decomposition model needs a more historical data and takes longer to configure. If onboarding speed is critical, launch first with the naive model as baseline and move to seasonal decomposition model once you have enough data.

​
System assumptions and limitations

Key assumptions:

  • Pattern continuity: Seasonal and hourly distribution patterns evolve gradually
  • Historical relevance: Past patterns continue in similar forms

When forecasts may be less accurate:

  • Major promotional campaigns altering conversion rates
  • Business strategy changes (new channels, geographic expansion)
  • Market disruptions breaking historical patterns
  • Limited historical data for new businesses

During significant business changes, treat forecasts as directional guidance until patterns stabilize.