February 02, 2026
Comparison of demand forecasting models for the electronics retail sector
Carlos Eduardo Amorim de Oliveira; Elisa Morandé Sales
Summary prepared by the ResumeAI tool, an artificial intelligence solution developed by Instituto Pecege focused on synthesis and writing.
This work develops and evaluates time series-based demand forecasting models for the IT category in a retail company, comparing statistical approaches with a focus on accuracy and applicability for inventory management. Demand forecasting is fundamental in supply chain management, especially in volatile sectors like consumer electronics. The IT category, with items such as notebooks and components, presents unique challenges: strong seasonality (Black Friday, back to school), rapid technological obsolescence, and the impact of promotional campaigns that generate abrupt fluctuations in consumption.
The absence of a robust forecasting system entails financial and operational consequences. Overestimating demand leads to excess inventory, tying up capital and increasing storage costs and obsolescence risk. Underestimating results in stockouts, which compromise revenue and customer loyalty. In this context, applying statistical models to analyze historical sales data is a strategic solution to mitigate risks. The ability to transform data into predictive intelligence is a competitive advantage (Fawcett & Provost, 2013).
Time series models are consolidated tools for sequential data analysis, based on the premise that past patterns contain information about future behavior (Morettin & Toloi, 2006). Techniques such as Holt-Winters and ARIMA (Autoregressive Integrated Moving Average) decompose a series into trend, seasonality, and noise to project future values. The advantage of these approaches is their ability to generate accurate forecasts based solely on the history of the variable of interest, which is useful when external data is scarce (Nielsen, 2020).
The application of these models in computer retail allows for the quantification of consumption patterns, offering a basis for purchasing decisions and inventory allocation. The systematic analysis of historical data enables the construction of models with greater predictive power and fosters a data-driven culture. By adjusting the models to reflect category-specific sales cycles, it is possible to anticipate demand peaks and troughs, allowing for more effective campaign planning and working capital management.
The development of an effective model requires a methodological process that begins with an exploratory data analysis, including graphical inspection, checking assumptions such as stationarity, and series decomposition to understand its structure (Hamilton, 1994). The choice of evaluation metrics aligned with business objectives is also a critical step. This study, therefore, seeks to build an analytical framework that can be understood, validated, and replicated for other product categories of the company.
The methodology adopts a quantitative approach, aligned with the principles of data science applied to business problems (Fawcett & Provost, 2013). The database used consists of the monthly sales history of the IT line, covering 30 months, from January 2023 to June 2025. The data were aggregated monthly using the Pandas library in Python (McKinney, 2018). To ensure the integrity of the series, months with no sales were filled with the value zero, reflecting the absence of demand.
Before modeling, an exploratory analysis was conducted. Each time series was decomposed into its three components: Trend, Seasonality, and Residuals (Morettin & Toloi, 2006). Subsequently, to assess the adequacy of ARIMA models, the Augmented Dickey-Fuller (ADF) Test was applied to verify the stationarity of the series, a prerequisite for many stochastic models (Fávero et al., 2014; Hamilton, 1994). A series is non-stationary when its statistical properties, such as mean and variance, change over time.
Four models with different levels of complexity were implemented, a recommended practice to avoid unnecessary complexity (Nielsen, 2020). The Naive Seasonal (SNAIVE) was used as a benchmark, predicting that the future value will be equal to that of the same period of the previous season. Simple Exponential Smoothing (SEX) was applied as a base model. The Additive Holt-Winters Method was chosen for its ability to model level, trend, and seasonality. Finally, the Auto-ARIMA approach was employed, using the auto_arima algorithm to find the combination of parameters (p,d,q)(P,D,Q) that minimizes the Akaike Information Criterion (AIC), ensuring a parsimonious model.
For validation, the series was divided into a training set (January 2023 to December 2024) and a test set (January 2025 to June 2025). The models were trained on the first period to predict the six months of the test period. Performance was evaluated by comparing predictions with actual values, using the Root Mean Squared Error (RMSE) and the Mean Absolute Percentage Error (MAPE). RMSE was prioritized for being more robust and more severely penalizing large errors, being less sensitive to demand values close to zero. MAPE can present inflated values in series with low demand, a common phenomenon in retail (Makridakis et al., 1998).
The application of the methodology revealed significant heterogeneity in the behavior of product families. The ADF Stationarity Test showed that, of the seven families analyzed, only two (“CORE I5” and “CORE I7”) were stationary (p-value ≤ 0.05), while the remaining five (“ACCESSORIES”, “GAMER NOTEBOOK”, “MONITOR”, “CORE I3”, and “PRINTERS”) were non-stationary. This result validated the need for models such as Auto-ARIMA and Holt-Winters, capable of handling non-stationarity (Hamilton, 1994). The decomposition of the aggregated series confirmed the presence of a growth trend and strong annual seasonality.
The quantitative evaluation of the models’ performance showed that none stood out as universally superior. The MAPE metric proved unstable, exceeding 100% for categories such as “PRINTERS” and “GAMER NOTEBOOK”. This behavior occurs when the mean absolute error is greater than the average actual value, a common artifact in low-volume series that makes MAPE an unreliable indicator. The RMSE distribution, in contrast, was more stable, reinforcing the decision to prioritize it as the main evaluation metric (Nielsen, 2020).
The analysis of the decomposition of each series, crossed with the RMSE results, revealed the main finding of the study: the best forecasting model is directly correlated with the series’ demand pattern, a pillar of data science applied to business (Fawcett & Provost, 2013). Three patterns emerged. The first was that of series dominated by a strong growth trend, such as the “CORE I5” family. For this, the Auto-ARIMA model presented the lowest RMSE (895.6), as its structure robustly captures and projects growth trajectories.
The second pattern was that of series dominated by strong seasonality, with a less defined trend, such as the “GAMER NOTEBOOK” family. The decomposition showed clear and recurring sales peaks. In this scenario, the Seasonal Naive (SNAIVE) model, despite its simplicity, had the best performance, with an RMSE of 584.1. SNAIVE’s premise that the previous year’s pattern will repeat itself aligned with the cyclical nature of demand, outperforming more complex models. This highlights that the suitability to the data pattern, rather than model complexity, is the determining factor.
The third product segment was that of intermittent or sporadic demand. The “PRINTERS” family was the main example, with a time series characterized by very low sales, many periods with zero value, and an absence of a clear trend or seasonality pattern. As a consequence, none of the traditional models managed to generate accurate predictions. The best result, obtained by Simple Exponential Smoothing, still presented a MAPE of 344.7% and an RMSE of 272.3. This result is not a modeling failure, but a crucial analytical finding: classic time series techniques are ineffective for intermittent demand data.
The discovery that different demand patterns require different modeling approaches has strategic implications. Instead of seeking a single model for the entire portfolio, the analysis points to the need for a segmented forecasting strategy. Business value lies in the ability to intelligently allocate modeling resources: apply simple models like SNAIVE for seasonal products; concentrate complex models like Auto-ARIMA on strategic products with a strong trend; and recognize intermittent demand products, which should be managed by other inventory policies, such as reorder points.
The results partially corroborate studies such as Simanungkalit (2024), which highlight the effectiveness of ARIMA models for series with a trend. However, this work advances by proposing a segmented strategy based on a prior diagnosis of the series pattern. Furthermore, it highlights the limitation of traditional models in scenarios of intermittent demand, paving the way for the exploration of more suitable models for this type of data, such as the Croston method, or the incorporation of exogenous variables.
This work demonstrated that the diagnostic and exploratory data analysis stage is as important as, or more important than, predictive modeling. Time series decomposition was the tool that revealed the portfolio’s heterogeneity, showing that it is divided into segments with distinct demand patterns: dominated by trend, by seasonality, and intermittent. This finding explains why no single model performed best in all scenarios. The study’s main contribution is not the identification of a superior algorithm, but the proposition of a two-phase analytical process — diagnose the series pattern and then select the appropriate model — as the most robust approach for demand forecasting in retail. The implementation of this framework allows the company to optimize resources, applying simple models where they are effective, complex ones where they are necessary, and recognizing cases where traditional forecasting is inadequate. It is concluded that the objective was achieved: it was demonstrated that the most effective approach for demand forecasting in the IT category is a segmented strategy, in which the choice of the predictive model is conditioned by the specific demand pattern of each product family, overcoming the search for a single, universal model.
References:
Favero, L. P.; Belfiore, P.; Silva, F. L.; Chan, B. L. 2014. Manual de Análise de Dados. 1ed. Elsevier, Rio de Janeiro, RJ, Brasil.
Fawcett, T.; Provost, F. 2013. Data Science for Business: What You Need to Know about Data Mining and Data-Analytic Thinking. O’Reilly Media, Sebastopol, CA, USA.
Hamilton, J. D. 1994. Time Series Analysis. 1ed. Princeton University Press, Princeton, NJ, USA.
Lazzeri, F. 2020. Machine Learning for Time Series Forecasting with Python. 1ed. Wiley, Hoboken, NJ, USA.
Makridakis, S.; Wheelwright, S. C.; Hyndman, R. J. 1998. Forecasting: Methods and Applications. 3ed. John Wiley & Sons, New York, NY, USA.
Mckinney, W. 2018. Python para Análise de Dados: tratamento de dados com Pandas, NumPy e IPython. 2ed. Novatec Editora, São Paulo, SP, Brasil.
Morettin, P. A.; Toloi, C. M. C. 2006. Análise de Séries Temporais. 2ed. Edgard Blücher, São Paulo, SP, Brasil.
Nielsen, A. 2020. Análise Prática de Séries Temporais: Predição com Estatística e Aprendizado de Máquina. 1ed. O’Reilly Media, Sebastopol, CA, USA.
Simanungkalit, E. 2024. Evaluation of ARIMA model performance in projecting future sales: case study on electronic products. Journal of Intelligent Decision Support System 7(4): 309-317.
Singer, C. D. P.; Motta, J. 2020. Análise de dados categorizados. 1ed. Edgard Blücher, São Paulo, SP, Brasil.
Executive summary from the Final Course Work of the Specialization in Data Science and Analytics of the MBA USP/Esalq
Learn more about the course; click here: