A Sales Forecasting Model for the Consumer Goods with Holiday Effects

Mu Zhang, Xiaonan Huang, Chang-bing YangJournal of Risk Analysis and Crisis Response 2020 · DOI

Abstract: Predicting consumer goods sales is challenging due to the simultaneous presence of long-term growth trends and intense seasonal volatility, often driven by "holiday effects." This paper proposes a hybrid forecasting framework that utilizes seasonal decomposition to isolate these components. By applying an ARIMA model to the seasonally adjusted data and then re-integrating the seasonal factors, the model achieves superior predictive accuracy. Testing on a real-world cigarette sales dataset demonstrates an Absolute Average Error (AAE) of 4.20%, significantly outperforming standard time-series models during peak holiday periods.

Introduction

In the fast-moving consumer goods (FMCG) sector, accurate sales forecasting is vital for inventory management and strategic planning. However, these datasets typically exhibit a "double trend" characteristic: a steady long-term growth trajectory overlaid with sharp, periodic fluctuations. These fluctuations are largely driven by the "holiday effect"—a surge in consumption behavior surrounding major cultural events. Traditional linear models often fail to capture these non-linear shocks, leading to significant forecasting errors during the most critical sales periods.

The Challenge of Holiday Effects

The holiday effect represents a "consumption behavior impulse." In regions like China, festivals such as the Spring Festival and National Day trigger dramatic spikes in sales for products like cigarettes and alcohol. These spikes are not merely random noise; they are systematic but volatile.

The original sales series of cigarettes in G City shows significant volatility and clear seasonal spikes.

As shown in the historical data for G City (Figure 3), sales peaks consistently occur in January and September, corresponding to major holidays. Standard quantitative models, such as basic ARIMA, often struggle with these sharp shocks because they assume a level of stationarity or gradual change that the holiday effect violates.

Methodology: A Three-Stage Pipeline

To address these challenges, the authors propose a hybrid approach designed to "clean" the data before forecasting.

1. Holiday Effect Testing

Before applying complex models, the researchers use a dummy variable model to statistically verify the holiday effect. By assigning binary values to pre-holiday periods, they confirm whether the observed spikes are statistically significant. This provides a rigorous mathematical foundation for proceeding with seasonal decomposition.

2. Seasonal Decomposition

The core of the approach is breaking the raw time-series data ($Y_t$) into four components: Trend ($T$), Seasonal ($S$), Cycle ($C$), and Irregular ($I$). Given that seasonal spikes in consumer goods often scale with total volume, a multiplicative model ($Y = T \times S \times C \times I$) is employed.

Comparison between the original volatile sales series (blue) and the extracted smoothed trend-cycle series (pink).

By isolating the seasonal factor ($S$), the researchers produce a "seasonally adjusted" series. Figure 4 illustrates how this process removes the violent holiday swings to reveal the underlying market momentum (the smoothed trend-cycle).

3. ARIMA Modeling and Re-composition

An ARIMA(1, 1, 1) model is then fitted to the "clean," seasonally adjusted data. Because the noise has been removed, the ARIMA model can more accurately project the future trend. Finally, the forecast is multiplied by the previously isolated seasonal factor to produce a final "holiday-aware" prediction.

Results and Discussion

The model was evaluated using 48 months of cigarette sales data from the G City Branch of the Guizhou Tobacco Company. The hybrid model was tested against actual sales from early 2011 and compared to several industry benchmarks.

The results were compelling:

  • Proposed Hybrid Model: 4.20% Absolute Average Error (AAE).
  • Product Seasonal ARIMA: 4.30% AAE.
  • Winters Multiplicative Model: 5.49% AAE.
  • Non-seasonal ARIMA: 11.05% AAE.

The most dramatic difference appeared during peak months. In January 2011, a major holiday month, the non-seasonal ARIMA model underestimated sales by 32.92%, while the proposed hybrid model restricted the error to just -7.60%.

Conclusion

This research demonstrates that for consumer goods, "cleaning" data of seasonal volatility is a prerequisite for accurate forecasting. By combining the statistical rigor of seasonal decomposition with the predictive power of ARIMA, the proposed model offers a robust tool for navigating the complexities of holiday-driven markets. Future work may involve exploring ensemble methods that blend multiple models to further mitigate prediction risks in increasingly dynamic consumer environments.

Summary updated

Abstract

model seasonal factor
BibTeX
@article{zhang2020sales,
  author = {Mu Zhang and Xiaonan Huang and Chang-bing Yang},
  title = {{A Sales Forecasting Model for the Consumer Goods with Holiday Effects}},
  journal = {Journal of Risk Analysis and Crisis Response},
  year = {2020},
  doi = {10.2991/jracr.k.200709.001},
}
Figure